--- /dev/null
- run: echo "LIBRARY_PATH=$(cat gcc_path)" >> $GITHUB_ENV
+name: CI
+
+on:
+ - push
+ - pull_request
+
+jobs:
+ build:
+ runs-on: ubuntu-latest
+
+ strategy:
+ fail-fast: false
+
+ steps:
+ - uses: actions/checkout@v2
+
+ - name: Install packages
+ run: sudo apt-get install ninja-build ripgrep
+
+ - name: Download artifact
+ uses: dawidd6/action-download-artifact@v2
+ with:
+ workflow: main.yml
+ name: libgccjit.so
+ path: gcc-build
+ repo: antoyo/gcc
+
+ - name: Setup path to libgccjit
+ run: |
+ echo $(readlink -f gcc-build) > gcc_path
+ ln gcc-build/libgccjit.so gcc-build/libgccjit.so.0
+
+ - name: Set LIBRARY_PATH
++ run: |
++ echo "LIBRARY_PATH=$(cat gcc_path)" >> $GITHUB_ENV
++ echo "LD_LIBRARY_PATH=$(cat gcc_path)" >> $GITHUB_ENV
+
+ # https://github.com/actions/cache/issues/133
+ - name: Fixup owner of ~/.cargo/
+ # Don't remove the trailing /. It is necessary to follow the symlink.
+ run: sudo chown -R $(whoami):$(id -ng) ~/.cargo/
+
+ - name: Cache cargo installed crates
+ uses: actions/cache@v1.1.2
+ with:
+ path: ~/.cargo/bin
+ key: cargo-installed-crates2-ubuntu-latest
+
+ - name: Cache cargo registry
+ uses: actions/cache@v1
+ with:
+ path: ~/.cargo/registry
+ key: ${{ runner.os }}-cargo-registry2-${{ hashFiles('**/Cargo.lock') }}
+
+ - name: Cache cargo index
+ uses: actions/cache@v1
+ with:
+ path: ~/.cargo/git
+ key: ${{ runner.os }}-cargo-index-${{ hashFiles('**/Cargo.lock') }}
+
+ - name: Cache cargo target dir
+ uses: actions/cache@v1.1.2
+ with:
+ path: target
+ key: ${{ runner.os }}-cargo-build-target-${{ hashFiles('rust-toolchain') }}
+
+ - name: Build
+ run: |
+ ./prepare_build.sh
+ ./build.sh
++ cargo test
+ ./clean_all.sh
+
+ - name: Prepare dependencies
+ run: |
+ git config --global user.email "user@example.com"
+ git config --global user.name "User"
+ ./prepare.sh
+
+ # Compile is a separate step, as the actions-rs/cargo action supports error annotations
+ - name: Compile
+ uses: actions-rs/cargo@v1.0.3
+ with:
+ command: build
+ args: --release
+
+ - name: Test
+ run: |
+ # Enable backtraces for easier debugging
+ export RUST_BACKTRACE=1
+
+ # Reduce amount of benchmark runs as they are slow
+ export COMPILE_RUNS=2
+ export RUN_RUNS=2
+
+ ./test.sh --release
--- /dev/null
+target
+**/*.rs.bk
+*.rlib
+*.o
+perf.data
+perf.data.old
+*.events
+*.string*
+/build_sysroot/sysroot
++/build_sysroot/sysroot_src
+/build_sysroot/Cargo.lock
+/build_sysroot/test_target/Cargo.lock
+/rust
++/simple-raytracer
+/regex
+gimple*
+*asm
+res
+test-backend
++gcc_path
--- /dev/null
- source = "git+https://github.com/antoyo/gccjit.rs#0572117c7ffdfcb0e6c6526d45266c3f34796bea"
+# This file is automatically @generated by Cargo.
+# It is not intended for manual editing.
+version = 3
+
+[[package]]
+name = "aho-corasick"
+version = "0.7.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1e37cfd5e7657ada45f742d6e99ca5788580b5c529dc78faf11ece6dc702656f"
+dependencies = [
+ "memchr",
+]
+
+[[package]]
+name = "ar"
+version = "0.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "450575f58f7bee32816abbff470cbc47797397c2a81e0eaced4b98436daf52e1"
+
+[[package]]
+name = "autocfg"
+version = "1.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cdb031dd78e28731d87d56cc8ffef4a8f36ca26c38fe2de700543e627f8a464a"
+
+[[package]]
+name = "bitflags"
+version = "1.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cf1de2fe8c75bc145a2f577add951f8134889b4795d47466a54a5c846d691693"
+
+[[package]]
+name = "cfg-if"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "baf1de4339761588bc0619e3cbc0120ee582ebb74b53b4efbf79117bd2da40fd"
+
+[[package]]
+name = "crc32fast"
+version = "1.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "81156fece84ab6a9f2afdb109ce3ae577e42b1228441eded99bd77f627953b1a"
+dependencies = [
+ "cfg-if",
+]
+
+[[package]]
+name = "fm"
+version = "0.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "68fda3cff2cce84c19e5dfa5179a4b35d2c0f18b893f108002b8a6a54984acca"
+dependencies = [
+ "regex",
+]
+
+[[package]]
+name = "gccjit"
+version = "1.0.0"
- source = "git+https://github.com/antoyo/gccjit.rs#0572117c7ffdfcb0e6c6526d45266c3f34796bea"
++source = "git+https://github.com/antoyo/gccjit.rs#54be27e41fff7b6ab532e2e21a82df50a12b9ad3"
+dependencies = [
+ "gccjit_sys",
+]
+
+[[package]]
+name = "gccjit_sys"
+version = "0.0.1"
++source = "git+https://github.com/antoyo/gccjit.rs#54be27e41fff7b6ab532e2e21a82df50a12b9ad3"
+dependencies = [
+ "libc 0.1.12",
+]
+
+[[package]]
+name = "getopts"
+version = "0.2.21"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "14dbbfd5c71d70241ecf9e6f13737f7b5ce823821063188d7e46c41d371eebd5"
+dependencies = [
+ "unicode-width",
+]
+
+[[package]]
+name = "getrandom"
+version = "0.2.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7fcd999463524c52659517fe2cea98493cfe485d10565e7b0fb07dbba7ad2753"
+dependencies = [
+ "cfg-if",
+ "libc 0.2.98",
+ "wasi",
+]
+
+[[package]]
+name = "hashbrown"
+version = "0.11.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ab5ef0d4909ef3724cc8cce6ccc8572c5c817592e9285f5464f8e86f8bd3726e"
+
+[[package]]
+name = "hermit-abi"
+version = "0.1.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "62b467343b94ba476dcb2500d242dadbb39557df889310ac77c5d99100aaac33"
+dependencies = [
+ "libc 0.2.98",
+]
+
+[[package]]
+name = "indexmap"
+version = "1.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bc633605454125dec4b66843673f01c7df2b89479b32e0ed634e43a91cff62a5"
+dependencies = [
+ "autocfg",
+ "hashbrown",
+]
+
+[[package]]
+name = "lang_tester"
+version = "0.3.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "96bd995a092cac79868250589869b5a5d656b02a02bd74c8ebdc566dc7203090"
+dependencies = [
+ "fm",
+ "getopts",
+ "libc 0.2.98",
+ "num_cpus",
+ "termcolor",
+ "threadpool",
+ "wait-timeout",
+ "walkdir",
+]
+
+[[package]]
+name = "libc"
+version = "0.1.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e32a70cf75e5846d53a673923498228bbec6a8624708a9ea5645f075d6276122"
+
+[[package]]
+name = "libc"
+version = "0.2.98"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "320cfe77175da3a483efed4bc0adc1968ca050b098ce4f2f1c13a56626128790"
+
+[[package]]
+name = "memchr"
+version = "2.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b16bd47d9e329435e309c58469fe0791c2d0d1ba96ec0954152a5ae2b04387dc"
+
+[[package]]
+name = "num_cpus"
+version = "1.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "05499f3756671c15885fee9034446956fff3f243d6077b91e5767df161f766b3"
+dependencies = [
+ "hermit-abi",
+ "libc 0.2.98",
+]
+
+[[package]]
+name = "object"
+version = "0.25.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a38f2be3697a57b4060074ff41b44c16870d916ad7877c17696e063257482bc7"
+dependencies = [
+ "crc32fast",
+ "indexmap",
+ "memchr",
+]
+
+[[package]]
+name = "ppv-lite86"
+version = "0.2.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ac74c624d6b2d21f425f752262f42188365d7b8ff1aff74c82e45136510a4857"
+
+[[package]]
+name = "rand"
+version = "0.8.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2e7573632e6454cf6b99d7aac4ccca54be06da05aca2ef7423d22d27d4d4bcd8"
+dependencies = [
+ "libc 0.2.98",
+ "rand_chacha",
+ "rand_core",
+ "rand_hc",
+]
+
+[[package]]
+name = "rand_chacha"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e6c10a63a0fa32252be49d21e7709d4d4baf8d231c2dbce1eaa8141b9b127d88"
+dependencies = [
+ "ppv-lite86",
+ "rand_core",
+]
+
+[[package]]
+name = "rand_core"
+version = "0.6.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d34f1408f55294453790c48b2f1ebbb1c5b4b7563eb1f418bcfcfdbb06ebb4e7"
+dependencies = [
+ "getrandom",
+]
+
+[[package]]
+name = "rand_hc"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d51e9f596de227fda2ea6c84607f5558e196eeaf43c986b724ba4fb8fdf497e7"
+dependencies = [
+ "rand_core",
+]
+
+[[package]]
+name = "redox_syscall"
+version = "0.2.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5ab49abadf3f9e1c4bc499e8845e152ad87d2ad2d30371841171169e9d75feee"
+dependencies = [
+ "bitflags",
+]
+
+[[package]]
+name = "regex"
+version = "1.5.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d07a8629359eb56f1e2fb1652bb04212c072a87ba68546a04065d525673ac461"
+dependencies = [
+ "aho-corasick",
+ "memchr",
+ "regex-syntax",
+]
+
+[[package]]
+name = "regex-syntax"
+version = "0.6.25"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f497285884f3fcff424ffc933e56d7cbca511def0c9831a7f9b5f6153e3cc89b"
+
+[[package]]
+name = "remove_dir_all"
+version = "0.5.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3acd125665422973a33ac9d3dd2df85edad0f4ae9b00dafb1a05e43a9f5ef8e7"
+dependencies = [
+ "winapi",
+]
+
+[[package]]
+name = "rustc_codegen_gcc"
+version = "0.1.0"
+dependencies = [
+ "ar",
+ "gccjit",
+ "lang_tester",
+ "object",
+ "target-lexicon",
+ "tempfile",
+]
+
+[[package]]
+name = "same-file"
+version = "1.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "93fc1dc3aaa9bfed95e02e6eadabb4baf7e3078b0bd1b4d7b6b0b68378900502"
+dependencies = [
+ "winapi-util",
+]
+
+[[package]]
+name = "target-lexicon"
+version = "0.10.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ab0e7238dcc7b40a7be719a25365910f6807bd864f4cce6b2e6b873658e2b19d"
+
+[[package]]
+name = "tempfile"
+version = "3.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dac1c663cfc93810f88aed9b8941d48cabf856a1b111c29a40439018d870eb22"
+dependencies = [
+ "cfg-if",
+ "libc 0.2.98",
+ "rand",
+ "redox_syscall",
+ "remove_dir_all",
+ "winapi",
+]
+
+[[package]]
+name = "termcolor"
+version = "1.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2dfed899f0eb03f32ee8c6a0aabdb8a7949659e3466561fc0adf54e26d88c5f4"
+dependencies = [
+ "winapi-util",
+]
+
+[[package]]
+name = "threadpool"
+version = "1.8.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d050e60b33d41c19108b32cea32164033a9013fe3b46cbd4457559bfbf77afaa"
+dependencies = [
+ "num_cpus",
+]
+
+[[package]]
+name = "unicode-width"
+version = "0.1.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9337591893a19b88d8d87f2cec1e73fad5cdfd10e5a6f349f498ad6ea2ffb1e3"
+
+[[package]]
+name = "wait-timeout"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9f200f5b12eb75f8c1ed65abd4b2db8a6e1b138a20de009dacee265a2498f3f6"
+dependencies = [
+ "libc 0.2.98",
+]
+
+[[package]]
+name = "walkdir"
+version = "2.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "808cf2735cd4b6866113f648b791c6adc5714537bc222d9347bb203386ffda56"
+dependencies = [
+ "same-file",
+ "winapi",
+ "winapi-util",
+]
+
+[[package]]
+name = "wasi"
+version = "0.10.2+wasi-snapshot-preview1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fd6fbd9a79829dd1ad0cc20627bf1ed606756a7f77edff7b66b7064f9cb327c6"
+
+[[package]]
+name = "winapi"
+version = "0.3.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5c839a674fcd7a98952e593242ea400abe93992746761e38641405d28b00f419"
+dependencies = [
+ "winapi-i686-pc-windows-gnu",
+ "winapi-x86_64-pc-windows-gnu",
+]
+
+[[package]]
+name = "winapi-i686-pc-windows-gnu"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ac3b87c63620426dd9b991e5ce0329eff545bccbbb34f3be09ff6fb6ab51b7b6"
+
+[[package]]
+name = "winapi-util"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "70ec6ce85bb158151cae5e5c87f95a8e97d2c0c4b001223f33a334e3ce5de178"
+dependencies = [
+ "winapi",
+]
+
+[[package]]
+name = "winapi-x86_64-pc-windows-gnu"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "712e227841d057c1ee1cd2fb22fa7e5a5461ae8e48fa2ca79ec42cfc1931183f"
--- /dev/null
- $ git clone https://github.com/antoyo/rustc_codegen_gcc.git
+# WIP libgccjit codegen backend for rust
+
+This is a GCC codegen for rustc, which means it can be loaded by the existing rustc frontend, but benefits from GCC: more architectures are supported and GCC's optimizations are used.
+
+**Despite its name, libgccjit can be used for ahead-of-time compilation, as is used here.**
+
+## Motivation
+
+The primary goal of this project is to be able to compile Rust code on platforms unsupported by LLVM.
+A secondary goal is to check if using the gcc backend will provide any run-time speed improvement for the programs compiled using rustc.
+
+## Building
+
+**This requires a patched libgccjit in order to work.
+The patches in [this repostory](https://github.com/antoyo/libgccjit-patches) need to be applied.
+(Those patches should work when applied on master, but in case it doesn't work, they are known to work when applied on 079c23cfe079f203d5df83fea8e92a60c7d7e878.)
+You can also use my [fork of gcc](https://github.com/antoyo/gcc) which already includes these patches.**
+
+**Put the path to your custom build of libgccjit in the file `gcc_path`.**
+
+```bash
- * Build the stage1 compiler (`rustup toolchain link debug-current stage2 build/x86_64-unknown-linux-gnu/stage1`).
++$ git clone https://github.com/rust-lang/rustc_codegen_gcc.git
+$ cd rustc_codegen_gcc
+$ ./prepare_build.sh # download and patch sysroot src
+$ ./build.sh --release
+```
+
+To run the tests:
+
+```bash
+$ ./prepare.sh # download and patch sysroot src and install hyperfine for benchmarking
+$ ./test.sh --release
+```
+
+## Usage
+
+`$cg_gccjit_dir` is the directory you cloned this repo into in the following instructions.
+
+### Cargo
+
+```bash
+$ CHANNEL="release" $cg_gccjit_dir/cargo.sh run
+```
+
+If you compiled cg_gccjit in debug mode (aka you didn't pass `--release` to `./test.sh`) you should use `CHANNEL="debug"` instead or omit `CHANNEL="release"` completely.
+
+### Rustc
+
+> You should prefer using the Cargo method.
+
+```bash
+$ rustc +$(cat $cg_gccjit_dir/rust-toolchain) -Cpanic=abort -Zcodegen-backend=$cg_gccjit_dir/target/release/librustc_codegen_gcc.so --sysroot $cg_gccjit_dir/build_sysroot/sysroot my_crate.rs
+```
+
+## Env vars
+
+<dl>
+ <dt>CG_GCCJIT_INCR_CACHE_DISABLED</dt>
+ <dd>Don't cache object files in the incremental cache. Useful during development of cg_gccjit
+ to make it possible to use incremental mode for all analyses performed by rustc without caching
+ object files when their content should have been changed by a change to cg_gccjit.</dd>
+ <dt>CG_GCCJIT_DISPLAY_CG_TIME</dt>
+ <dd>Display the time it took to perform codegen for a crate</dd>
+</dl>
+
+## Debugging
+
+Sometimes, libgccjit will crash and output an error like this:
+
+```
+during RTL pass: expand
+libgccjit.so: error: in expmed_mode_index, at expmed.h:249
+0x7f0da2e61a35 expmed_mode_index
+ ../../../gcc/gcc/expmed.h:249
+0x7f0da2e61aa4 expmed_op_cost_ptr
+ ../../../gcc/gcc/expmed.h:271
+0x7f0da2e620dc sdiv_cost_ptr
+ ../../../gcc/gcc/expmed.h:540
+0x7f0da2e62129 sdiv_cost
+ ../../../gcc/gcc/expmed.h:558
+0x7f0da2e73c12 expand_divmod(int, tree_code, machine_mode, rtx_def*, rtx_def*, rtx_def*, int)
+ ../../../gcc/gcc/expmed.c:4335
+0x7f0da2ea1423 expand_expr_real_2(separate_ops*, rtx_def*, machine_mode, expand_modifier)
+ ../../../gcc/gcc/expr.c:9240
+0x7f0da2cd1a1e expand_gimple_stmt_1
+ ../../../gcc/gcc/cfgexpand.c:3796
+0x7f0da2cd1c30 expand_gimple_stmt
+ ../../../gcc/gcc/cfgexpand.c:3857
+0x7f0da2cd90a9 expand_gimple_basic_block
+ ../../../gcc/gcc/cfgexpand.c:5898
+0x7f0da2cdade8 execute
+ ../../../gcc/gcc/cfgexpand.c:6582
+```
+
+To see the code which causes this error, call the following function:
+
+```c
+gcc_jit_context_dump_to_file(ctxt, "/tmp/output.c", 1 /* update_locations */)
+```
+
+This will create a C-like file and add the locations into the IR pointing to this C file.
+Then, rerun the program and it will output the location in the second line:
+
+```
+libgccjit.so: /tmp/something.c:61322:0: error: in expmed_mode_index, at expmed.h:249
+```
+
+Or add a breakpoint to `add_error` in gdb and print the line number using:
+
+```
+p loc->m_line
+```
+
+### How to use a custom-build rustc
+
- * Add `~/.rustup/toolchains/debug-current/lib/rustlib/x86_64-unknown-linux-gnu/lib` to `LD_LIBRARY_PATH`.
++ * Build the stage2 compiler (`rustup toolchain link debug-current build/x86_64-unknown-linux-gnu/stage2`).
+ * Clean and rebuild the codegen with `debug-current` in the file `rust-toolchain`.
--- /dev/null
- export GCC_PATH=$(cat gcc_path)
+#!/bin/bash
+
+#set -x
+set -e
+
++if [ -f ./gcc_path ]; then
++ export GCC_PATH=$(cat gcc_path)
++else
++ echo 'Please put the path to your custom build of libgccjit in the file `gcc_path`, see Readme.md for details'
++ exit 1
++fi
+
+export LD_LIBRARY_PATH="$GCC_PATH"
+export LIBRARY_PATH="$GCC_PATH"
+
+if [[ "$1" == "--release" ]]; then
+ export CHANNEL='release'
+ CARGO_INCREMENTAL=1 cargo rustc --release
+else
+ echo $LD_LIBRARY_PATH
+ export CHANNEL='debug'
+ cargo rustc
+fi
+
+source config.sh
+
+rm -r target/out || true
+mkdir -p target/out/gccjit
+
+echo "[BUILD] sysroot"
+time ./build_sysroot/build_sysroot.sh $CHANNEL
--- /dev/null
- RUSTDOCFLAGS=$RUSTFLAGS cargo +${TOOLCHAIN} $cmd --target $TARGET_TRIPLE $@
+#!/bin/bash
+
+if [ -z $CHANNEL ]; then
+export CHANNEL='debug'
+fi
+
+pushd $(dirname "$0") >/dev/null
+source config.sh
+
+# read nightly compiler from rust-toolchain file
+TOOLCHAIN=$(cat rust-toolchain)
+
+popd >/dev/null
+
+if [[ $(rustc -V) != $(rustc +${TOOLCHAIN} -V) ]]; then
+ echo "rustc_codegen_gcc is build for $(rustc +${TOOLCHAIN} -V) but the default rustc version is $(rustc -V)."
+ echo "Using $(rustc +${TOOLCHAIN} -V)."
+fi
+
+cmd=$1
+shift
+
++RUSTDOCFLAGS="$RUSTFLAGS" cargo +${TOOLCHAIN} $cmd --target $TARGET_TRIPLE $@
--- /dev/null
- export GCC_PATH=$(cat gcc_path)
+set -e
+
+export CARGO_INCREMENTAL=0
+
- export RUSTFLAGS=$linker' -Cpanic=abort -Cdebuginfo=2 -Zpanic-abort-tests -Zcodegen-backend='$(pwd)'/target/'$CHANNEL'/librustc_codegen_gcc.'$dylib_ext' --sysroot '$(pwd)'/build_sysroot/sysroot'
++if [ -f ./gcc_path ]; then
++ export GCC_PATH=$(cat gcc_path)
++else
++ echo 'Please put the path to your custom build of libgccjit in the file `gcc_path`, see Readme.md for details'
++ exit 1
++fi
+
+unamestr=`uname`
+if [[ "$unamestr" == 'Linux' ]]; then
+ dylib_ext='so'
+elif [[ "$unamestr" == 'Darwin' ]]; then
+ dylib_ext='dylib'
+else
+ echo "Unsupported os"
+ exit 1
+fi
+
+HOST_TRIPLE=$(rustc -vV | grep host | cut -d: -f2 | tr -d " ")
+TARGET_TRIPLE=$HOST_TRIPLE
+#TARGET_TRIPLE="aarch64-unknown-linux-gnu"
+
+linker=''
+RUN_WRAPPER=''
+if [[ "$HOST_TRIPLE" != "$TARGET_TRIPLE" ]]; then
+ if [[ "$TARGET_TRIPLE" == "aarch64-unknown-linux-gnu" ]]; then
+ # We are cross-compiling for aarch64. Use the correct linker and run tests in qemu.
+ linker='-Clinker=aarch64-linux-gnu-gcc'
+ RUN_WRAPPER='qemu-aarch64 -L /usr/aarch64-linux-gnu'
+ else
+ echo "Unknown non-native platform"
+ fi
+fi
+
++export RUSTFLAGS="$linker -Cpanic=abort -Cdebuginfo=2 -Clto=off -Zpanic-abort-tests -Zcodegen-backend=$(pwd)/target/${CHANNEL:-debug}/librustc_codegen_gcc.$dylib_ext --sysroot $(pwd)/build_sysroot/sysroot"
+
+# FIXME(antoyo): remove once the atomic shim is gone
+if [[ `uname` == 'Darwin' ]]; then
+ export RUSTFLAGS="$RUSTFLAGS -Clink-arg=-undefined -Clink-arg=dynamic_lookup"
+fi
+
+RUSTC="rustc $RUSTFLAGS -L crate=target/out --out-dir target/out"
+export RUSTC_LOG=warn # display metadata load errors
+
+export LD_LIBRARY_PATH="$(pwd)/target/out:$(pwd)/build_sysroot/sysroot/lib/rustlib/$TARGET_TRIPLE/lib:$GCC_PATH"
+export DYLD_LIBRARY_PATH=$LD_LIBRARY_PATH
--- /dev/null
- nightly-2021-08-12
++nightly-2021-09-17
--- /dev/null
- pub(crate) unsafe fn codegen(tcx: TyCtxt<'_>, mods: &mut GccContext, kind: AllocatorKind, has_alloc_error_handler: bool) {
+use gccjit::{FunctionType, ToRValue};
+use rustc_ast::expand::allocator::{AllocatorKind, AllocatorTy, ALLOCATOR_METHODS};
+use rustc_middle::bug;
+use rustc_middle::ty::TyCtxt;
+use rustc_span::symbol::sym;
+
+use crate::GccContext;
+
++pub(crate) unsafe fn codegen(tcx: TyCtxt<'_>, mods: &mut GccContext, _module_name: &str, kind: AllocatorKind, has_alloc_error_handler: bool) {
+ let context = &mods.context;
+ let usize =
+ match tcx.sess.target.pointer_width {
+ 16 => context.new_type::<u16>(),
+ 32 => context.new_type::<u32>(),
+ 64 => context.new_type::<u64>(),
+ tws => bug!("Unsupported target word size for int: {}", tws),
+ };
+ let i8 = context.new_type::<i8>();
+ let i8p = i8.make_pointer();
+ let void = context.new_type::<()>();
+
+ for method in ALLOCATOR_METHODS {
+ let mut types = Vec::with_capacity(method.inputs.len());
+ for ty in method.inputs.iter() {
+ match *ty {
+ AllocatorTy::Layout => {
+ types.push(usize);
+ types.push(usize);
+ }
+ AllocatorTy::Ptr => types.push(i8p),
+ AllocatorTy::Usize => types.push(usize),
+
+ AllocatorTy::ResultPtr | AllocatorTy::Unit => panic!("invalid allocator arg"),
+ }
+ }
+ let output = match method.output {
+ AllocatorTy::ResultPtr => Some(i8p),
+ AllocatorTy::Unit => None,
+
+ AllocatorTy::Layout | AllocatorTy::Usize | AllocatorTy::Ptr => {
+ panic!("invalid allocator output")
+ }
+ };
+ let name = format!("__rust_{}", method.name);
+
+ let args: Vec<_> = types.iter().enumerate()
+ .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index)))
+ .collect();
+ let func = context.new_function(None, FunctionType::Exported, output.unwrap_or(void), &args, name, false);
+
+ if tcx.sess.target.options.default_hidden_visibility {
+ // TODO(antoyo): set visibility.
+ }
+ if tcx.sess.must_emit_unwind_tables() {
+ // TODO(antoyo): emit unwind tables.
+ }
+
+ let callee = kind.fn_name(method.name);
+ let args: Vec<_> = types.iter().enumerate()
+ .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index)))
+ .collect();
+ let callee = context.new_function(None, FunctionType::Extern, output.unwrap_or(void), &args, callee, false);
+ // TODO(antoyo): set visibility.
+
+ let block = func.new_block("entry");
+
+ let args = args
+ .iter()
+ .enumerate()
+ .map(|(i, _)| func.get_param(i as i32).to_rvalue())
+ .collect::<Vec<_>>();
+ let ret = context.new_call(None, callee, &args);
+ //llvm::LLVMSetTailCall(ret, True);
+ if output.is_some() {
+ block.end_with_return(None, ret);
+ }
+ else {
+ block.end_with_void_return(None);
+ }
++
++ // TODO(@Commeownist): Check if we need to emit some extra debugging info in certain circumstances
++ // as described in https://github.com/rust-lang/rust/commit/77a96ed5646f7c3ee8897693decc4626fe380643
+ }
+
+ let types = [usize, usize];
+ let name = "__rust_alloc_error_handler".to_string();
+ let args: Vec<_> = types.iter().enumerate()
+ .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index)))
+ .collect();
+ let func = context.new_function(None, FunctionType::Exported, void, &args, name, false);
+
+ let kind =
+ if has_alloc_error_handler {
+ AllocatorKind::Global
+ }
+ else {
+ AllocatorKind::Default
+ };
+ let callee = kind.fn_name(sym::oom);
+ let args: Vec<_> = types.iter().enumerate()
+ .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index)))
+ .collect();
+ let callee = context.new_function(None, FunctionType::Extern, void, &args, callee, false);
+ //llvm::LLVMRustSetVisibility(callee, llvm::Visibility::Hidden);
+
+ let block = func.new_block("entry");
+
+ let args = args
+ .iter()
+ .enumerate()
+ .map(|(i, _)| func.get_param(i as i32).to_rvalue())
+ .collect::<Vec<_>>();
+ let _ret = context.new_call(None, callee, &args);
+ //llvm::LLVMSetTailCall(ret, True);
+ block.end_with_void_return(None);
+}
--- /dev/null
- use rustc_codegen_ssa::back::archive::{find_library, ArchiveBuilder};
- use rustc_codegen_ssa::METADATA_FILENAME;
+use std::fs::File;
+use std::path::{Path, PathBuf};
+
+use rustc_session::Session;
- use rustc_span::symbol::Symbol;
++use rustc_codegen_ssa::back::archive::ArchiveBuilder;
++
+use rustc_data_structures::temp_dir::MaybeTempDir;
+use rustc_middle::middle::cstore::DllImport;
- lib_search_paths: Vec<PathBuf>,
++
+
+struct ArchiveConfig<'a> {
+ sess: &'a Session,
+ dst: PathBuf,
- use rustc_codegen_ssa::back::link::archive_search_paths;
+ use_native_ar: bool,
+ use_gnu_style_archive: bool,
+}
+
+#[derive(Debug)]
+enum ArchiveEntry {
+ FromArchive {
+ archive_index: usize,
+ entry_index: usize,
+ },
+ File(PathBuf),
+}
+
+pub struct ArArchiveBuilder<'a> {
+ config: ArchiveConfig<'a>,
+ src_archives: Vec<(PathBuf, ar::Archive<File>)>,
+ // Don't use `HashMap` here, as the order is important. `rust.metadata.bin` must always be at
+ // the end of an archive for linkers to not get confused.
+ entries: Vec<(String, ArchiveEntry)>,
+}
+
+impl<'a> ArchiveBuilder<'a> for ArArchiveBuilder<'a> {
+ fn new(sess: &'a Session, output: &Path, input: Option<&Path>) -> Self {
- lib_search_paths: archive_search_paths(sess),
+ let config = ArchiveConfig {
+ sess,
+ dst: output.to_path_buf(),
- fn add_native_library(&mut self, name: Symbol, verbatim: bool) {
- let location = find_library(name, verbatim, &self.config.lib_search_paths, self.config.sess);
- self.add_archive(location.clone(), |_| false)
- .unwrap_or_else(|e| {
- panic!(
- "failed to add native library {}: {}",
- location.to_string_lossy(),
- e
- );
- });
- }
-
- fn add_rlib(
- &mut self,
- rlib: &Path,
- name: &str,
- lto: bool,
- skip_objects: bool,
- ) -> std::io::Result<()> {
- let obj_start = name.to_owned();
-
- self.add_archive(rlib.to_owned(), move |fname: &str| {
- // Ignore metadata files, no matter the name.
- if fname == METADATA_FILENAME {
- return true;
- }
-
- // Don't include Rust objects if LTO is enabled
- if lto && fname.starts_with(&obj_start) && fname.ends_with(".o") {
- return true;
- }
+ use_native_ar: false,
+ // FIXME test for linux and System V derivatives instead
+ use_gnu_style_archive: sess.target.options.archive_format == "gnu",
+ };
+
+ let (src_archives, entries) = if let Some(input) = input {
+ let mut archive = ar::Archive::new(File::open(input).unwrap());
+ let mut entries = Vec::new();
+
+ let mut i = 0;
+ while let Some(entry) = archive.next_entry() {
+ let entry = entry.unwrap();
+ entries.push((
+ String::from_utf8(entry.header().identifier().to_vec()).unwrap(),
+ ArchiveEntry::FromArchive {
+ archive_index: 0,
+ entry_index: i,
+ },
+ ));
+ i += 1;
+ }
+
+ (vec![(input.to_owned(), archive)], entries)
+ } else {
+ (vec![], Vec::new())
+ };
+
+ ArArchiveBuilder {
+ config,
+ src_archives,
+ entries,
+ }
+ }
+
+ fn src_files(&mut self) -> Vec<String> {
+ self.entries.iter().map(|(name, _)| name.clone()).collect()
+ }
+
+ fn remove_file(&mut self, name: &str) {
+ let index = self
+ .entries
+ .iter()
+ .position(|(entry_name, _)| entry_name == name)
+ .expect("Tried to remove file not existing in src archive");
+ self.entries.remove(index);
+ }
+
+ fn add_file(&mut self, file: &Path) {
+ self.entries.push((
+ file.file_name().unwrap().to_str().unwrap().to_string(),
+ ArchiveEntry::File(file.to_owned()),
+ ));
+ }
+
- // Otherwise if this is *not* a rust object and we're skipping
- // objects then skip this file
- if skip_objects && (!fname.starts_with(&obj_start) || !fname.ends_with(".o")) {
- return true;
++ fn add_archive<F>(&mut self, archive_path: &Path, mut skip: F) -> std::io::Result<()>
++ where
++ F: FnMut(&str) -> bool + 'static,
++ {
++ let mut archive = ar::Archive::new(std::fs::File::open(&archive_path)?);
++ let archive_index = self.src_archives.len();
+
- // ok, don't skip this
- return false;
- })
++ let mut i = 0;
++ while let Some(entry) = archive.next_entry() {
++ let entry = entry?;
++ let file_name = String::from_utf8(entry.header().identifier().to_vec())
++ .map_err(|err| std::io::Error::new(std::io::ErrorKind::InvalidData, err))?;
++ if !skip(&file_name) {
++ self.entries
++ .push((file_name, ArchiveEntry::FromArchive { archive_index, entry_index: i }));
+ }
++ i += 1;
++ }
+
-
- impl<'a> ArArchiveBuilder<'a> {
- fn add_archive<F>(&mut self, archive_path: PathBuf, mut skip: F) -> std::io::Result<()>
- where
- F: FnMut(&str) -> bool + 'static,
- {
- let mut archive = ar::Archive::new(std::fs::File::open(&archive_path)?);
- let archive_index = self.src_archives.len();
-
- let mut i = 0;
- while let Some(entry) = archive.next_entry() {
- let entry = entry.unwrap();
- let file_name = String::from_utf8(entry.header().identifier().to_vec()).unwrap();
- if !skip(&file_name) {
- self.entries.push((
- file_name,
- ArchiveEntry::FromArchive {
- archive_index,
- entry_index: i,
- },
- ));
- }
- i += 1;
- }
-
- self.src_archives.push((archive_path, archive));
- Ok(())
- }
- }
++ self.src_archives.push((archive_path.to_owned(), archive));
++ Ok(())
+ }
+
+ fn update_symbols(&mut self) {
+ }
+
+ fn build(mut self) {
+ use std::process::Command;
+
+ fn add_file_using_ar(archive: &Path, file: &Path) {
+ Command::new("ar")
+ .arg("r") // add or replace file
+ .arg("-c") // silence created file message
+ .arg(archive)
+ .arg(&file)
+ .status()
+ .unwrap();
+ }
+
+ enum BuilderKind<'a> {
+ Bsd(ar::Builder<File>),
+ Gnu(ar::GnuBuilder<File>),
+ NativeAr(&'a Path),
+ }
+
+ let mut builder = if self.config.use_native_ar {
+ BuilderKind::NativeAr(&self.config.dst)
+ } else if self.config.use_gnu_style_archive {
+ BuilderKind::Gnu(ar::GnuBuilder::new(
+ File::create(&self.config.dst).unwrap(),
+ self.entries
+ .iter()
+ .map(|(name, _)| name.as_bytes().to_vec())
+ .collect(),
+ ))
+ } else {
+ BuilderKind::Bsd(ar::Builder::new(File::create(&self.config.dst).unwrap()))
+ };
+
+ // Add all files
+ for (entry_name, entry) in self.entries.into_iter() {
+ match entry {
+ ArchiveEntry::FromArchive {
+ archive_index,
+ entry_index,
+ } => {
+ let (ref src_archive_path, ref mut src_archive) =
+ self.src_archives[archive_index];
+ let entry = src_archive.jump_to_entry(entry_index).unwrap();
+ let header = entry.header().clone();
+
+ match builder {
+ BuilderKind::Bsd(ref mut builder) => {
+ builder.append(&header, entry).unwrap()
+ }
+ BuilderKind::Gnu(ref mut builder) => {
+ builder.append(&header, entry).unwrap()
+ }
+ BuilderKind::NativeAr(archive_file) => {
+ Command::new("ar")
+ .arg("x")
+ .arg(src_archive_path)
+ .arg(&entry_name)
+ .status()
+ .unwrap();
+ add_file_using_ar(archive_file, Path::new(&entry_name));
+ std::fs::remove_file(entry_name).unwrap();
+ }
+ }
+ }
+ ArchiveEntry::File(file) =>
+ match builder {
+ BuilderKind::Bsd(ref mut builder) => {
+ builder
+ .append_file(entry_name.as_bytes(), &mut File::open(file).expect("file for bsd builder"))
+ .unwrap()
+ },
+ BuilderKind::Gnu(ref mut builder) => {
+ builder
+ .append_file(entry_name.as_bytes(), &mut File::open(&file).expect(&format!("file {:?} for gnu builder", file)))
+ .unwrap()
+ },
+ BuilderKind::NativeAr(archive_file) => add_file_using_ar(archive_file, &file),
+ },
+ }
+ }
+
+ // Finalize archive
+ std::mem::drop(builder);
+
+ // Run ranlib to be able to link the archive
+ let status = std::process::Command::new("ranlib")
+ .arg(self.config.dst)
+ .status()
+ .expect("Couldn't run ranlib");
+
+ if !status.success() {
+ self.config.sess.fatal(&format!("Ranlib exited with code {:?}", status.code()));
+ }
+ }
+
+ fn inject_dll_import_lib(&mut self, _lib_name: &str, _dll_imports: &[DllImport], _tmpdir: &MaybeTempDir) {
+ unimplemented!();
+ }
+}
--- /dev/null
- use gccjit::{RValue, ToRValue, Type};
++use gccjit::{LValue, RValue, ToRValue, Type};
+use rustc_ast::ast::{InlineAsmOptions, InlineAsmTemplatePiece};
+use rustc_codegen_ssa::mir::operand::OperandValue;
+use rustc_codegen_ssa::mir::place::PlaceRef;
+use rustc_codegen_ssa::traits::{AsmBuilderMethods, AsmMethods, BaseTypeMethods, BuilderMethods, GlobalAsmOperandRef, InlineAsmOperandRef};
- use rustc_data_structures::fx::FxHashMap;
++
+use rustc_hir::LlvmInlineAsmInner;
- use rustc_middle::bug;
++use rustc_middle::{bug, ty::Instance};
+use rustc_span::Span;
+use rustc_target::asm::*;
+
++use std::borrow::Cow;
++
+use crate::builder::Builder;
+use crate::context::CodegenCx;
+use crate::type_of::LayoutGccExt;
+
++
++// Rust asm! and GCC Extended Asm semantics differ substantially.
++//
++// 1. Rust asm operands go along as one list of operands. Operands themselves indicate
++// if they're "in" or "out". "In" and "out" operands can interleave. One operand can be
++// both "in" and "out" (`inout(reg)`).
++//
++// GCC asm has two different lists for "in" and "out" operands. In terms of gccjit,
++// this means that all "out" operands must go before "in" operands. "In" and "out" operands
++// cannot interleave.
++//
++// 2. Operand lists in both Rust and GCC are indexed. Index starts from 0. Indexes are important
++// because the asm template refers to operands by index.
++//
++// Mapping from Rust to GCC index would be 1-1 if it wasn't for...
++//
++// 3. Clobbers. GCC has a separate list of clobbers, and clobbers don't have indexes.
++// Contrary, Rust expresses clobbers through "out" operands that aren't tied to
++// a variable (`_`), and such "clobbers" do have index.
++//
++// 4. Furthermore, GCC Extended Asm does not support explicit register constraints
++// (like `out("eax")`) directly, offering so-called "local register variables"
++// as a workaround. These variables need to be declared and initialized *before*
++// the Extended Asm block but *after* normal local variables
++// (see comment in `codegen_inline_asm` for explanation).
++//
++// With that in mind, let's see how we translate Rust syntax to GCC
++// (from now on, `CC` stands for "constraint code"):
++//
++// * `out(reg_class) var` -> translated to output operand: `"=CC"(var)`
++// * `inout(reg_class) var` -> translated to output operand: `"+CC"(var)`
++// * `in(reg_class) var` -> translated to input operand: `"CC"(var)`
++//
++// * `out(reg_class) _` -> translated to one `=r(tmp)`, where "tmp" is a temporary unused variable
++//
++// * `out("explicit register") _` -> not translated to any operands, register is simply added to clobbers list
++//
++// * `inout(reg_class) in_var => out_var` -> translated to two operands:
++// output: `"=CC"(in_var)`
++// input: `"num"(out_var)` where num is the GCC index
++// of the corresponding output operand
++//
++// * `inout(reg_class) in_var => _` -> same as `inout(reg_class) in_var => tmp`,
++// where "tmp" is a temporary unused variable
++//
++// * `out/in/inout("explicit register") var` -> translated to one or two operands as described above
++// with `"r"(var)` constraint,
++// and one register variable assigned to the desired register.
++//
++
++const ATT_SYNTAX_INS: &str = ".att_syntax noprefix\n\t";
++const INTEL_SYNTAX_INS: &str = "\n\t.intel_syntax noprefix";
++
++
++struct AsmOutOperand<'a, 'tcx, 'gcc> {
++ rust_idx: usize,
++ constraint: &'a str,
++ late: bool,
++ readwrite: bool,
++
++ tmp_var: LValue<'gcc>,
++ out_place: Option<PlaceRef<'tcx, RValue<'gcc>>>
++}
++
++struct AsmInOperand<'a, 'tcx> {
++ rust_idx: usize,
++ constraint: Cow<'a, str>,
++ val: RValue<'tcx>
++}
++
++impl AsmOutOperand<'_, '_, '_> {
++ fn to_constraint(&self) -> String {
++ let mut res = String::with_capacity(self.constraint.len() + self.late as usize + 1);
++
++ let sign = if self.readwrite { '+' } else { '=' };
++ res.push(sign);
++ if !self.late {
++ res.push('&');
++ }
++
++ res.push_str(&self.constraint);
++ res
++ }
++}
++
++enum ConstraintOrRegister {
++ Constraint(&'static str),
++ Register(&'static str)
++}
++
++
+impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> {
- fn codegen_llvm_inline_asm(&mut self, _ia: &LlvmInlineAsmInner, _outputs: Vec<PlaceRef<'tcx, RValue<'gcc>>>, mut _inputs: Vec<RValue<'gcc>>, _span: Span) -> bool {
- // TODO(antoyo)
- return true;
++ fn codegen_llvm_inline_asm(&mut self, _ia: &LlvmInlineAsmInner, _outputs: Vec<PlaceRef<'tcx, RValue<'gcc>>>, _inputs: Vec<RValue<'gcc>>, span: Span) -> bool {
++ self.sess().struct_span_err(span, "GCC backend does not support `llvm_asm!`")
++ .help("consider using the `asm!` macro instead")
++ .emit();
++
++ // We return `true` even if we've failed to generate the asm
++ // because we want to suppress the "malformed inline assembly" error
++ // generated by the frontend.
++ true
+ }
+
- fn codegen_inline_asm(&mut self, template: &[InlineAsmTemplatePiece], operands: &[InlineAsmOperandRef<'tcx, Self>], options: InlineAsmOptions, _span: &[Span]) {
++ fn codegen_inline_asm(&mut self, template: &[InlineAsmTemplatePiece], rust_operands: &[InlineAsmOperandRef<'tcx, Self>], options: InlineAsmOptions, _span: &[Span]) {
+ let asm_arch = self.tcx.sess.asm_arch.unwrap();
++ let is_x86 = matches!(asm_arch, InlineAsmArch::X86 | InlineAsmArch::X86_64);
++ let att_dialect = is_x86 && options.contains(InlineAsmOptions::ATT_SYNTAX);
++ let intel_dialect = is_x86 && !options.contains(InlineAsmOptions::ATT_SYNTAX);
+
- let intel_dialect =
- match asm_arch {
- InlineAsmArch::X86 | InlineAsmArch::X86_64 if !options.contains(InlineAsmOptions::ATT_SYNTAX) => true,
- _ => false,
- };
++ // GCC index of an output operand equals its position in the array
++ let mut outputs = vec![];
++
++ // GCC index of an input operand equals its position in the array
++ // added to `outputs.len()`
++ let mut inputs = vec![];
++
++ // Clobbers collected from `out("explicit register") _` and `inout("expl_reg") var => _`
++ let mut clobbers = vec![];
++
++ // We're trying to preallocate space for the template
++ let mut constants_len = 0;
+
- // Collect the types of output operands
- // FIXME(antoyo): we do this here instead of later because of a bug in libgccjit where creating the
- // variable after the extended asm expression causes a segfault:
- // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100380
- let mut output_vars = FxHashMap::default();
- let mut operand_numbers = FxHashMap::default();
- let mut current_number = 0;
- for (idx, op) in operands.iter().enumerate() {
++ // There are rules we must adhere to if we want GCC to do the right thing:
++ //
++ // * Every local variable that the asm block uses as an output must be declared *before*
++ // the asm block.
++ // * There must be no instructions whatsoever between the register variables and the asm.
++ //
++ // Therefore, the backend must generate the instructions strictly in this order:
++ //
++ // 1. Output variables.
++ // 2. Register variables.
++ // 3. The asm block.
++ //
++ // We also must make sure that no input operands are emitted before output operands.
++ //
++ // This is why we work in passes, first emitting local vars, then local register vars.
++ // Also, we don't emit any asm operands immediately; we save them to
++ // the one of the buffers to be emitted later.
++
++ // 1. Normal variables (and saving operands to buffers).
++ for (rust_idx, op) in rust_operands.iter().enumerate() {
+ match *op {
- InlineAsmOperandRef::Out { place, .. } => {
- let ty =
- match place {
- Some(place) => place.layout.gcc_type(self.cx, false),
- None => {
- // If the output is discarded, we don't really care what
- // type is used. We're just using this to tell GCC to
- // reserve the register.
- //dummy_output_type(self.cx, reg.reg_class())
-
- // NOTE: if no output value, we should not create one (it will be a
- // clobber).
- continue;
- },
- };
- let var = self.current_func().new_local(None, ty, "output_register");
- operand_numbers.insert(idx, current_number);
- current_number += 1;
- output_vars.insert(idx, var);
++ InlineAsmOperandRef::Out { reg, late, place } => {
++ use ConstraintOrRegister::*;
++
++ let (constraint, ty) = match (reg_to_gcc(reg), place) {
++ (Constraint(constraint), Some(place)) => (constraint, place.layout.gcc_type(self.cx, false)),
++ // When `reg` is a class and not an explicit register but the out place is not specified,
++ // we need to create an unused output variable to assign the output to. This var
++ // needs to be of a type that's "compatible" with the register class, but specific type
++ // doesn't matter.
++ (Constraint(constraint), None) => (constraint, dummy_output_type(self.cx, reg.reg_class())),
++ (Register(_), Some(_)) => {
++ // left for the next pass
++ continue
++ },
++ (Register(reg_name), None) => {
++ clobbers.push(reg_name);
++ continue
++ }
++ };
++
++ let tmp_var = self.current_func().new_local(None, ty, "output_register");
++ outputs.push(AsmOutOperand {
++ constraint,
++ rust_idx,
++ late,
++ readwrite: false,
++ tmp_var,
++ out_place: place
++ });
+ }
- InlineAsmOperandRef::InOut { out_place, .. } => {
- let ty =
- match out_place {
- Some(place) => place.layout.gcc_type(self.cx, false),
- None => {
- // NOTE: if no output value, we should not create one.
- continue;
- },
- };
- operand_numbers.insert(idx, current_number);
- current_number += 1;
- let var = self.current_func().new_local(None, ty, "output_register");
- output_vars.insert(idx, var);
++
++ InlineAsmOperandRef::In { reg, value } => {
++ if let ConstraintOrRegister::Constraint(constraint) = reg_to_gcc(reg) {
++ inputs.push(AsmInOperand {
++ constraint: Cow::Borrowed(constraint),
++ rust_idx,
++ val: value.immediate()
++ });
++ }
++ else {
++ // left for the next pass
++ continue
++ }
++ }
++
++ InlineAsmOperandRef::InOut { reg, late, in_value, out_place } => {
++ let constraint = if let ConstraintOrRegister::Constraint(constraint) = reg_to_gcc(reg) {
++ constraint
++ }
++ else {
++ // left for the next pass
++ continue
++ };
++
++ // Rustc frontend guarantees that input and output types are "compatible",
++ // so we can just use input var's type for the output variable.
++ //
++ // This decision is also backed by the fact that LLVM needs in and out
++ // values to be of *exactly the same type*, not just "compatible".
++ // I'm not sure if GCC is so picky too, but better safe than sorry.
++ let ty = in_value.layout.gcc_type(self.cx, false);
++ let tmp_var = self.current_func().new_local(None, ty, "output_register");
++
++ // If the out_place is None (i.e `inout(reg) _` syntax was used), we translate
++ // it to one "readwrite (+) output variable", otherwise we translate it to two
++ // "out and tied in" vars as described above.
++ let readwrite = out_place.is_none();
++ outputs.push(AsmOutOperand {
++ constraint,
++ rust_idx,
++ late,
++ readwrite,
++ tmp_var,
++ out_place,
++ });
++
++ if !readwrite {
++ let out_gcc_idx = outputs.len() - 1;
++ let constraint = Cow::Owned(out_gcc_idx.to_string());
++
++ inputs.push(AsmInOperand {
++ constraint,
++ rust_idx,
++ val: in_value.immediate()
++ });
++ }
++ }
++
++ InlineAsmOperandRef::Const { ref string } => {
++ constants_len += string.len() + att_dialect as usize;
++ }
++
++ InlineAsmOperandRef::SymFn { instance } => {
++ constants_len += self.tcx.symbol_name(instance).name.len();
++ }
++ InlineAsmOperandRef::SymStatic { def_id } => {
++ constants_len += self.tcx.symbol_name(Instance::mono(self.tcx, def_id)).name.len();
+ }
- _ => {}
+ }
+ }
+
- // All output operands must come before the input operands, hence the 2 loops.
- for (idx, op) in operands.iter().enumerate() {
++ // 2. Register variables.
++ for (rust_idx, op) in rust_operands.iter().enumerate() {
+ match *op {
- InlineAsmOperandRef::In { .. } | InlineAsmOperandRef::InOut { .. } => {
- operand_numbers.insert(idx, current_number);
- current_number += 1;
- },
- _ => (),
++ // `out("explicit register") var`
++ InlineAsmOperandRef::Out { reg, late, place } => {
++ if let ConstraintOrRegister::Register(reg_name) = reg_to_gcc(reg) {
++ let out_place = if let Some(place) = place {
++ place
++ }
++ else {
++ // processed in the previous pass
++ continue
++ };
++
++ let ty = out_place.layout.gcc_type(self.cx, false);
++ let tmp_var = self.current_func().new_local(None, ty, "output_register");
++ tmp_var.set_register_name(reg_name);
++
++ outputs.push(AsmOutOperand {
++ constraint: "r".into(),
++ rust_idx,
++ late,
++ readwrite: false,
++ tmp_var,
++ out_place: Some(out_place)
++ });
++ }
++
++ // processed in the previous pass
++ }
++
++ // `in("explicit register") var`
++ InlineAsmOperandRef::In { reg, value } => {
++ if let ConstraintOrRegister::Register(reg_name) = reg_to_gcc(reg) {
++ let ty = value.layout.gcc_type(self.cx, false);
++ let reg_var = self.current_func().new_local(None, ty, "input_register");
++ reg_var.set_register_name(reg_name);
++ self.llbb().add_assignment(None, reg_var, value.immediate());
++
++ inputs.push(AsmInOperand {
++ constraint: "r".into(),
++ rust_idx,
++ val: reg_var.to_rvalue()
++ });
++ }
++
++ // processed in the previous pass
++ }
++
++ // `inout("explicit register") in_var => out_var`
++ InlineAsmOperandRef::InOut { reg, late, in_value, out_place } => {
++ if let ConstraintOrRegister::Register(reg_name) = reg_to_gcc(reg) {
++ let out_place = if let Some(place) = out_place {
++ place
++ }
++ else {
++ // processed in the previous pass
++ continue
++ };
++
++ // See explanation in the first pass.
++ let ty = in_value.layout.gcc_type(self.cx, false);
++ let tmp_var = self.current_func().new_local(None, ty, "output_register");
++ tmp_var.set_register_name(reg_name);
++
++ outputs.push(AsmOutOperand {
++ constraint: "r".into(),
++ rust_idx,
++ late,
++ readwrite: false,
++ tmp_var,
++ out_place: Some(out_place)
++ });
++
++ let constraint = Cow::Owned((outputs.len() - 1).to_string());
++ inputs.push(AsmInOperand {
++ constraint,
++ rust_idx,
++ val: in_value.immediate()
++ });
++ }
++
++ // processed in the previous pass
++ }
++
++ InlineAsmOperandRef::Const { .. }
++ | InlineAsmOperandRef::SymFn { .. }
++ | InlineAsmOperandRef::SymStatic { .. } => {
++ // processed in the previous pass
++ }
+ }
+ }
+
- // Build the template string
- let mut template_str = String::new();
++ // 3. Build the template string
++
++ let mut template_str = String::with_capacity(estimate_template_length(template, constants_len, att_dialect));
++ if !intel_dialect {
++ template_str.push_str(ATT_SYNTAX_INS);
++ }
++
+ for piece in template {
+ match *piece {
+ InlineAsmTemplatePiece::String(ref string) => {
- if string.contains('%') {
- for c in string.chars() {
- if c == '%' {
- template_str.push_str("%%");
- }
- else {
- template_str.push(c);
- }
- }
++ // TODO(@Commeownist): switch to `Iterator::intersperse` once it's stable
++ let mut iter = string.split('%');
++ if let Some(s) = iter.next() {
++ template_str.push_str(s);
+ }
- else {
- template_str.push_str(string)
++
++ for s in iter {
++ template_str.push_str("%%");
++ template_str.push_str(s);
+ }
+ }
+ InlineAsmTemplatePiece::Placeholder { operand_idx, modifier, span: _ } => {
- match operands[operand_idx] {
- InlineAsmOperandRef::Out { reg, place: Some(_), .. } => {
++ let mut push_to_template = |modifier, gcc_idx| {
++ use std::fmt::Write;
++
++ template_str.push('%');
++ if let Some(modifier) = modifier {
++ template_str.push(modifier);
++ }
++ write!(template_str, "{}", gcc_idx).expect("pushing to string failed");
++ };
++
++ match rust_operands[operand_idx] {
++ InlineAsmOperandRef::Out { reg, .. } => {
+ let modifier = modifier_to_gcc(asm_arch, reg.reg_class(), modifier);
- if let Some(modifier) = modifier {
- template_str.push_str(&format!("%{}{}", modifier, operand_numbers[&operand_idx]));
- } else {
- template_str.push_str(&format!("%{}", operand_numbers[&operand_idx]));
- }
- },
- InlineAsmOperandRef::Out { place: None, .. } => {
- unimplemented!("Out None");
- },
- InlineAsmOperandRef::In { reg, .. }
- | InlineAsmOperandRef::InOut { reg, .. } => {
++ let gcc_index = outputs.iter()
++ .position(|op| operand_idx == op.rust_idx)
++ .expect("wrong rust index");
++ push_to_template(modifier, gcc_index);
++ }
++
++ InlineAsmOperandRef::In { reg, .. } => {
+ let modifier = modifier_to_gcc(asm_arch, reg.reg_class(), modifier);
- if let Some(modifier) = modifier {
- template_str.push_str(&format!("%{}{}", modifier, operand_numbers[&operand_idx]));
- } else {
- template_str.push_str(&format!("%{}", operand_numbers[&operand_idx]));
- }
++ let in_gcc_index = inputs.iter()
++ .position(|op| operand_idx == op.rust_idx)
++ .expect("wrong rust index");
++ let gcc_index = in_gcc_index + outputs.len();
++ push_to_template(modifier, gcc_index);
+ }
++
++ InlineAsmOperandRef::InOut { reg, .. } => {
++ let modifier = modifier_to_gcc(asm_arch, reg.reg_class(), modifier);
++
++ // The input register is tied to the output, so we can just use the index of the output register
++ let gcc_index = outputs.iter()
++ .position(|op| operand_idx == op.rust_idx)
++ .expect("wrong rust index");
++ push_to_template(modifier, gcc_index);
++ }
++
++ InlineAsmOperandRef::SymFn { instance } => {
++ let name = self.tcx.symbol_name(instance).name;
++ template_str.push_str(name);
++ }
++
++ InlineAsmOperandRef::SymStatic { def_id } => {
++ // TODO(@Commeownist): This may not be sufficient for all kinds of statics.
++ // Some statics may need the `@plt` suffix, like thread-local vars.
++ let instance = Instance::mono(self.tcx, def_id);
++ let name = self.tcx.symbol_name(instance).name;
++ template_str.push_str(name);
++ }
++
+ InlineAsmOperandRef::Const { ref string } => {
+ // Const operands get injected directly into the template
++ if att_dialect {
++ template_str.push('$');
++ }
+ template_str.push_str(string);
+ }
- InlineAsmOperandRef::SymFn { .. }
- | InlineAsmOperandRef::SymStatic { .. } => {
- unimplemented!();
- // Only emit the raw symbol name
- //template_str.push_str(&format!("${{{}:c}}", op_idx[&operand_idx]));
- }
+ }
+ }
+ }
+ }
+
++ if !intel_dialect {
++ template_str.push_str(INTEL_SYNTAX_INS);
++ }
++
++ // 4. Generate Extended Asm block
++
+ let block = self.llbb();
- let template_str =
- if intel_dialect {
- template_str
- }
- else {
- // FIXME(antoyo): this might break the "m" memory constraint:
- // https://stackoverflow.com/a/9347957/389119
- // TODO(antoyo): only set on x86 platforms.
- format!(".att_syntax noprefix\n\t{}\n\t.intel_syntax noprefix", template_str)
- };
+ let extended_asm = block.add_extended_asm(None, &template_str);
+
- // Collect the types of output operands
- let mut output_types = vec![];
- for (idx, op) in operands.iter().enumerate() {
- match *op {
- InlineAsmOperandRef::Out { reg, late, place } => {
- let ty =
- match place {
- Some(place) => place.layout.gcc_type(self.cx, false),
- None => {
- // If the output is discarded, we don't really care what
- // type is used. We're just using this to tell GCC to
- // reserve the register.
- dummy_output_type(self.cx, reg.reg_class())
- },
- };
- output_types.push(ty);
- let prefix = if late { "=" } else { "=&" };
- let constraint = format!("{}{}", prefix, reg_to_gcc(reg));
++ for op in &outputs {
++ extended_asm.add_output_operand(None, &op.to_constraint(), op.tmp_var);
++ }
+
- if place.is_some() {
- let var = output_vars[&idx];
- extended_asm.add_output_operand(None, &constraint, var);
- }
- else {
- // NOTE: reg.to_string() returns the register name with quotes around it so
- // remove them.
- extended_asm.add_clobber(reg.to_string().trim_matches('"'));
- }
- }
- InlineAsmOperandRef::InOut { reg, late, in_value, out_place } => {
- let ty =
- match out_place {
- Some(out_place) => out_place.layout.gcc_type(self.cx, false),
- None => dummy_output_type(self.cx, reg.reg_class())
- };
- output_types.push(ty);
- // TODO(antoyo): prefix of "+" for reading and writing?
- let prefix = if late { "=" } else { "=&" };
- let constraint = format!("{}{}", prefix, reg_to_gcc(reg));
-
- if out_place.is_some() {
- let var = output_vars[&idx];
- // TODO(antoyo): also specify an output operand when out_place is none: that would
- // be the clobber but clobbers do not support general constraint like reg;
- // they only support named registers.
- // Not sure how we can do this. And the LLVM backend does not seem to add a
- // clobber.
- extended_asm.add_output_operand(None, &constraint, var);
- }
++ for op in &inputs {
++ extended_asm.add_input_operand(None, &op.constraint, op.val);
++ }
+
- let constraint = reg_to_gcc(reg);
- extended_asm.add_input_operand(None, &constraint, in_value.immediate());
- }
- InlineAsmOperandRef::In { reg, value } => {
- let constraint = reg_to_gcc(reg);
- extended_asm.add_input_operand(None, &constraint, value.immediate());
- }
- _ => {}
++ for clobber in clobbers.iter() {
++ extended_asm.add_clobber(clobber);
++ }
++
++ if !options.contains(InlineAsmOptions::PRESERVES_FLAGS) {
++ // TODO(@Commeownist): I'm not 100% sure this one clobber is sufficient
++ // on all architectures. For instance, what about FP stack?
++ extended_asm.add_clobber("cc");
++ }
++ if !options.contains(InlineAsmOptions::NOMEM) {
++ extended_asm.add_clobber("memory");
++ }
++ if !options.contains(InlineAsmOptions::PURE) {
++ extended_asm.set_volatile_flag(true);
++ }
++ if !options.contains(InlineAsmOptions::NOSTACK) {
++ // TODO(@Commeownist): figure out how to align stack
++ }
++ if options.contains(InlineAsmOptions::NORETURN) {
++ let builtin_unreachable = self.context.get_builtin_function("__builtin_unreachable");
++ let builtin_unreachable: RValue<'gcc> = unsafe { std::mem::transmute(builtin_unreachable) };
++ self.call(self.type_void(), builtin_unreachable, &[], None);
++ }
++
++ // Write results to outputs.
++ //
++ // We need to do this because:
++ // 1. Turning `PlaceRef` into `RValue` is error-prone and has nasty edge cases
++ // (especially with current `rustc_backend_ssa` API).
++ // 2. Not every output operand has an `out_place`, and it's required by `add_output_operand`.
++ //
++ // Instead, we generate a temporary output variable for each output operand, and then this loop,
++ // generates `out_place = tmp_var;` assignments if out_place exists.
++ for op in &outputs {
++ if let Some(place) = op.out_place {
++ OperandValue::Immediate(op.tmp_var.to_rvalue()).store(self, place);
+ }
+ }
+
- // Write results to outputs
- for (idx, op) in operands.iter().enumerate() {
- if let InlineAsmOperandRef::Out { place: Some(place), .. }
- | InlineAsmOperandRef::InOut { out_place: Some(place), .. } = *op
- {
- OperandValue::Immediate(output_vars[&idx].to_rvalue()).store(self, place);
++ }
++}
++
++fn estimate_template_length(template: &[InlineAsmTemplatePiece], constants_len: usize, att_dialect: bool) -> usize {
++ let len: usize = template.iter().map(|piece| {
++ match *piece {
++ InlineAsmTemplatePiece::String(ref string) => {
++ string.len()
++ }
++ InlineAsmTemplatePiece::Placeholder { .. } => {
++ // '%' + 1 char modifier + 1 char index
++ 3
+ }
+ }
++ })
++ .sum();
++
++ // increase it by 5% to account for possible '%' signs that'll be duplicated
++ // I pulled the number out of blue, but should be fair enough
++ // as the upper bound
++ let mut res = (len as f32 * 1.05) as usize + constants_len;
++
++ if att_dialect {
++ res += INTEL_SYNTAX_INS.len() + ATT_SYNTAX_INS.len();
+ }
++ res
+}
+
+/// Converts a register class to a GCC constraint code.
- // TODO(antoyo): return &'static str instead?
- fn reg_to_gcc(reg: InlineAsmRegOrRegClass) -> String {
- match reg {
++fn reg_to_gcc(reg: InlineAsmRegOrRegClass) -> ConstraintOrRegister {
++ let constraint = match reg {
+ // For vector registers LLVM wants the register name to match the type size.
+ InlineAsmRegOrRegClass::Reg(reg) => {
+ // TODO(antoyo): add support for vector register.
- let constraint =
- match reg.name() {
- "ax" => "a",
- "bx" => "b",
- "cx" => "c",
- "dx" => "d",
- "si" => "S",
- "di" => "D",
- // TODO(antoyo): for registers like r11, we have to create a register variable: https://stackoverflow.com/a/31774784/389119
- // TODO(antoyo): in this case though, it's a clobber, so it should work as r11.
- // Recent nightly supports clobber() syntax, so update to it. It does not seem
- // like it's implemented yet.
- name => name, // FIXME(antoyo): probably wrong.
- };
- constraint.to_string()
++ match reg.name() {
++ "ax" => "a",
++ "bx" => "b",
++ "cx" => "c",
++ "dx" => "d",
++ "si" => "S",
++ "di" => "D",
++ // For registers like r11, we have to create a register variable: https://stackoverflow.com/a/31774784/389119
++ name => return ConstraintOrRegister::Register(name),
++ }
+ },
+ InlineAsmRegOrRegClass::RegClass(reg) => match reg {
+ InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::preg) => unimplemented!(),
+ InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => unimplemented!(),
+ InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg) => unimplemented!(),
+ InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => unimplemented!(),
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg) => unimplemented!(),
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg_thumb) => unimplemented!(),
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8) => unimplemented!(),
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => unimplemented!(),
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg) => unimplemented!(),
+ InlineAsmRegClass::Bpf(_) => unimplemented!(),
+ InlineAsmRegClass::Hexagon(HexagonInlineAsmRegClass::reg) => unimplemented!(),
+ InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg) => unimplemented!(),
+ InlineAsmRegClass::Mips(MipsInlineAsmRegClass::freg) => unimplemented!(),
+ InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg16) => unimplemented!(),
+ InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg32) => unimplemented!(),
+ InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg64) => unimplemented!(),
+ InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg) => unimplemented!(),
+ InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg_nonzero) => unimplemented!(),
+ InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::freg) => unimplemented!(),
++ InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::cr)
++ | InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::xer) => {
++ unreachable!("clobber-only")
++ },
+ InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) => unimplemented!(),
+ InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => unimplemented!(),
+ InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::vreg) => unimplemented!(),
- InlineAsmRegClass::X86(X86InlineAsmRegClass::mmx_reg) => unimplemented!(),
+ InlineAsmRegClass::X86(X86InlineAsmRegClass::reg) => "r",
- InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => unimplemented!(),
- InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => unimplemented!(),
++ InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => "Q",
++ InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => "q",
+ InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg)
- | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg) => unimplemented!(),
- InlineAsmRegClass::X86(X86InlineAsmRegClass::x87_reg) => unimplemented!(),
- InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => unimplemented!(),
++ | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg) => "x",
++ InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => "v",
+ InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => unimplemented!(),
+ InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => unimplemented!(),
++ InlineAsmRegClass::X86(
++ X86InlineAsmRegClass::x87_reg | X86InlineAsmRegClass::mmx_reg,
++ ) => unreachable!("clobber-only"),
+ InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => {
+ bug!("GCC backend does not support SPIR-V")
+ }
++ InlineAsmRegClass::S390x(S390xInlineAsmRegClass::reg) => unimplemented!(),
++ InlineAsmRegClass::S390x(S390xInlineAsmRegClass::freg) => unimplemented!(),
+ InlineAsmRegClass::Err => unreachable!(),
+ }
- .to_string(),
- }
++ };
++
++ ConstraintOrRegister::Constraint(constraint)
+}
+
+/// Type to use for outputs that are discarded. It doesn't really matter what
+/// the type is, as long as it is valid for the constraint code.
+fn dummy_output_type<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, reg: InlineAsmRegClass) -> Type<'gcc> {
+ match reg {
+ InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => cx.type_i32(),
+ InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::preg) => unimplemented!(),
+ InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg)
+ | InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => {
+ unimplemented!()
+ }
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg_thumb) => cx.type_i32(),
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16) => cx.type_f32(),
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8) => cx.type_f64(),
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => {
+ unimplemented!()
+ }
+ InlineAsmRegClass::Bpf(_) => unimplemented!(),
+ InlineAsmRegClass::Hexagon(HexagonInlineAsmRegClass::reg) => cx.type_i32(),
+ InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg) => cx.type_i32(),
+ InlineAsmRegClass::Mips(MipsInlineAsmRegClass::freg) => cx.type_f32(),
+ InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg16) => cx.type_i16(),
+ InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg32) => cx.type_i32(),
+ InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg64) => cx.type_i64(),
+ InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg) => cx.type_i32(),
+ InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg_nonzero) => cx.type_i32(),
+ InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::freg) => cx.type_f64(),
++ InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::cr)
++ | InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::xer) => {
++ unreachable!("clobber-only")
++ },
+ InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) => cx.type_i32(),
+ InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => cx.type_f32(),
+ InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::vreg) => cx.type_f32(),
+ InlineAsmRegClass::X86(X86InlineAsmRegClass::reg)
+ | InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => cx.type_i32(),
+ InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => cx.type_i8(),
+ InlineAsmRegClass::X86(X86InlineAsmRegClass::mmx_reg) => unimplemented!(),
+ InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg)
+ | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg)
+ | InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => cx.type_f32(),
+ InlineAsmRegClass::X86(X86InlineAsmRegClass::x87_reg) => unimplemented!(),
+ InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => cx.type_i16(),
+ InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => cx.type_i32(),
+ InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => {
+ bug!("LLVM backend does not support SPIR-V")
+ },
++ InlineAsmRegClass::S390x(S390xInlineAsmRegClass::reg) => cx.type_i32(),
++ InlineAsmRegClass::S390x(S390xInlineAsmRegClass::freg) => cx.type_f64(),
+ InlineAsmRegClass::Err => unreachable!(),
+ }
+}
+
+impl<'gcc, 'tcx> AsmMethods for CodegenCx<'gcc, 'tcx> {
+ fn codegen_global_asm(&self, template: &[InlineAsmTemplatePiece], operands: &[GlobalAsmOperandRef], options: InlineAsmOptions, _line_spans: &[Span]) {
+ let asm_arch = self.tcx.sess.asm_arch.unwrap();
+
+ // Default to Intel syntax on x86
+ let intel_syntax = matches!(asm_arch, InlineAsmArch::X86 | InlineAsmArch::X86_64)
+ && !options.contains(InlineAsmOptions::ATT_SYNTAX);
+
+ // Build the template string
+ let mut template_str = String::new();
+ for piece in template {
+ match *piece {
+ InlineAsmTemplatePiece::String(ref string) => {
+ for line in string.lines() {
+ // NOTE: gcc does not allow inline comment, so remove them.
+ let line =
+ if let Some(index) = line.rfind("//") {
+ &line[..index]
+ }
+ else {
+ line
+ };
+ template_str.push_str(line);
+ template_str.push('\n');
+ }
+ },
+ InlineAsmTemplatePiece::Placeholder { operand_idx, modifier: _, span: _ } => {
+ match operands[operand_idx] {
+ GlobalAsmOperandRef::Const { ref string } => {
+ // Const operands get injected directly into the
- // template. Note that we don't need to escape $
++ // template. Note that we don't need to escape %
+ // here unlike normal inline assembly.
+ template_str.push_str(string);
+ }
+ }
+ }
+ }
+ }
+
+ let template_str =
+ if intel_syntax {
+ format!("{}\n\t.intel_syntax noprefix", template_str)
+ }
+ else {
+ format!(".att_syntax\n\t{}\n\t.intel_syntax noprefix", template_str)
+ };
+ // NOTE: seems like gcc will put the asm in the wrong section, so set it to .text manually.
+ let template_str = format!(".pushsection .text\n{}\n.popsection", template_str);
+ self.context.add_top_level_asm(None, &template_str);
+ }
+}
+
+fn modifier_to_gcc(arch: InlineAsmArch, reg: InlineAsmRegClass, modifier: Option<char>) -> Option<char> {
+ match reg {
+ InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => modifier,
+ InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::preg) => modifier,
+ InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg)
+ | InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => {
+ unimplemented!()
+ }
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg_thumb) => unimplemented!(),
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16) => unimplemented!(),
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8) => unimplemented!(),
+ InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8)
+ | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => {
+ unimplemented!()
+ }
+ InlineAsmRegClass::Bpf(_) => unimplemented!(),
+ InlineAsmRegClass::Hexagon(_) => unimplemented!(),
+ InlineAsmRegClass::Mips(_) => unimplemented!(),
+ InlineAsmRegClass::Nvptx(_) => unimplemented!(),
+ InlineAsmRegClass::PowerPC(_) => unimplemented!(),
+ InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg)
+ | InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => unimplemented!(),
+ InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::vreg) => unimplemented!(),
+ InlineAsmRegClass::X86(X86InlineAsmRegClass::reg)
+ | InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => match modifier {
- None if arch == InlineAsmArch::X86_64 => Some('q'),
- None => Some('k'),
++ None => if arch == InlineAsmArch::X86_64 { Some('q') } else { Some('k') },
+ Some('l') => Some('b'),
+ Some('h') => Some('h'),
+ Some('x') => Some('w'),
+ Some('e') => Some('k'),
+ Some('r') => Some('q'),
+ _ => unreachable!(),
+ },
- InlineAsmRegClass::X86(X86InlineAsmRegClass::mmx_reg) => unimplemented!(),
- InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => unimplemented!(),
- InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg)
- | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg)
- | InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => unimplemented!(),
- InlineAsmRegClass::X86(X86InlineAsmRegClass::x87_reg) => unimplemented!(),
- InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => unimplemented!(),
++ InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => None,
++ InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::xmm_reg)
++ | InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::ymm_reg)
++ | InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::zmm_reg) => match (reg, modifier) {
++ (X86InlineAsmRegClass::xmm_reg, None) => Some('x'),
++ (X86InlineAsmRegClass::ymm_reg, None) => Some('t'),
++ (X86InlineAsmRegClass::zmm_reg, None) => Some('g'),
++ (_, Some('x')) => Some('x'),
++ (_, Some('y')) => Some('t'),
++ (_, Some('z')) => Some('g'),
++ _ => unreachable!(),
++ },
++ InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => None,
++ InlineAsmRegClass::X86(X86InlineAsmRegClass::x87_reg | X86InlineAsmRegClass::mmx_reg) => {
++ unreachable!("clobber-only")
++ }
+ InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => unimplemented!(),
+ InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => {
+ bug!("LLVM backend does not support SPIR-V")
+ },
++ InlineAsmRegClass::S390x(S390xInlineAsmRegClass::reg) => unimplemented!(),
++ InlineAsmRegClass::S390x(S390xInlineAsmRegClass::freg) => unimplemented!(),
+ InlineAsmRegClass::Err => unreachable!(),
+ }
+}
--- /dev/null
- use std::ops::{Deref, Range};
+use std::borrow::Cow;
+use std::cell::Cell;
+use std::convert::TryFrom;
- use rustc_middle::ty::layout::{HasParamEnv, HasTyCtxt, TyAndLayout};
++use std::ops::Deref;
+
+use gccjit::FunctionType;
+use gccjit::{
+ BinaryOp,
+ Block,
+ ComparisonOp,
+ Function,
+ LValue,
+ RValue,
+ ToRValue,
+ Type,
+ UnaryOp,
+};
+use rustc_codegen_ssa::MemFlags;
+use rustc_codegen_ssa::common::{AtomicOrdering, AtomicRmwBinOp, IntPredicate, RealPredicate, SynchronizationScope};
+use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
+use rustc_codegen_ssa::mir::place::PlaceRef;
+use rustc_codegen_ssa::traits::{
+ BackendTypes,
+ BaseTypeMethods,
+ BuilderMethods,
+ ConstMethods,
+ DerivedTypeMethods,
+ LayoutTypeMethods,
+ HasCodegen,
+ OverflowOp,
+ StaticBuilderMethods,
+};
+use rustc_middle::ty::{ParamEnv, Ty, TyCtxt};
- LayoutOf,
++use rustc_middle::ty::layout::{HasParamEnv, HasTyCtxt, LayoutError, LayoutOfHelpers, TyAndLayout};
+use rustc_span::Span;
+use rustc_span::def_id::DefId;
+use rustc_target::abi::{
+ self,
+ Align,
+ HasDataLayout,
- impl<'tcx> LayoutOf for Builder<'_, '_, 'tcx> {
- type Ty = Ty<'tcx>;
- type TyAndLayout = TyAndLayout<'tcx>;
+ Size,
+ TargetDataLayout,
++ WrappingRange,
+};
+use rustc_target::spec::{HasTargetSpec, Target};
+
+use crate::common::{SignType, TypeReflection, type_is_pointer};
+use crate::context::CodegenCx;
+use crate::type_of::LayoutGccExt;
+
+// TODO(antoyo)
+type Funclet = ();
+
+// TODO(antoyo): remove this variable.
+static mut RETURN_VALUE_COUNT: usize = 0;
+
+enum ExtremumOperation {
+ Max,
+ Min,
+}
+
+trait EnumClone {
+ fn clone(&self) -> Self;
+}
+
+impl EnumClone for AtomicOrdering {
+ fn clone(&self) -> Self {
+ match *self {
+ AtomicOrdering::NotAtomic => AtomicOrdering::NotAtomic,
+ AtomicOrdering::Unordered => AtomicOrdering::Unordered,
+ AtomicOrdering::Monotonic => AtomicOrdering::Monotonic,
+ AtomicOrdering::Acquire => AtomicOrdering::Acquire,
+ AtomicOrdering::Release => AtomicOrdering::Release,
+ AtomicOrdering::AcquireRelease => AtomicOrdering::AcquireRelease,
+ AtomicOrdering::SequentiallyConsistent => AtomicOrdering::SequentiallyConsistent,
+ }
+ }
+}
+
+pub struct Builder<'a: 'gcc, 'gcc, 'tcx> {
+ pub cx: &'a CodegenCx<'gcc, 'tcx>,
+ pub block: Option<Block<'gcc>>,
+ stack_var_count: Cell<usize>,
+}
+
+impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
+ fn with_cx(cx: &'a CodegenCx<'gcc, 'tcx>) -> Self {
+ Builder {
+ cx,
+ block: None,
+ stack_var_count: Cell::new(0),
+ }
+ }
+
+ fn atomic_extremum(&mut self, operation: ExtremumOperation, dst: RValue<'gcc>, src: RValue<'gcc>, order: AtomicOrdering) -> RValue<'gcc> {
+ let size = self.cx.int_width(src.get_type()) / 8;
+
+ let func = self.current_func();
+
+ let load_ordering =
+ match order {
+ // TODO(antoyo): does this make sense?
+ AtomicOrdering::AcquireRelease | AtomicOrdering::Release => AtomicOrdering::Acquire,
+ _ => order.clone(),
+ };
+ let previous_value = self.atomic_load(dst.get_type(), dst, load_ordering.clone(), Size::from_bytes(size));
+ let previous_var = func.new_local(None, previous_value.get_type(), "previous_value");
+ let return_value = func.new_local(None, previous_value.get_type(), "return_value");
+ self.llbb().add_assignment(None, previous_var, previous_value);
+ self.llbb().add_assignment(None, return_value, previous_var.to_rvalue());
+
+ let while_block = func.new_block("while");
+ let after_block = func.new_block("after_while");
+ self.llbb().end_with_jump(None, while_block);
+
+ // NOTE: since jumps were added and compare_exchange doesn't expect this, the current blocks in the
+ // state need to be updated.
+ self.block = Some(while_block);
+ *self.cx.current_block.borrow_mut() = Some(while_block);
+
+ let comparison_operator =
+ match operation {
+ ExtremumOperation::Max => ComparisonOp::LessThan,
+ ExtremumOperation::Min => ComparisonOp::GreaterThan,
+ };
+
+ let cond1 = self.context.new_comparison(None, comparison_operator, previous_var.to_rvalue(), self.context.new_cast(None, src, previous_value.get_type()));
+ let compare_exchange = self.compare_exchange(dst, previous_var, src, order, load_ordering, false);
+ let cond2 = self.cx.context.new_unary_op(None, UnaryOp::LogicalNegate, compare_exchange.get_type(), compare_exchange);
+ let cond = self.cx.context.new_binary_op(None, BinaryOp::LogicalAnd, self.cx.bool_type, cond1, cond2);
+
+ while_block.end_with_conditional(None, cond, while_block, after_block);
+
+ // NOTE: since jumps were added in a place rustc does not expect, the current blocks in the
+ // state need to be updated.
+ self.block = Some(after_block);
+ *self.cx.current_block.borrow_mut() = Some(after_block);
+
+ return_value.to_rvalue()
+ }
+
+ fn compare_exchange(&self, dst: RValue<'gcc>, cmp: LValue<'gcc>, src: RValue<'gcc>, order: AtomicOrdering, failure_order: AtomicOrdering, weak: bool) -> RValue<'gcc> {
+ let size = self.cx.int_width(src.get_type());
+ let compare_exchange = self.context.get_builtin_function(&format!("__atomic_compare_exchange_{}", size / 8));
+ let order = self.context.new_rvalue_from_int(self.i32_type, order.to_gcc());
+ let failure_order = self.context.new_rvalue_from_int(self.i32_type, failure_order.to_gcc());
+ let weak = self.context.new_rvalue_from_int(self.bool_type, weak as i32);
+
+ let void_ptr_type = self.context.new_type::<*mut ()>();
+ let volatile_void_ptr_type = void_ptr_type.make_volatile();
+ let dst = self.context.new_cast(None, dst, volatile_void_ptr_type);
+ let expected = self.context.new_cast(None, cmp.get_address(None), void_ptr_type);
+
+ // NOTE: not sure why, but we have the wrong type here.
+ let int_type = compare_exchange.get_param(2).to_rvalue().get_type();
+ let src = self.context.new_cast(None, src, int_type);
+ self.context.new_call(None, compare_exchange, &[dst, expected, src, weak, order, failure_order])
+ }
+
+ pub fn assign(&self, lvalue: LValue<'gcc>, value: RValue<'gcc>) {
+ self.llbb().add_assignment(None, lvalue, value);
+ }
+
+ fn check_call<'b>(&mut self, _typ: &str, func: Function<'gcc>, args: &'b [RValue<'gcc>]) -> Cow<'b, [RValue<'gcc>]> {
+ let mut all_args_match = true;
+ let mut param_types = vec![];
+ let param_count = func.get_param_count();
+ for (index, arg) in args.iter().enumerate().take(param_count) {
+ let param = func.get_param(index as i32);
+ let param = param.to_rvalue().get_type();
+ if param != arg.get_type() {
+ all_args_match = false;
+ }
+ param_types.push(param);
+ }
+
+ if all_args_match {
+ return Cow::Borrowed(args);
+ }
+
+ let casted_args: Vec<_> = param_types
+ .into_iter()
+ .zip(args.iter())
+ .enumerate()
+ .map(|(_i, (expected_ty, &actual_val))| {
+ let actual_ty = actual_val.get_type();
+ if expected_ty != actual_ty {
+ self.bitcast(actual_val, expected_ty)
+ }
+ else {
+ actual_val
+ }
+ })
+ .collect();
+
+ Cow::Owned(casted_args)
+ }
+
+ fn check_ptr_call<'b>(&mut self, _typ: &str, func_ptr: RValue<'gcc>, args: &'b [RValue<'gcc>]) -> Cow<'b, [RValue<'gcc>]> {
+ let mut all_args_match = true;
+ let mut param_types = vec![];
+ let gcc_func = func_ptr.get_type().is_function_ptr_type().expect("function ptr");
+ for (index, arg) in args.iter().enumerate().take(gcc_func.get_param_count()) {
+ let param = gcc_func.get_param_type(index);
+ if param != arg.get_type() {
+ all_args_match = false;
+ }
+ param_types.push(param);
+ }
+
+ if all_args_match {
+ return Cow::Borrowed(args);
+ }
+
+ let casted_args: Vec<_> = param_types
+ .into_iter()
+ .zip(args.iter())
+ .enumerate()
+ .map(|(_i, (expected_ty, &actual_val))| {
+ let actual_ty = actual_val.get_type();
+ if expected_ty != actual_ty {
+ self.bitcast(actual_val, expected_ty)
+ }
+ else {
+ actual_val
+ }
+ })
+ .collect();
+
+ Cow::Owned(casted_args)
+ }
+
+ fn check_store(&mut self, val: RValue<'gcc>, ptr: RValue<'gcc>) -> RValue<'gcc> {
+ let dest_ptr_ty = self.cx.val_ty(ptr).make_pointer(); // TODO(antoyo): make sure make_pointer() is okay here.
+ let stored_ty = self.cx.val_ty(val);
+ let stored_ptr_ty = self.cx.type_ptr_to(stored_ty);
+
+ if dest_ptr_ty == stored_ptr_ty {
+ ptr
+ }
+ else {
+ self.bitcast(ptr, stored_ptr_ty)
+ }
+ }
+
+ pub fn current_func(&self) -> Function<'gcc> {
+ self.block.expect("block").get_function()
+ }
+
+ fn function_call(&mut self, func: RValue<'gcc>, args: &[RValue<'gcc>], _funclet: Option<&Funclet>) -> RValue<'gcc> {
+ // TODO(antoyo): remove when the API supports a different type for functions.
+ let func: Function<'gcc> = self.cx.rvalue_as_function(func);
+ let args = self.check_call("call", func, args);
+
+ // gccjit requires to use the result of functions, even when it's not used.
+ // That's why we assign the result to a local or call add_eval().
+ let return_type = func.get_return_type();
+ let current_block = self.current_block.borrow().expect("block");
+ let void_type = self.context.new_type::<()>();
+ let current_func = current_block.get_function();
+ if return_type != void_type {
+ unsafe { RETURN_VALUE_COUNT += 1 };
+ let result = current_func.new_local(None, return_type, &format!("returnValue{}", unsafe { RETURN_VALUE_COUNT }));
+ current_block.add_assignment(None, result, self.cx.context.new_call(None, func, &args));
+ result.to_rvalue()
+ }
+ else {
+ current_block.add_eval(None, self.cx.context.new_call(None, func, &args));
+ // Return dummy value when not having return value.
+ self.context.new_rvalue_from_long(self.isize_type, 0)
+ }
+ }
+
+ fn function_ptr_call(&mut self, func_ptr: RValue<'gcc>, args: &[RValue<'gcc>], _funclet: Option<&Funclet>) -> RValue<'gcc> {
+ let args = self.check_ptr_call("call", func_ptr, args);
+
+ // gccjit requires to use the result of functions, even when it's not used.
+ // That's why we assign the result to a local or call add_eval().
+ let gcc_func = func_ptr.get_type().is_function_ptr_type().expect("function ptr");
+ let mut return_type = gcc_func.get_return_type();
+ let current_block = self.current_block.borrow().expect("block");
+ let void_type = self.context.new_type::<()>();
+ let current_func = current_block.get_function();
+
+ // FIXME(antoyo): As a temporary workaround for unsupported LLVM intrinsics.
+ if gcc_func.get_param_count() == 0 && format!("{:?}", func_ptr) == "__builtin_ia32_pmovmskb128" {
+ return_type = self.int_type;
+ }
+
+ if return_type != void_type {
+ unsafe { RETURN_VALUE_COUNT += 1 };
+ let result = current_func.new_local(None, return_type, &format!("returnValue{}", unsafe { RETURN_VALUE_COUNT }));
+ current_block.add_assignment(None, result, self.cx.context.new_call_through_ptr(None, func_ptr, &args));
+ result.to_rvalue()
+ }
+ else {
+ if gcc_func.get_param_count() == 0 {
+ // FIXME(antoyo): As a temporary workaround for unsupported LLVM intrinsics.
+ current_block.add_eval(None, self.cx.context.new_call_through_ptr(None, func_ptr, &[]));
+ }
+ else {
+ current_block.add_eval(None, self.cx.context.new_call_through_ptr(None, func_ptr, &args));
+ }
+ // Return dummy value when not having return value.
+ let result = current_func.new_local(None, self.isize_type, "dummyValueThatShouldNeverBeUsed");
+ current_block.add_assignment(None, result, self.context.new_rvalue_from_long(self.isize_type, 0));
+ result.to_rvalue()
+ }
+ }
+
+ pub fn overflow_call(&mut self, func: Function<'gcc>, args: &[RValue<'gcc>], _funclet: Option<&Funclet>) -> RValue<'gcc> {
+ // gccjit requires to use the result of functions, even when it's not used.
+ // That's why we assign the result to a local.
+ let return_type = self.context.new_type::<bool>();
+ let current_block = self.current_block.borrow().expect("block");
+ let current_func = current_block.get_function();
+ // TODO(antoyo): return the new_call() directly? Since the overflow function has no side-effects.
+ unsafe { RETURN_VALUE_COUNT += 1 };
+ let result = current_func.new_local(None, return_type, &format!("returnValue{}", unsafe { RETURN_VALUE_COUNT }));
+ current_block.add_assignment(None, result, self.cx.context.new_call(None, func, &args));
+ result.to_rvalue()
+ }
+}
+
+impl<'gcc, 'tcx> HasCodegen<'tcx> for Builder<'_, 'gcc, 'tcx> {
+ type CodegenCx = CodegenCx<'gcc, 'tcx>;
+}
+
+impl<'tcx> HasTyCtxt<'tcx> for Builder<'_, '_, 'tcx> {
+ fn tcx(&self) -> TyCtxt<'tcx> {
+ self.cx.tcx()
+ }
+}
+
+impl HasDataLayout for Builder<'_, '_, '_> {
+ fn data_layout(&self) -> &TargetDataLayout {
+ self.cx.data_layout()
+ }
+}
+
- fn layout_of(&self, ty: Ty<'tcx>) -> Self::TyAndLayout {
- self.cx.layout_of(ty)
++impl<'tcx> LayoutOfHelpers<'tcx> for Builder<'_, '_, 'tcx> {
++ type LayoutOfResult = TyAndLayout<'tcx>;
+
- let range = scalar.valid_range_exclusive(bx);
- if range.start != range.end {
- bx.range_metadata(load, range);
++ #[inline]
++ fn handle_layout_err(&self, err: LayoutError<'tcx>, span: Span, ty: Ty<'tcx>) -> ! {
++ self.cx.handle_layout_err(err, span, ty)
+ }
+}
+
+impl<'gcc, 'tcx> Deref for Builder<'_, 'gcc, 'tcx> {
+ type Target = CodegenCx<'gcc, 'tcx>;
+
+ fn deref(&self) -> &Self::Target {
+ self.cx
+ }
+}
+
+impl<'gcc, 'tcx> BackendTypes for Builder<'_, 'gcc, 'tcx> {
+ type Value = <CodegenCx<'gcc, 'tcx> as BackendTypes>::Value;
+ type Function = <CodegenCx<'gcc, 'tcx> as BackendTypes>::Function;
+ type BasicBlock = <CodegenCx<'gcc, 'tcx> as BackendTypes>::BasicBlock;
+ type Type = <CodegenCx<'gcc, 'tcx> as BackendTypes>::Type;
+ type Funclet = <CodegenCx<'gcc, 'tcx> as BackendTypes>::Funclet;
+
+ type DIScope = <CodegenCx<'gcc, 'tcx> as BackendTypes>::DIScope;
+ type DILocation = <CodegenCx<'gcc, 'tcx> as BackendTypes>::DILocation;
+ type DIVariable = <CodegenCx<'gcc, 'tcx> as BackendTypes>::DIVariable;
+}
+
+impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
+ fn build(cx: &'a CodegenCx<'gcc, 'tcx>, block: Block<'gcc>) -> Self {
+ let mut bx = Builder::with_cx(cx);
+ *cx.current_block.borrow_mut() = Some(block);
+ bx.block = Some(block);
+ bx
+ }
+
+ fn build_sibling_block(&mut self, name: &str) -> Self {
+ let block = self.append_sibling_block(name);
+ Self::build(self.cx, block)
+ }
+
+ fn llbb(&self) -> Block<'gcc> {
+ self.block.expect("block")
+ }
+
+ fn append_block(cx: &'a CodegenCx<'gcc, 'tcx>, func: RValue<'gcc>, name: &str) -> Block<'gcc> {
+ let func = cx.rvalue_as_function(func);
+ func.new_block(name)
+ }
+
+ fn append_sibling_block(&mut self, name: &str) -> Block<'gcc> {
+ let func = self.current_func();
+ func.new_block(name)
+ }
+
+ fn ret_void(&mut self) {
+ self.llbb().end_with_void_return(None)
+ }
+
+ fn ret(&mut self, value: RValue<'gcc>) {
+ let value =
+ if self.structs_as_pointer.borrow().contains(&value) {
+ // NOTE: hack to workaround a limitation of the rustc API: see comment on
+ // CodegenCx.structs_as_pointer
+ value.dereference(None).to_rvalue()
+ }
+ else {
+ value
+ };
+ self.llbb().end_with_return(None, value);
+ }
+
+ fn br(&mut self, dest: Block<'gcc>) {
+ self.llbb().end_with_jump(None, dest)
+ }
+
+ fn cond_br(&mut self, cond: RValue<'gcc>, then_block: Block<'gcc>, else_block: Block<'gcc>) {
+ self.llbb().end_with_conditional(None, cond, then_block, else_block)
+ }
+
+ fn switch(&mut self, value: RValue<'gcc>, default_block: Block<'gcc>, cases: impl ExactSizeIterator<Item = (u128, Block<'gcc>)>) {
+ let mut gcc_cases = vec![];
+ let typ = self.val_ty(value);
+ for (on_val, dest) in cases {
+ let on_val = self.const_uint_big(typ, on_val);
+ gcc_cases.push(self.context.new_case(on_val, on_val, dest));
+ }
+ self.block.expect("block").end_with_switch(None, value, default_block, &gcc_cases);
+ }
+
+ fn invoke(&mut self, _typ: Type<'gcc>, _func: RValue<'gcc>, _args: &[RValue<'gcc>], then: Block<'gcc>, catch: Block<'gcc>, _funclet: Option<&Funclet>) -> RValue<'gcc> {
+ let condition = self.context.new_rvalue_from_int(self.bool_type, 0);
+ self.llbb().end_with_conditional(None, condition, then, catch);
+ self.context.new_rvalue_from_int(self.int_type, 0)
+
+ // TODO(antoyo)
+ }
+
+ fn unreachable(&mut self) {
+ let func = self.context.get_builtin_function("__builtin_unreachable");
+ let block = self.block.expect("block");
+ block.add_eval(None, self.context.new_call(None, func, &[]));
+ let return_type = block.get_function().get_return_type();
+ let void_type = self.context.new_type::<()>();
+ if return_type == void_type {
+ block.end_with_void_return(None)
+ }
+ else {
+ let return_value = self.current_func()
+ .new_local(None, return_type, "unreachableReturn");
+ block.end_with_return(None, return_value)
+ }
+ }
+
+ fn add(&mut self, a: RValue<'gcc>, mut b: RValue<'gcc>) -> RValue<'gcc> {
+ // FIXME(antoyo): this should not be required.
+ if format!("{:?}", a.get_type()) != format!("{:?}", b.get_type()) {
+ b = self.context.new_cast(None, b, a.get_type());
+ }
+ a + b
+ }
+
+ fn fadd(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a + b
+ }
+
+ fn sub(&mut self, a: RValue<'gcc>, mut b: RValue<'gcc>) -> RValue<'gcc> {
+ if a.get_type() != b.get_type() {
+ b = self.context.new_cast(None, b, a.get_type());
+ }
+ a - b
+ }
+
+ fn fsub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a - b
+ }
+
+ fn mul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a * b
+ }
+
+ fn fmul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a * b
+ }
+
+ fn udiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): convert the arguments to unsigned?
+ a / b
+ }
+
+ fn exactudiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): convert the arguments to unsigned?
+ // TODO(antoyo): poison if not exact.
+ a / b
+ }
+
+ fn sdiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): convert the arguments to signed?
+ a / b
+ }
+
+ fn exactsdiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): posion if not exact.
+ // FIXME(antoyo): rustc_codegen_ssa::mir::intrinsic uses different types for a and b but they
+ // should be the same.
+ let typ = a.get_type().to_signed(self);
+ let a = self.context.new_cast(None, a, typ);
+ let b = self.context.new_cast(None, b, typ);
+ a / b
+ }
+
+ fn fdiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a / b
+ }
+
+ fn urem(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a % b
+ }
+
+ fn srem(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a % b
+ }
+
+ fn frem(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ if a.get_type() == self.cx.float_type {
+ let fmodf = self.context.get_builtin_function("fmodf");
+ // FIXME(antoyo): this seems to produce the wrong result.
+ return self.context.new_call(None, fmodf, &[a, b]);
+ }
+ assert_eq!(a.get_type(), self.cx.double_type);
+
+ let fmod = self.context.get_builtin_function("fmod");
+ return self.context.new_call(None, fmod, &[a, b]);
+ }
+
+ fn shl(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ // FIXME(antoyo): remove the casts when libgccjit can shift an unsigned number by an unsigned number.
+ let a_type = a.get_type();
+ let b_type = b.get_type();
+ if a_type.is_unsigned(self) && b_type.is_signed(self) {
+ let a = self.context.new_cast(None, a, b_type);
+ let result = a << b;
+ self.context.new_cast(None, result, a_type)
+ }
+ else if a_type.is_signed(self) && b_type.is_unsigned(self) {
+ let b = self.context.new_cast(None, b, a_type);
+ a << b
+ }
+ else {
+ a << b
+ }
+ }
+
+ fn lshr(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ // FIXME(antoyo): remove the casts when libgccjit can shift an unsigned number by an unsigned number.
+ // TODO(antoyo): cast to unsigned to do a logical shift if that does not work.
+ let a_type = a.get_type();
+ let b_type = b.get_type();
+ if a_type.is_unsigned(self) && b_type.is_signed(self) {
+ let a = self.context.new_cast(None, a, b_type);
+ let result = a >> b;
+ self.context.new_cast(None, result, a_type)
+ }
+ else if a_type.is_signed(self) && b_type.is_unsigned(self) {
+ let b = self.context.new_cast(None, b, a_type);
+ a >> b
+ }
+ else {
+ a >> b
+ }
+ }
+
+ fn ashr(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): check whether behavior is an arithmetic shift for >> .
+ // FIXME(antoyo): remove the casts when libgccjit can shift an unsigned number by an unsigned number.
+ let a_type = a.get_type();
+ let b_type = b.get_type();
+ if a_type.is_unsigned(self) && b_type.is_signed(self) {
+ let a = self.context.new_cast(None, a, b_type);
+ let result = a >> b;
+ self.context.new_cast(None, result, a_type)
+ }
+ else if a_type.is_signed(self) && b_type.is_unsigned(self) {
+ let b = self.context.new_cast(None, b, a_type);
+ a >> b
+ }
+ else {
+ a >> b
+ }
+ }
+
+ fn and(&mut self, a: RValue<'gcc>, mut b: RValue<'gcc>) -> RValue<'gcc> {
+ // FIXME(antoyo): hack by putting the result in a variable to workaround this bug:
+ // https://gcc.gnu.org/bugzilla//show_bug.cgi?id=95498
+ if a.get_type() != b.get_type() {
+ b = self.context.new_cast(None, b, a.get_type());
+ }
+ let res = self.current_func().new_local(None, b.get_type(), "andResult");
+ self.llbb().add_assignment(None, res, a & b);
+ res.to_rvalue()
+ }
+
+ fn or(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ // FIXME(antoyo): hack by putting the result in a variable to workaround this bug:
+ // https://gcc.gnu.org/bugzilla//show_bug.cgi?id=95498
+ let res = self.current_func().new_local(None, b.get_type(), "orResult");
+ self.llbb().add_assignment(None, res, a | b);
+ res.to_rvalue()
+ }
+
+ fn xor(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a ^ b
+ }
+
+ fn neg(&mut self, a: RValue<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): use new_unary_op()?
+ self.cx.context.new_rvalue_from_long(a.get_type(), 0) - a
+ }
+
+ fn fneg(&mut self, a: RValue<'gcc>) -> RValue<'gcc> {
+ self.cx.context.new_unary_op(None, UnaryOp::Minus, a.get_type(), a)
+ }
+
+ fn not(&mut self, a: RValue<'gcc>) -> RValue<'gcc> {
+ let operation =
+ if a.get_type().is_bool() {
+ UnaryOp::LogicalNegate
+ }
+ else {
+ UnaryOp::BitwiseNegate
+ };
+ self.cx.context.new_unary_op(None, operation, a.get_type(), a)
+ }
+
+ fn unchecked_sadd(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a + b
+ }
+
+ fn unchecked_uadd(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a + b
+ }
+
+ fn unchecked_ssub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a - b
+ }
+
+ fn unchecked_usub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): should generate poison value?
+ a - b
+ }
+
+ fn unchecked_smul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a * b
+ }
+
+ fn unchecked_umul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+ a * b
+ }
+
+ fn fadd_fast(&mut self, _lhs: RValue<'gcc>, _rhs: RValue<'gcc>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn fsub_fast(&mut self, _lhs: RValue<'gcc>, _rhs: RValue<'gcc>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn fmul_fast(&mut self, _lhs: RValue<'gcc>, _rhs: RValue<'gcc>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn fdiv_fast(&mut self, _lhs: RValue<'gcc>, _rhs: RValue<'gcc>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn frem_fast(&mut self, _lhs: RValue<'gcc>, _rhs: RValue<'gcc>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn checked_binop(&mut self, oop: OverflowOp, typ: Ty<'_>, lhs: Self::Value, rhs: Self::Value) -> (Self::Value, Self::Value) {
+ use rustc_middle::ty::{Int, IntTy::*, Uint, UintTy::*};
+
+ let new_kind =
+ match typ.kind() {
+ Int(t @ Isize) => Int(t.normalize(self.tcx.sess.target.pointer_width)),
+ Uint(t @ Usize) => Uint(t.normalize(self.tcx.sess.target.pointer_width)),
+ t @ (Uint(_) | Int(_)) => t.clone(),
+ _ => panic!("tried to get overflow intrinsic for op applied to non-int type"),
+ };
+
+ // TODO(antoyo): remove duplication with intrinsic?
+ let name =
+ match oop {
+ OverflowOp::Add =>
+ match new_kind {
+ Int(I8) => "__builtin_add_overflow",
+ Int(I16) => "__builtin_add_overflow",
+ Int(I32) => "__builtin_sadd_overflow",
+ Int(I64) => "__builtin_saddll_overflow",
+ Int(I128) => "__builtin_add_overflow",
+
+ Uint(U8) => "__builtin_add_overflow",
+ Uint(U16) => "__builtin_add_overflow",
+ Uint(U32) => "__builtin_uadd_overflow",
+ Uint(U64) => "__builtin_uaddll_overflow",
+ Uint(U128) => "__builtin_add_overflow",
+
+ _ => unreachable!(),
+ },
+ OverflowOp::Sub =>
+ match new_kind {
+ Int(I8) => "__builtin_sub_overflow",
+ Int(I16) => "__builtin_sub_overflow",
+ Int(I32) => "__builtin_ssub_overflow",
+ Int(I64) => "__builtin_ssubll_overflow",
+ Int(I128) => "__builtin_sub_overflow",
+
+ Uint(U8) => "__builtin_sub_overflow",
+ Uint(U16) => "__builtin_sub_overflow",
+ Uint(U32) => "__builtin_usub_overflow",
+ Uint(U64) => "__builtin_usubll_overflow",
+ Uint(U128) => "__builtin_sub_overflow",
+
+ _ => unreachable!(),
+ },
+ OverflowOp::Mul =>
+ match new_kind {
+ Int(I8) => "__builtin_mul_overflow",
+ Int(I16) => "__builtin_mul_overflow",
+ Int(I32) => "__builtin_smul_overflow",
+ Int(I64) => "__builtin_smulll_overflow",
+ Int(I128) => "__builtin_mul_overflow",
+
+ Uint(U8) => "__builtin_mul_overflow",
+ Uint(U16) => "__builtin_mul_overflow",
+ Uint(U32) => "__builtin_umul_overflow",
+ Uint(U64) => "__builtin_umulll_overflow",
+ Uint(U128) => "__builtin_mul_overflow",
+
+ _ => unreachable!(),
+ },
+ };
+
+ let intrinsic = self.context.get_builtin_function(&name);
+ let res = self.current_func()
+ // TODO(antoyo): is it correct to use rhs type instead of the parameter typ?
+ .new_local(None, rhs.get_type(), "binopResult")
+ .get_address(None);
+ let overflow = self.overflow_call(intrinsic, &[lhs, rhs, res], None);
+ (res.dereference(None).to_rvalue(), overflow)
+ }
+
+ fn alloca(&mut self, ty: Type<'gcc>, align: Align) -> RValue<'gcc> {
+ // FIXME(antoyo): this check that we don't call get_aligned() a second time on a type.
+ // Ideally, we shouldn't need to do this check.
+ let aligned_type =
+ if ty == self.cx.u128_type || ty == self.cx.i128_type {
+ ty
+ }
+ else {
+ ty.get_aligned(align.bytes())
+ };
+ // TODO(antoyo): It might be better to return a LValue, but fixing the rustc API is non-trivial.
+ self.stack_var_count.set(self.stack_var_count.get() + 1);
+ self.current_func().new_local(None, aligned_type, &format!("stack_var_{}", self.stack_var_count.get())).get_address(None)
+ }
+
+ fn dynamic_alloca(&mut self, _ty: Type<'gcc>, _align: Align) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn array_alloca(&mut self, _ty: Type<'gcc>, _len: RValue<'gcc>, _align: Align) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn load(&mut self, _ty: Type<'gcc>, ptr: RValue<'gcc>, _align: Align) -> RValue<'gcc> {
+ // TODO(antoyo): use ty.
+ let block = self.llbb();
+ let function = block.get_function();
+ // NOTE: instead of returning the dereference here, we have to assign it to a variable in
+ // the current basic block. Otherwise, it could be used in another basic block, causing a
+ // dereference after a drop, for instance.
+ // TODO(antoyo): handle align.
+ let deref = ptr.dereference(None).to_rvalue();
+ let value_type = deref.get_type();
+ unsafe { RETURN_VALUE_COUNT += 1 };
+ let loaded_value = function.new_local(None, value_type, &format!("loadedValue{}", unsafe { RETURN_VALUE_COUNT }));
+ block.add_assignment(None, loaded_value, deref);
+ loaded_value.to_rvalue()
+ }
+
+ fn volatile_load(&mut self, _ty: Type<'gcc>, ptr: RValue<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): use ty.
+ let ptr = self.context.new_cast(None, ptr, ptr.get_type().make_volatile());
+ ptr.dereference(None).to_rvalue()
+ }
+
+ fn atomic_load(&mut self, _ty: Type<'gcc>, ptr: RValue<'gcc>, order: AtomicOrdering, size: Size) -> RValue<'gcc> {
+ // TODO(antoyo): use ty.
+ // TODO(antoyo): handle alignment.
+ let atomic_load = self.context.get_builtin_function(&format!("__atomic_load_{}", size.bytes()));
+ let ordering = self.context.new_rvalue_from_int(self.i32_type, order.to_gcc());
+
+ let volatile_const_void_ptr_type = self.context.new_type::<*mut ()>().make_const().make_volatile();
+ let ptr = self.context.new_cast(None, ptr, volatile_const_void_ptr_type);
+ self.context.new_call(None, atomic_load, &[ptr, ordering])
+ }
+
+ fn load_operand(&mut self, place: PlaceRef<'tcx, RValue<'gcc>>) -> OperandRef<'tcx, RValue<'gcc>> {
+ assert_eq!(place.llextra.is_some(), place.layout.is_unsized());
+
+ if place.layout.is_zst() {
+ return OperandRef::new_zst(self, place.layout);
+ }
+
+ fn scalar_load_metadata<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>, load: RValue<'gcc>, scalar: &abi::Scalar) {
+ let vr = scalar.valid_range.clone();
+ match scalar.value {
+ abi::Int(..) => {
- abi::Pointer if vr.start() < vr.end() && !vr.contains(&0) => {
++ if !scalar.is_always_valid(bx) {
++ bx.range_metadata(load, scalar.valid_range);
+ }
+ }
- fn range_metadata(&mut self, _load: RValue<'gcc>, _range: Range<u128>) {
++ abi::Pointer if vr.start < vr.end && !vr.contains(0) => {
+ bx.nonnull_metadata(load);
+ }
+ _ => {}
+ }
+ }
+
+ let val =
+ if let Some(llextra) = place.llextra {
+ OperandValue::Ref(place.llval, Some(llextra), place.align)
+ }
+ else if place.layout.is_gcc_immediate() {
+ let load = self.load(place.llval.get_type(), place.llval, place.align);
+ if let abi::Abi::Scalar(ref scalar) = place.layout.abi {
+ scalar_load_metadata(self, load, scalar);
+ }
+ OperandValue::Immediate(self.to_immediate(load, place.layout))
+ }
+ else if let abi::Abi::ScalarPair(ref a, ref b) = place.layout.abi {
+ let b_offset = a.value.size(self).align_to(b.value.align(self).abi);
+ let pair_type = place.layout.gcc_type(self, false);
+
+ let mut load = |i, scalar: &abi::Scalar, align| {
+ let llptr = self.struct_gep(pair_type, place.llval, i as u64);
+ let load = self.load(llptr.get_type(), llptr, align);
+ scalar_load_metadata(self, load, scalar);
+ if scalar.is_bool() { self.trunc(load, self.type_i1()) } else { load }
+ };
+
+ OperandValue::Pair(
+ load(0, a, place.align),
+ load(1, b, place.align.restrict_for_offset(b_offset)),
+ )
+ }
+ else {
+ OperandValue::Ref(place.llval, None, place.align)
+ };
+
+ OperandRef { val, layout: place.layout }
+ }
+
+ fn write_operand_repeatedly(mut self, cg_elem: OperandRef<'tcx, RValue<'gcc>>, count: u64, dest: PlaceRef<'tcx, RValue<'gcc>>) -> Self {
+ let zero = self.const_usize(0);
+ let count = self.const_usize(count);
+ let start = dest.project_index(&mut self, zero).llval;
+ let end = dest.project_index(&mut self, count).llval;
+
+ let mut header_bx = self.build_sibling_block("repeat_loop_header");
+ let mut body_bx = self.build_sibling_block("repeat_loop_body");
+ let next_bx = self.build_sibling_block("repeat_loop_next");
+
+ let ptr_type = start.get_type();
+ let current = self.llbb().get_function().new_local(None, ptr_type, "loop_var");
+ let current_val = current.to_rvalue();
+ self.assign(current, start);
+
+ self.br(header_bx.llbb());
+
+ let keep_going = header_bx.icmp(IntPredicate::IntNE, current_val, end);
+ header_bx.cond_br(keep_going, body_bx.llbb(), next_bx.llbb());
+
+ let align = dest.align.restrict_for_offset(dest.layout.field(self.cx(), 0).size);
+ cg_elem.val.store(&mut body_bx, PlaceRef::new_sized_aligned(current_val, cg_elem.layout, align));
+
+ let next = body_bx.inbounds_gep(self.backend_type(cg_elem.layout), current.to_rvalue(), &[self.const_usize(1)]);
+ body_bx.llbb().add_assignment(None, current, next);
+ body_bx.br(header_bx.llbb());
+
+ next_bx
+ }
+
- fn to_immediate_scalar(&mut self, val: Self::Value, scalar: &abi::Scalar) -> Self::Value {
++ fn range_metadata(&mut self, _load: RValue<'gcc>, _range: WrappingRange) {
+ // TODO(antoyo)
+ }
+
+ fn nonnull_metadata(&mut self, _load: RValue<'gcc>) {
+ // TODO(antoyo)
+ }
+
+ fn store(&mut self, val: RValue<'gcc>, ptr: RValue<'gcc>, align: Align) -> RValue<'gcc> {
+ self.store_with_flags(val, ptr, align, MemFlags::empty())
+ }
+
+ fn store_with_flags(&mut self, val: RValue<'gcc>, ptr: RValue<'gcc>, _align: Align, _flags: MemFlags) -> RValue<'gcc> {
+ let ptr = self.check_store(val, ptr);
+ self.llbb().add_assignment(None, ptr.dereference(None), val);
+ // TODO(antoyo): handle align and flags.
+ // NOTE: dummy value here since it's never used. FIXME(antoyo): API should not return a value here?
+ self.cx.context.new_rvalue_zero(self.type_i32())
+ }
+
+ fn atomic_store(&mut self, value: RValue<'gcc>, ptr: RValue<'gcc>, order: AtomicOrdering, size: Size) {
+ // TODO(antoyo): handle alignment.
+ let atomic_store = self.context.get_builtin_function(&format!("__atomic_store_{}", size.bytes()));
+ let ordering = self.context.new_rvalue_from_int(self.i32_type, order.to_gcc());
+ let volatile_const_void_ptr_type = self.context.new_type::<*mut ()>().make_const().make_volatile();
+ let ptr = self.context.new_cast(None, ptr, volatile_const_void_ptr_type);
+
+ // FIXME(antoyo): fix libgccjit to allow comparing an integer type with an aligned integer type because
+ // the following cast is required to avoid this error:
+ // gcc_jit_context_new_call: mismatching types for argument 2 of function "__atomic_store_4": assignment to param arg1 (type: int) from loadedValue3577 (type: unsigned int __attribute__((aligned(4))))
+ let int_type = atomic_store.get_param(1).to_rvalue().get_type();
+ let value = self.context.new_cast(None, value, int_type);
+ self.llbb()
+ .add_eval(None, self.context.new_call(None, atomic_store, &[ptr, value, ordering]));
+ }
+
+ fn gep(&mut self, _typ: Type<'gcc>, ptr: RValue<'gcc>, indices: &[RValue<'gcc>]) -> RValue<'gcc> {
+ let mut result = ptr;
+ for index in indices {
+ result = self.context.new_array_access(None, result, *index).get_address(None).to_rvalue();
+ }
+ result
+ }
+
+ fn inbounds_gep(&mut self, _typ: Type<'gcc>, ptr: RValue<'gcc>, indices: &[RValue<'gcc>]) -> RValue<'gcc> {
+ // FIXME(antoyo): would be safer if doing the same thing (loop) as gep.
+ // TODO(antoyo): specify inbounds somehow.
+ match indices.len() {
+ 1 => {
+ self.context.new_array_access(None, ptr, indices[0]).get_address(None)
+ },
+ 2 => {
+ let array = ptr.dereference(None); // TODO(antoyo): assert that first index is 0?
+ self.context.new_array_access(None, array, indices[1]).get_address(None)
+ },
+ _ => unimplemented!(),
+ }
+ }
+
+ fn struct_gep(&mut self, value_type: Type<'gcc>, ptr: RValue<'gcc>, idx: u64) -> RValue<'gcc> {
+ // FIXME(antoyo): it would be better if the API only called this on struct, not on arrays.
+ assert_eq!(idx as usize as u64, idx);
+ let value = ptr.dereference(None).to_rvalue();
+
+ if value_type.is_array().is_some() {
+ let index = self.context.new_rvalue_from_long(self.u64_type, i64::try_from(idx).expect("i64::try_from"));
+ let element = self.context.new_array_access(None, value, index);
+ element.get_address(None)
+ }
+ else if let Some(vector_type) = value_type.is_vector() {
+ let array_type = vector_type.get_element_type().make_pointer();
+ let array = self.bitcast(ptr, array_type);
+ let index = self.context.new_rvalue_from_long(self.u64_type, i64::try_from(idx).expect("i64::try_from"));
+ let element = self.context.new_array_access(None, array, index);
+ element.get_address(None)
+ }
+ else if let Some(struct_type) = value_type.is_struct() {
+ ptr.dereference_field(None, struct_type.get_field(idx as i32)).get_address(None)
+ }
+ else {
+ panic!("Unexpected type {:?}", value_type);
+ }
+ }
+
+ /* Casts */
+ fn trunc(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): check that it indeed truncate the value.
+ self.context.new_cast(None, value, dest_ty)
+ }
+
+ fn sext(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): check that it indeed sign extend the value.
+ if dest_ty.is_vector().is_some() {
+ // TODO(antoyo): nothing to do as it is only for LLVM?
+ return value;
+ }
+ self.context.new_cast(None, value, dest_ty)
+ }
+
+ fn fptoui(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ self.context.new_cast(None, value, dest_ty)
+ }
+
+ fn fptosi(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ self.context.new_cast(None, value, dest_ty)
+ }
+
+ fn uitofp(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ self.context.new_cast(None, value, dest_ty)
+ }
+
+ fn sitofp(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ self.context.new_cast(None, value, dest_ty)
+ }
+
+ fn fptrunc(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): make sure it truncates.
+ self.context.new_cast(None, value, dest_ty)
+ }
+
+ fn fpext(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ self.context.new_cast(None, value, dest_ty)
+ }
+
+ fn ptrtoint(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ self.cx.ptrtoint(self.block.expect("block"), value, dest_ty)
+ }
+
+ fn inttoptr(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ self.cx.inttoptr(self.block.expect("block"), value, dest_ty)
+ }
+
+ fn bitcast(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ self.cx.const_bitcast(value, dest_ty)
+ }
+
+ fn intcast(&mut self, value: RValue<'gcc>, dest_typ: Type<'gcc>, _is_signed: bool) -> RValue<'gcc> {
+ // NOTE: is_signed is for value, not dest_typ.
+ self.cx.context.new_cast(None, value, dest_typ)
+ }
+
+ fn pointercast(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ let val_type = value.get_type();
+ match (type_is_pointer(val_type), type_is_pointer(dest_ty)) {
+ (false, true) => {
+ // NOTE: Projecting a field of a pointer type will attemp a cast from a signed char to
+ // a pointer, which is not supported by gccjit.
+ return self.cx.context.new_cast(None, self.inttoptr(value, val_type.make_pointer()), dest_ty);
+ },
+ (false, false) => {
+ // When they are not pointers, we want a transmute (or reinterpret_cast).
+ self.bitcast(value, dest_ty)
+ },
+ (true, true) => self.cx.context.new_cast(None, value, dest_ty),
+ (true, false) => unimplemented!(),
+ }
+ }
+
+ /* Comparisons */
+ fn icmp(&mut self, op: IntPredicate, lhs: RValue<'gcc>, mut rhs: RValue<'gcc>) -> RValue<'gcc> {
+ if lhs.get_type() != rhs.get_type() {
+ // NOTE: hack because we try to cast a vector type to the same vector type.
+ if format!("{:?}", lhs.get_type()) != format!("{:?}", rhs.get_type()) {
+ rhs = self.context.new_cast(None, rhs, lhs.get_type());
+ }
+ }
+ self.context.new_comparison(None, op.to_gcc_comparison(), lhs, rhs)
+ }
+
+ fn fcmp(&mut self, op: RealPredicate, lhs: RValue<'gcc>, rhs: RValue<'gcc>) -> RValue<'gcc> {
+ self.context.new_comparison(None, op.to_gcc_comparison(), lhs, rhs)
+ }
+
+ /* Miscellaneous instructions */
+ fn memcpy(&mut self, dst: RValue<'gcc>, dst_align: Align, src: RValue<'gcc>, src_align: Align, size: RValue<'gcc>, flags: MemFlags) {
+ if flags.contains(MemFlags::NONTEMPORAL) {
+ // HACK(nox): This is inefficient but there is no nontemporal memcpy.
+ let val = self.load(src.get_type(), src, src_align);
+ let ptr = self.pointercast(dst, self.type_ptr_to(self.val_ty(val)));
+ self.store_with_flags(val, ptr, dst_align, flags);
+ return;
+ }
+ let size = self.intcast(size, self.type_size_t(), false);
+ let _is_volatile = flags.contains(MemFlags::VOLATILE);
+ let dst = self.pointercast(dst, self.type_i8p());
+ let src = self.pointercast(src, self.type_ptr_to(self.type_void()));
+ let memcpy = self.context.get_builtin_function("memcpy");
+ let block = self.block.expect("block");
+ // TODO(antoyo): handle aligns and is_volatile.
+ block.add_eval(None, self.context.new_call(None, memcpy, &[dst, src, size]));
+ }
+
+ fn memmove(&mut self, dst: RValue<'gcc>, dst_align: Align, src: RValue<'gcc>, src_align: Align, size: RValue<'gcc>, flags: MemFlags) {
+ if flags.contains(MemFlags::NONTEMPORAL) {
+ // HACK(nox): This is inefficient but there is no nontemporal memmove.
+ let val = self.load(src.get_type(), src, src_align);
+ let ptr = self.pointercast(dst, self.type_ptr_to(self.val_ty(val)));
+ self.store_with_flags(val, ptr, dst_align, flags);
+ return;
+ }
+ let size = self.intcast(size, self.type_size_t(), false);
+ let _is_volatile = flags.contains(MemFlags::VOLATILE);
+ let dst = self.pointercast(dst, self.type_i8p());
+ let src = self.pointercast(src, self.type_ptr_to(self.type_void()));
+
+ let memmove = self.context.get_builtin_function("memmove");
+ let block = self.block.expect("block");
+ // TODO(antoyo): handle is_volatile.
+ block.add_eval(None, self.context.new_call(None, memmove, &[dst, src, size]));
+ }
+
+ fn memset(&mut self, ptr: RValue<'gcc>, fill_byte: RValue<'gcc>, size: RValue<'gcc>, _align: Align, flags: MemFlags) {
+ let _is_volatile = flags.contains(MemFlags::VOLATILE);
+ let ptr = self.pointercast(ptr, self.type_i8p());
+ let memset = self.context.get_builtin_function("memset");
+ let block = self.block.expect("block");
+ // TODO(antoyo): handle align and is_volatile.
+ let fill_byte = self.context.new_cast(None, fill_byte, self.i32_type);
+ let size = self.intcast(size, self.type_size_t(), false);
+ block.add_eval(None, self.context.new_call(None, memset, &[ptr, fill_byte, size]));
+ }
+
+ fn select(&mut self, cond: RValue<'gcc>, then_val: RValue<'gcc>, mut else_val: RValue<'gcc>) -> RValue<'gcc> {
+ let func = self.current_func();
+ let variable = func.new_local(None, then_val.get_type(), "selectVar");
+ let then_block = func.new_block("then");
+ let else_block = func.new_block("else");
+ let after_block = func.new_block("after");
+ self.llbb().end_with_conditional(None, cond, then_block, else_block);
+
+ then_block.add_assignment(None, variable, then_val);
+ then_block.end_with_jump(None, after_block);
+
+ if then_val.get_type() != else_val.get_type() {
+ else_val = self.context.new_cast(None, else_val, then_val.get_type());
+ }
+ else_block.add_assignment(None, variable, else_val);
+ else_block.end_with_jump(None, after_block);
+
+ // NOTE: since jumps were added in a place rustc does not expect, the current blocks in the
+ // state need to be updated.
+ self.block = Some(after_block);
+ *self.cx.current_block.borrow_mut() = Some(after_block);
+
+ variable.to_rvalue()
+ }
+
+ #[allow(dead_code)]
+ fn va_arg(&mut self, _list: RValue<'gcc>, _ty: Type<'gcc>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn extract_element(&mut self, _vec: RValue<'gcc>, _idx: RValue<'gcc>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn vector_splat(&mut self, _num_elts: usize, _elt: RValue<'gcc>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn extract_value(&mut self, aggregate_value: RValue<'gcc>, idx: u64) -> RValue<'gcc> {
+ // FIXME(antoyo): it would be better if the API only called this on struct, not on arrays.
+ assert_eq!(idx as usize as u64, idx);
+ let value_type = aggregate_value.get_type();
+
+ if value_type.is_array().is_some() {
+ let index = self.context.new_rvalue_from_long(self.u64_type, i64::try_from(idx).expect("i64::try_from"));
+ let element = self.context.new_array_access(None, aggregate_value, index);
+ element.get_address(None)
+ }
+ else if value_type.is_vector().is_some() {
+ panic!();
+ }
+ else if let Some(pointer_type) = value_type.get_pointee() {
+ if let Some(struct_type) = pointer_type.is_struct() {
+ // NOTE: hack to workaround a limitation of the rustc API: see comment on
+ // CodegenCx.structs_as_pointer
+ aggregate_value.dereference_field(None, struct_type.get_field(idx as i32)).to_rvalue()
+ }
+ else {
+ panic!("Unexpected type {:?}", value_type);
+ }
+ }
+ else if let Some(struct_type) = value_type.is_struct() {
+ aggregate_value.access_field(None, struct_type.get_field(idx as i32)).to_rvalue()
+ }
+ else {
+ panic!("Unexpected type {:?}", value_type);
+ }
+ }
+
+ fn insert_value(&mut self, aggregate_value: RValue<'gcc>, value: RValue<'gcc>, idx: u64) -> RValue<'gcc> {
+ // FIXME(antoyo): it would be better if the API only called this on struct, not on arrays.
+ assert_eq!(idx as usize as u64, idx);
+ let value_type = aggregate_value.get_type();
+
+ let lvalue =
+ if value_type.is_array().is_some() {
+ let index = self.context.new_rvalue_from_long(self.u64_type, i64::try_from(idx).expect("i64::try_from"));
+ self.context.new_array_access(None, aggregate_value, index)
+ }
+ else if value_type.is_vector().is_some() {
+ panic!();
+ }
+ else if let Some(pointer_type) = value_type.get_pointee() {
+ if let Some(struct_type) = pointer_type.is_struct() {
+ // NOTE: hack to workaround a limitation of the rustc API: see comment on
+ // CodegenCx.structs_as_pointer
+ aggregate_value.dereference_field(None, struct_type.get_field(idx as i32))
+ }
+ else {
+ panic!("Unexpected type {:?}", value_type);
+ }
+ }
+ else {
+ panic!("Unexpected type {:?}", value_type);
+ };
+ self.llbb().add_assignment(None, lvalue, value);
+
+ aggregate_value
+ }
+
+ fn landing_pad(&mut self, _ty: Type<'gcc>, _pers_fn: RValue<'gcc>, _num_clauses: usize) -> RValue<'gcc> {
+ let field1 = self.context.new_field(None, self.u8_type, "landing_pad_field_1");
+ let field2 = self.context.new_field(None, self.i32_type, "landing_pad_field_1");
+ let struct_type = self.context.new_struct_type(None, "landing_pad", &[field1, field2]);
+ self.current_func().new_local(None, struct_type.as_type(), "landing_pad")
+ .to_rvalue()
+ // TODO(antoyo): Properly implement unwinding.
+ // the above is just to make the compilation work as it seems
+ // rustc_codegen_ssa now calls the unwinding builder methods even on panic=abort.
+ }
+
+ fn set_cleanup(&mut self, _landing_pad: RValue<'gcc>) {
+ // TODO(antoyo)
+ }
+
+ fn resume(&mut self, _exn: RValue<'gcc>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn cleanup_pad(&mut self, _parent: Option<RValue<'gcc>>, _args: &[RValue<'gcc>]) -> Funclet {
+ unimplemented!();
+ }
+
+ fn cleanup_ret(&mut self, _funclet: &Funclet, _unwind: Option<Block<'gcc>>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn catch_pad(&mut self, _parent: RValue<'gcc>, _args: &[RValue<'gcc>]) -> Funclet {
+ unimplemented!();
+ }
+
+ fn catch_switch(&mut self, _parent: Option<RValue<'gcc>>, _unwind: Option<Block<'gcc>>, _num_handlers: usize) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn add_handler(&mut self, _catch_switch: RValue<'gcc>, _handler: Block<'gcc>) {
+ unimplemented!();
+ }
+
+ fn set_personality_fn(&mut self, _personality: RValue<'gcc>) {
+ // TODO(antoyo)
+ }
+
+ // Atomic Operations
+ fn atomic_cmpxchg(&mut self, dst: RValue<'gcc>, cmp: RValue<'gcc>, src: RValue<'gcc>, order: AtomicOrdering, failure_order: AtomicOrdering, weak: bool) -> RValue<'gcc> {
+ let expected = self.current_func().new_local(None, cmp.get_type(), "expected");
+ self.llbb().add_assignment(None, expected, cmp);
+ let success = self.compare_exchange(dst, expected, src, order, failure_order, weak);
+
+ let pair_type = self.cx.type_struct(&[src.get_type(), self.bool_type], false);
+ let result = self.current_func().new_local(None, pair_type, "atomic_cmpxchg_result");
+ let align = Align::from_bits(64).expect("align"); // TODO(antoyo): use good align.
+
+ let value_type = result.to_rvalue().get_type();
+ if let Some(struct_type) = value_type.is_struct() {
+ self.store(success, result.access_field(None, struct_type.get_field(1)).get_address(None), align);
+ // NOTE: since success contains the call to the intrinsic, it must be stored before
+ // expected so that we store expected after the call.
+ self.store(expected.to_rvalue(), result.access_field(None, struct_type.get_field(0)).get_address(None), align);
+ }
+ // TODO(antoyo): handle when value is not a struct.
+
+ result.to_rvalue()
+ }
+
+ fn atomic_rmw(&mut self, op: AtomicRmwBinOp, dst: RValue<'gcc>, src: RValue<'gcc>, order: AtomicOrdering) -> RValue<'gcc> {
+ let size = self.cx.int_width(src.get_type()) / 8;
+ let name =
+ match op {
+ AtomicRmwBinOp::AtomicXchg => format!("__atomic_exchange_{}", size),
+ AtomicRmwBinOp::AtomicAdd => format!("__atomic_fetch_add_{}", size),
+ AtomicRmwBinOp::AtomicSub => format!("__atomic_fetch_sub_{}", size),
+ AtomicRmwBinOp::AtomicAnd => format!("__atomic_fetch_and_{}", size),
+ AtomicRmwBinOp::AtomicNand => format!("__atomic_fetch_nand_{}", size),
+ AtomicRmwBinOp::AtomicOr => format!("__atomic_fetch_or_{}", size),
+ AtomicRmwBinOp::AtomicXor => format!("__atomic_fetch_xor_{}", size),
+ AtomicRmwBinOp::AtomicMax => return self.atomic_extremum(ExtremumOperation::Max, dst, src, order),
+ AtomicRmwBinOp::AtomicMin => return self.atomic_extremum(ExtremumOperation::Min, dst, src, order),
+ AtomicRmwBinOp::AtomicUMax => return self.atomic_extremum(ExtremumOperation::Max, dst, src, order),
+ AtomicRmwBinOp::AtomicUMin => return self.atomic_extremum(ExtremumOperation::Min, dst, src, order),
+ };
+
+
+ let atomic_function = self.context.get_builtin_function(name);
+ let order = self.context.new_rvalue_from_int(self.i32_type, order.to_gcc());
+
+ let void_ptr_type = self.context.new_type::<*mut ()>();
+ let volatile_void_ptr_type = void_ptr_type.make_volatile();
+ let dst = self.context.new_cast(None, dst, volatile_void_ptr_type);
+ // FIXME(antoyo): not sure why, but we have the wrong type here.
+ let new_src_type = atomic_function.get_param(1).to_rvalue().get_type();
+ let src = self.context.new_cast(None, src, new_src_type);
+ let res = self.context.new_call(None, atomic_function, &[dst, src, order]);
+ self.context.new_cast(None, res, src.get_type())
+ }
+
+ fn atomic_fence(&mut self, order: AtomicOrdering, scope: SynchronizationScope) {
+ let name =
+ match scope {
+ SynchronizationScope::SingleThread => "__atomic_signal_fence",
+ SynchronizationScope::CrossThread => "__atomic_thread_fence",
+ };
+ let thread_fence = self.context.get_builtin_function(name);
+ let order = self.context.new_rvalue_from_int(self.i32_type, order.to_gcc());
+ self.llbb().add_eval(None, self.context.new_call(None, thread_fence, &[order]));
+ }
+
+ fn set_invariant_load(&mut self, load: RValue<'gcc>) {
+ // NOTE: Hack to consider vtable function pointer as non-global-variable function pointer.
+ self.normal_function_addresses.borrow_mut().insert(load);
+ // TODO(antoyo)
+ }
+
+ fn lifetime_start(&mut self, _ptr: RValue<'gcc>, _size: Size) {
+ // TODO(antoyo)
+ }
+
+ fn lifetime_end(&mut self, _ptr: RValue<'gcc>, _size: Size) {
+ // TODO(antoyo)
+ }
+
+ fn call(&mut self, _typ: Type<'gcc>, func: RValue<'gcc>, args: &[RValue<'gcc>], funclet: Option<&Funclet>) -> RValue<'gcc> {
+ // FIXME(antoyo): remove when having a proper API.
+ let gcc_func = unsafe { std::mem::transmute(func) };
+ if self.functions.borrow().values().find(|value| **value == gcc_func).is_some() {
+ self.function_call(func, args, funclet)
+ }
+ else {
+ // If it's a not function that was defined, it's a function pointer.
+ self.function_ptr_call(func, args, funclet)
+ }
+ }
+
+ fn zext(&mut self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> {
+ // FIXME(antoyo): this does not zero-extend.
+ if value.get_type().is_bool() && dest_typ.is_i8(&self.cx) {
+ // FIXME(antoyo): hack because base::from_immediate converts i1 to i8.
+ // Fix the code in codegen_ssa::base::from_immediate.
+ return value;
+ }
+ self.context.new_cast(None, value, dest_typ)
+ }
+
+ fn cx(&self) -> &CodegenCx<'gcc, 'tcx> {
+ self.cx
+ }
+
+ fn do_not_inline(&mut self, _llret: RValue<'gcc>) {
+ unimplemented!();
+ }
+
+ fn set_span(&mut self, _span: Span) {}
+
+ fn from_immediate(&mut self, val: Self::Value) -> Self::Value {
+ if self.cx().val_ty(val) == self.cx().type_i1() {
+ self.zext(val, self.cx().type_i8())
+ }
+ else {
+ val
+ }
+ }
+
++ fn to_immediate_scalar(&mut self, val: Self::Value, scalar: abi::Scalar) -> Self::Value {
+ if scalar.is_bool() {
+ return self.trunc(val, self.cx().type_i1());
+ }
+ val
+ }
+
+ fn fptoui_sat(&mut self, _val: RValue<'gcc>, _dest_ty: Type<'gcc>) -> Option<RValue<'gcc>> {
+ None
+ }
+
+ fn fptosi_sat(&mut self, _val: RValue<'gcc>, _dest_ty: Type<'gcc>) -> Option<RValue<'gcc>> {
+ None
+ }
+
+ fn instrprof_increment(&mut self, _fn_name: RValue<'gcc>, _hash: RValue<'gcc>, _num_counters: RValue<'gcc>, _index: RValue<'gcc>) {
+ unimplemented!();
+ }
+}
+
+impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
+ pub fn shuffle_vector(&mut self, v1: RValue<'gcc>, v2: RValue<'gcc>, mask: RValue<'gcc>) -> RValue<'gcc> {
+ let return_type = v1.get_type();
+ let params = [
+ self.context.new_parameter(None, return_type, "v1"),
+ self.context.new_parameter(None, return_type, "v2"),
+ self.context.new_parameter(None, mask.get_type(), "mask"),
+ ];
+ let shuffle = self.context.new_function(None, FunctionType::Extern, return_type, ¶ms, "_mm_shuffle_epi8", false);
+ self.context.new_call(None, shuffle, &[v1, v2, mask])
+ }
+}
+
+impl<'a, 'gcc, 'tcx> StaticBuilderMethods for Builder<'a, 'gcc, 'tcx> {
+ fn get_static(&mut self, def_id: DefId) -> RValue<'gcc> {
+ // Forward to the `get_static` method of `CodegenCx`
+ self.cx().get_static(def_id)
+ }
+}
+
+impl<'tcx> HasParamEnv<'tcx> for Builder<'_, '_, 'tcx> {
+ fn param_env(&self) -> ParamEnv<'tcx> {
+ self.cx.param_env()
+ }
+}
+
+impl<'tcx> HasTargetSpec for Builder<'_, '_, 'tcx> {
+ fn target_spec(&self) -> &Target {
+ &self.cx.target_spec()
+ }
+}
+
+trait ToGccComp {
+ fn to_gcc_comparison(&self) -> ComparisonOp;
+}
+
+impl ToGccComp for IntPredicate {
+ fn to_gcc_comparison(&self) -> ComparisonOp {
+ match *self {
+ IntPredicate::IntEQ => ComparisonOp::Equals,
+ IntPredicate::IntNE => ComparisonOp::NotEquals,
+ IntPredicate::IntUGT => ComparisonOp::GreaterThan,
+ IntPredicate::IntUGE => ComparisonOp::GreaterThanEquals,
+ IntPredicate::IntULT => ComparisonOp::LessThan,
+ IntPredicate::IntULE => ComparisonOp::LessThanEquals,
+ IntPredicate::IntSGT => ComparisonOp::GreaterThan,
+ IntPredicate::IntSGE => ComparisonOp::GreaterThanEquals,
+ IntPredicate::IntSLT => ComparisonOp::LessThan,
+ IntPredicate::IntSLE => ComparisonOp::LessThanEquals,
+ }
+ }
+}
+
+impl ToGccComp for RealPredicate {
+ fn to_gcc_comparison(&self) -> ComparisonOp {
+ // TODO(antoyo): check that ordered vs non-ordered is respected.
+ match *self {
+ RealPredicate::RealPredicateFalse => unreachable!(),
+ RealPredicate::RealOEQ => ComparisonOp::Equals,
+ RealPredicate::RealOGT => ComparisonOp::GreaterThan,
+ RealPredicate::RealOGE => ComparisonOp::GreaterThanEquals,
+ RealPredicate::RealOLT => ComparisonOp::LessThan,
+ RealPredicate::RealOLE => ComparisonOp::LessThanEquals,
+ RealPredicate::RealONE => ComparisonOp::NotEquals,
+ RealPredicate::RealORD => unreachable!(),
+ RealPredicate::RealUNO => unreachable!(),
+ RealPredicate::RealUEQ => ComparisonOp::Equals,
+ RealPredicate::RealUGT => ComparisonOp::GreaterThan,
+ RealPredicate::RealUGE => ComparisonOp::GreaterThan,
+ RealPredicate::RealULT => ComparisonOp::LessThan,
+ RealPredicate::RealULE => ComparisonOp::LessThan,
+ RealPredicate::RealUNE => ComparisonOp::NotEquals,
+ RealPredicate::RealPredicateTrue => unreachable!(),
+ }
+ }
+}
+
+#[repr(C)]
+#[allow(non_camel_case_types)]
+enum MemOrdering {
+ __ATOMIC_RELAXED,
+ __ATOMIC_CONSUME,
+ __ATOMIC_ACQUIRE,
+ __ATOMIC_RELEASE,
+ __ATOMIC_ACQ_REL,
+ __ATOMIC_SEQ_CST,
+}
+
+trait ToGccOrdering {
+ fn to_gcc(self) -> i32;
+}
+
+impl ToGccOrdering for AtomicOrdering {
+ fn to_gcc(self) -> i32 {
+ use MemOrdering::*;
+
+ let ordering =
+ match self {
+ AtomicOrdering::NotAtomic => __ATOMIC_RELAXED, // TODO(antoyo): check if that's the same.
+ AtomicOrdering::Unordered => __ATOMIC_RELAXED,
+ AtomicOrdering::Monotonic => __ATOMIC_RELAXED, // TODO(antoyo): check if that's the same.
+ AtomicOrdering::Acquire => __ATOMIC_ACQUIRE,
+ AtomicOrdering::Release => __ATOMIC_RELEASE,
+ AtomicOrdering::AcquireRelease => __ATOMIC_ACQ_REL,
+ AtomicOrdering::SequentiallyConsistent => __ATOMIC_SEQ_CST,
+ };
+ ordering as i32
+ }
+}
--- /dev/null
- assert!(!instance.substs.has_param_types_or_consts());
+use gccjit::{FunctionType, RValue};
+use rustc_codegen_ssa::traits::BaseTypeMethods;
+use rustc_middle::ty::{Instance, TypeFoldable};
+use rustc_middle::ty::layout::{FnAbiExt, HasTyCtxt};
+use rustc_target::abi::call::FnAbi;
+
+use crate::abi::FnAbiGccExt;
+use crate::context::CodegenCx;
+
+/// Codegens a reference to a fn/method item, monomorphizing and
+/// inlining as it goes.
+///
+/// # Parameters
+///
+/// - `cx`: the crate context
+/// - `instance`: the instance to be instantiated
+pub fn get_fn<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, instance: Instance<'tcx>) -> RValue<'gcc> {
+ let tcx = cx.tcx();
+
+ assert!(!instance.substs.needs_infer());
+ assert!(!instance.substs.has_escaping_bound_vars());
+
+ if let Some(&func) = cx.instances.borrow().get(&instance) {
+ return func;
+ }
+
+ let sym = tcx.symbol_name(instance).name;
+
+ let fn_abi = FnAbi::of_instance(cx, instance, &[]);
+
+ let func =
+ if let Some(func) = cx.get_declared_value(&sym) {
+ // Create a fn pointer with the new signature.
+ let ptrty = fn_abi.ptr_to_gcc_type(cx);
+
+ // This is subtle and surprising, but sometimes we have to bitcast
+ // the resulting fn pointer. The reason has to do with external
+ // functions. If you have two crates that both bind the same C
+ // library, they may not use precisely the same types: for
+ // example, they will probably each declare their own structs,
+ // which are distinct types from LLVM's point of view (nominal
+ // types).
+ //
+ // Now, if those two crates are linked into an application, and
+ // they contain inlined code, you can wind up with a situation
+ // where both of those functions wind up being loaded into this
+ // application simultaneously. In that case, the same function
+ // (from LLVM's point of view) requires two types. But of course
+ // LLVM won't allow one function to have two types.
+ //
+ // What we currently do, therefore, is declare the function with
+ // one of the two types (whichever happens to come first) and then
+ // bitcast as needed when the function is referenced to make sure
+ // it has the type we expect.
+ //
+ // This can occur on either a crate-local or crate-external
+ // reference. It also occurs when testing libcore and in some
+ // other weird situations. Annoying.
+ if cx.val_ty(func) != ptrty {
+ // TODO(antoyo): cast the pointer.
+ func
+ }
+ else {
+ func
+ }
+ }
+ else {
+ cx.linkage.set(FunctionType::Extern);
+ let func = cx.declare_fn(&sym, &fn_abi);
+
+ // TODO(antoyo): set linkage and attributes.
+ func
+ };
+
+ cx.instances.borrow_mut().insert(instance, func);
+
+ func
+}
--- /dev/null
- use rustc_middle::ty::{layout::TyAndLayout, ScalarInt};
- use rustc_mir::interpret::{Allocation, GlobalAlloc, Scalar};
+use std::convert::TryFrom;
+use std::convert::TryInto;
+
+use gccjit::{Block, CType, RValue, Type, ToRValue};
+use rustc_codegen_ssa::mir::place::PlaceRef;
+use rustc_codegen_ssa::traits::{
+ BaseTypeMethods,
+ ConstMethods,
+ DerivedTypeMethods,
+ MiscMethods,
+ StaticMethods,
+};
+use rustc_middle::bug;
+use rustc_middle::mir::Mutability;
- use rustc_target::abi::{self, HasDataLayout, LayoutOf, Pointer, Size};
++use rustc_middle::ty::ScalarInt;
++use rustc_middle::ty::layout::{TyAndLayout, LayoutOf};
++use rustc_middle::mir::interpret::{Allocation, GlobalAlloc, Scalar};
+use rustc_span::Symbol;
- fn scalar_to_backend(&self, cv: Scalar, layout: &abi::Scalar, ty: Type<'gcc>) -> RValue<'gcc> {
++use rustc_target::abi::{self, HasDataLayout, Pointer, Size};
+
+use crate::consts::const_alloc_to_gcc;
+use crate::context::CodegenCx;
+use crate::type_of::LayoutGccExt;
+
+impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
+ pub fn const_bytes(&self, bytes: &[u8]) -> RValue<'gcc> {
+ bytes_in_context(self, bytes)
+ }
+
+ fn const_cstr(&self, symbol: Symbol, _null_terminated: bool) -> RValue<'gcc> {
+ // TODO(antoyo): handle null_terminated.
+ if let Some(&value) = self.const_cstr_cache.borrow().get(&symbol) {
+ return value.to_rvalue();
+ }
+
+ let global = self.global_string(&*symbol.as_str());
+
+ self.const_cstr_cache.borrow_mut().insert(symbol, global.dereference(None));
+ global
+ }
+
+ fn global_string(&self, string: &str) -> RValue<'gcc> {
+ // TODO(antoyo): handle non-null-terminated strings.
+ let string = self.context.new_string_literal(&*string);
+ let sym = self.generate_local_symbol_name("str");
+ // NOTE: TLS is always off for a string litteral.
+ // NOTE: string litterals do not have a link section.
+ let global = self.define_global(&sym, self.val_ty(string), false, None)
+ .unwrap_or_else(|| bug!("symbol `{}` is already defined", sym));
+ self.global_init_block.add_assignment(None, global.dereference(None), string);
+ global.to_rvalue()
+ // TODO(antoyo): set linkage.
+ }
+
+ pub fn inttoptr(&self, block: Block<'gcc>, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ let func = block.get_function();
+ let local = func.new_local(None, value.get_type(), "intLocal");
+ block.add_assignment(None, local, value);
+ let value_address = local.get_address(None);
+
+ let ptr = self.context.new_cast(None, value_address, dest_ty.make_pointer());
+ ptr.dereference(None).to_rvalue()
+ }
+
+ pub fn ptrtoint(&self, block: Block<'gcc>, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): when libgccjit allow casting from pointer to int, remove this.
+ let func = block.get_function();
+ let local = func.new_local(None, value.get_type(), "ptrLocal");
+ block.add_assignment(None, local, value);
+ let ptr_address = local.get_address(None);
+
+ let ptr = self.context.new_cast(None, ptr_address, dest_ty.make_pointer());
+ ptr.dereference(None).to_rvalue()
+ }
+}
+
+pub fn bytes_in_context<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, bytes: &[u8]) -> RValue<'gcc> {
+ let context = &cx.context;
+ let typ = context.new_array_type(None, context.new_type::<u8>(), bytes.len() as i32);
+ let global = cx.declare_unnamed_global(typ);
+ global.global_set_initializer(bytes);
+ global.to_rvalue()
+}
+
+pub fn type_is_pointer<'gcc>(typ: Type<'gcc>) -> bool {
+ typ.get_pointee().is_some()
+}
+
+impl<'gcc, 'tcx> ConstMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
+ fn const_null(&self, typ: Type<'gcc>) -> RValue<'gcc> {
+ if type_is_pointer(typ) {
+ self.context.new_null(typ)
+ }
+ else {
+ self.const_int(typ, 0)
+ }
+ }
+
+ fn const_undef(&self, typ: Type<'gcc>) -> RValue<'gcc> {
+ let local = self.current_func.borrow().expect("func")
+ .new_local(None, typ, "undefined");
+ if typ.is_struct().is_some() {
+ // NOTE: hack to workaround a limitation of the rustc API: see comment on
+ // CodegenCx.structs_as_pointer
+ let pointer = local.get_address(None);
+ self.structs_as_pointer.borrow_mut().insert(pointer);
+ pointer
+ }
+ else {
+ local.to_rvalue()
+ }
+ }
+
+ fn const_int(&self, typ: Type<'gcc>, int: i64) -> RValue<'gcc> {
+ self.context.new_rvalue_from_long(typ, i64::try_from(int).expect("i64::try_from"))
+ }
+
+ fn const_uint(&self, typ: Type<'gcc>, int: u64) -> RValue<'gcc> {
+ self.context.new_rvalue_from_long(typ, u64::try_from(int).expect("u64::try_from") as i64)
+ }
+
+ fn const_uint_big(&self, typ: Type<'gcc>, num: u128) -> RValue<'gcc> {
+ let num64: Result<i64, _> = num.try_into();
+ if let Ok(num) = num64 {
+ // FIXME(antoyo): workaround for a bug where libgccjit is expecting a constant.
+ // The operations >> 64 and | low are making the normal case a non-constant.
+ return self.context.new_rvalue_from_long(typ, num as i64);
+ }
+
+ if num >> 64 != 0 {
+ // FIXME(antoyo): use a new function new_rvalue_from_unsigned_long()?
+ let low = self.context.new_rvalue_from_long(self.u64_type, num as u64 as i64);
+ let high = self.context.new_rvalue_from_long(typ, (num >> 64) as u64 as i64);
+
+ let sixty_four = self.context.new_rvalue_from_long(typ, 64);
+ (high << sixty_four) | self.context.new_cast(None, low, typ)
+ }
+ else if typ.is_i128(self) {
+ let num = self.context.new_rvalue_from_long(self.u64_type, num as u64 as i64);
+ self.context.new_cast(None, num, typ)
+ }
+ else {
+ self.context.new_rvalue_from_long(typ, num as u64 as i64)
+ }
+ }
+
+ fn const_bool(&self, val: bool) -> RValue<'gcc> {
+ self.const_uint(self.type_i1(), val as u64)
+ }
+
+ fn const_i32(&self, i: i32) -> RValue<'gcc> {
+ self.const_int(self.type_i32(), i as i64)
+ }
+
+ fn const_u32(&self, i: u32) -> RValue<'gcc> {
+ self.const_uint(self.type_u32(), i as u64)
+ }
+
+ fn const_u64(&self, i: u64) -> RValue<'gcc> {
+ self.const_uint(self.type_u64(), i)
+ }
+
+ fn const_usize(&self, i: u64) -> RValue<'gcc> {
+ let bit_size = self.data_layout().pointer_size.bits();
+ if bit_size < 64 {
+ // make sure it doesn't overflow
+ assert!(i < (1 << bit_size));
+ }
+
+ self.const_uint(self.usize_type, i)
+ }
+
+ fn const_u8(&self, _i: u8) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn const_real(&self, _t: Type<'gcc>, _val: f64) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn const_str(&self, s: Symbol) -> (RValue<'gcc>, RValue<'gcc>) {
+ let len = s.as_str().len();
+ let cs = self.const_ptrcast(self.const_cstr(s, false),
+ self.type_ptr_to(self.layout_of(self.tcx.types.str_).gcc_type(self, true)),
+ );
+ (cs, self.const_usize(len as u64))
+ }
+
+ fn const_struct(&self, values: &[RValue<'gcc>], packed: bool) -> RValue<'gcc> {
+ let fields: Vec<_> = values.iter()
+ .map(|value| value.get_type())
+ .collect();
+ // TODO(antoyo): cache the type? It's anonymous, so probably not.
+ let name = fields.iter().map(|typ| format!("{:?}", typ)).collect::<Vec<_>>().join("_");
+ let typ = self.type_struct(&fields, packed);
+ let structure = self.global_init_func.new_local(None, typ, &name);
+ let struct_type = typ.is_struct().expect("struct type");
+ for (index, value) in values.iter().enumerate() {
+ let field = struct_type.get_field(index as i32);
+ let field_lvalue = structure.access_field(None, field);
+ self.global_init_block.add_assignment(None, field_lvalue, *value);
+ }
+ self.lvalue_to_rvalue(structure)
+ }
+
+ fn const_to_opt_uint(&self, _v: RValue<'gcc>) -> Option<u64> {
+ // TODO(antoyo)
+ None
+ }
+
+ fn const_to_opt_u128(&self, _v: RValue<'gcc>, _sign_ext: bool) -> Option<u128> {
+ // TODO(antoyo)
+ None
+ }
+
++ fn scalar_to_backend(&self, cv: Scalar, layout: abi::Scalar, ty: Type<'gcc>) -> RValue<'gcc> {
+ let bitsize = if layout.is_bool() { 1 } else { layout.value.size(self).bits() };
+ match cv {
+ Scalar::Int(ScalarInt::ZST) => {
+ assert_eq!(0, layout.value.size(self).bytes());
+ self.const_undef(self.type_ix(0))
+ }
+ Scalar::Int(int) => {
+ let data = int.assert_bits(layout.value.size(self));
+
+ // FIXME(antoyo): there's some issues with using the u128 code that follows, so hard-code
+ // the paths for floating-point values.
+ if ty == self.float_type {
+ return self.context.new_rvalue_from_double(ty, f32::from_bits(data as u32) as f64);
+ }
+ else if ty == self.double_type {
+ return self.context.new_rvalue_from_double(ty, f64::from_bits(data as u64));
+ }
+
+ let value = self.const_uint_big(self.type_ix(bitsize), data);
+ if layout.value == Pointer {
+ self.inttoptr(self.current_block.borrow().expect("block"), value, ty)
+ } else {
+ self.const_bitcast(value, ty)
+ }
+ }
+ Scalar::Ptr(ptr, _size) => {
+ let (alloc_id, offset) = ptr.into_parts();
+ let base_addr =
+ match self.tcx.global_alloc(alloc_id) {
+ GlobalAlloc::Memory(alloc) => {
+ let init = const_alloc_to_gcc(self, alloc);
+ let value =
+ match alloc.mutability {
+ Mutability::Mut => self.static_addr_of_mut(init, alloc.align, None),
+ _ => self.static_addr_of(init, alloc.align, None),
+ };
+ if !self.sess().fewer_names() {
+ // TODO(antoyo): set value name.
+ }
+ value
+ },
+ GlobalAlloc::Function(fn_instance) => {
+ self.get_fn_addr(fn_instance)
+ },
+ GlobalAlloc::Static(def_id) => {
+ assert!(self.tcx.is_static(def_id));
+ self.get_static(def_id)
+ },
+ };
+ let ptr_type = base_addr.get_type();
+ let base_addr = self.const_bitcast(base_addr, self.usize_type);
+ let offset = self.context.new_rvalue_from_long(self.usize_type, offset.bytes() as i64);
+ let ptr = self.const_bitcast(base_addr + offset, ptr_type);
+ let value = ptr.dereference(None);
+ if layout.value != Pointer {
+ self.const_bitcast(value.to_rvalue(), ty)
+ }
+ else {
+ self.const_bitcast(value.get_address(None), ty)
+ }
+ }
+ }
+ }
+
+ fn const_data_from_alloc(&self, alloc: &Allocation) -> Self::Value {
+ const_alloc_to_gcc(self, alloc)
+ }
+
+ fn from_const_alloc(&self, layout: TyAndLayout<'tcx>, alloc: &Allocation, offset: Size) -> PlaceRef<'tcx, RValue<'gcc>> {
+ assert_eq!(alloc.align, layout.align.abi);
+ let ty = self.type_ptr_to(layout.gcc_type(self, true));
+ let value =
+ if layout.size == Size::ZERO {
+ let value = self.const_usize(alloc.align.bytes());
+ self.context.new_cast(None, value, ty)
+ }
+ else {
+ let init = const_alloc_to_gcc(self, alloc);
+ let base_addr = self.static_addr_of(init, alloc.align, None);
+
+ let array = self.const_bitcast(base_addr, self.type_i8p());
+ let value = self.context.new_array_access(None, array, self.const_usize(offset.bytes())).get_address(None);
+ self.const_bitcast(value, ty)
+ };
+ PlaceRef::new_sized(value, layout)
+ }
+
+ fn const_ptrcast(&self, val: RValue<'gcc>, ty: Type<'gcc>) -> RValue<'gcc> {
+ self.context.new_cast(None, val, ty)
+ }
+}
+
+pub trait SignType<'gcc, 'tcx> {
+ fn is_signed(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_unsigned(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn to_signed(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc>;
+}
+
+impl<'gcc, 'tcx> SignType<'gcc, 'tcx> for Type<'gcc> {
+ fn is_signed(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.is_i8(cx) || self.is_i16(cx) || self.is_i32(cx) || self.is_i64(cx) || self.is_i128(cx)
+ }
+
+ fn is_unsigned(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.is_u8(cx) || self.is_u16(cx) || self.is_u32(cx) || self.is_u64(cx) || self.is_u128(cx)
+ }
+
+ fn to_signed(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc> {
+ if self.is_u8(cx) {
+ cx.i8_type
+ }
+ else if self.is_u16(cx) {
+ cx.i16_type
+ }
+ else if self.is_u32(cx) {
+ cx.i32_type
+ }
+ else if self.is_u64(cx) {
+ cx.i64_type
+ }
+ else if self.is_u128(cx) {
+ cx.i128_type
+ }
+ else {
+ self.clone()
+ }
+ }
+}
+
+pub trait TypeReflection<'gcc, 'tcx> {
+ fn is_uchar(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_ushort(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_uint(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_ulong(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_ulonglong(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+
+ fn is_i8(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_u8(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_i16(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_u16(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_i32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_u32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_i64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_u64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_i128(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_u128(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+
+ fn is_f32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+ fn is_f64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+}
+
+impl<'gcc, 'tcx> TypeReflection<'gcc, 'tcx> for Type<'gcc> {
+ fn is_uchar(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.u8_type
+ }
+
+ fn is_ushort(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.u16_type
+ }
+
+ fn is_uint(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.uint_type
+ }
+
+ fn is_ulong(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.ulong_type
+ }
+
+ fn is_ulonglong(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.ulonglong_type
+ }
+
+ fn is_i8(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.i8_type
+ }
+
+ fn is_u8(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.u8_type
+ }
+
+ fn is_i16(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.i16_type
+ }
+
+ fn is_u16(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.u16_type
+ }
+
+ fn is_i32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.i32_type
+ }
+
+ fn is_u32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.u32_type
+ }
+
+ fn is_i64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.i64_type
+ }
+
+ fn is_u64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.u64_type
+ }
+
+ fn is_i128(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.context.new_c_type(CType::Int128t)
+ }
+
+ fn is_u128(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.context.new_c_type(CType::UInt128t)
+ }
+
+ fn is_f32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.context.new_type::<f32>()
+ }
+
+ fn is_f64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
+ self.unqualified() == cx.context.new_type::<f64>()
+ }
+}
--- /dev/null
- use rustc_mir::interpret::{self, Allocation, ErrorHandled, Scalar as InterpScalar, read_target_uint};
+use gccjit::{RValue, Type};
+use rustc_codegen_ssa::traits::{BaseTypeMethods, ConstMethods, DerivedTypeMethods, StaticMethods};
+use rustc_hir as hir;
+use rustc_hir::Node;
+use rustc_middle::{bug, span_bug};
+use rustc_middle::middle::codegen_fn_attrs::{CodegenFnAttrFlags, CodegenFnAttrs};
+use rustc_middle::mir::mono::MonoItem;
+use rustc_middle::ty::{self, Instance, Ty};
- use rustc_target::abi::{self, Align, HasDataLayout, LayoutOf, Primitive, Size};
++use rustc_middle::ty::layout::LayoutOf;
++use rustc_middle::mir::interpret::{self, Allocation, ErrorHandled, Scalar as InterpScalar, read_target_uint};
+use rustc_span::Span;
+use rustc_span::def_id::DefId;
- &abi::Scalar { value: Primitive::Pointer, valid_range: 0..=!0 },
++use rustc_target::abi::{self, Align, HasDataLayout, Primitive, Size, WrappingRange};
+
+use crate::base;
+use crate::context::CodegenCx;
+use crate::mangled_std_symbols::{ARGC, ARGV, ARGV_INIT_ARRAY};
+use crate::type_of::LayoutGccExt;
+
+impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
+ pub fn const_bitcast(&self, value: RValue<'gcc>, typ: Type<'gcc>) -> RValue<'gcc> {
+ if value.get_type() == self.bool_type.make_pointer() {
+ if let Some(pointee) = typ.get_pointee() {
+ if pointee.is_vector().is_some() {
+ panic!()
+ }
+ }
+ }
+ self.context.new_bitcast(None, value, typ)
+ }
+}
+
+impl<'gcc, 'tcx> StaticMethods for CodegenCx<'gcc, 'tcx> {
+ fn static_addr_of(&self, cv: RValue<'gcc>, align: Align, kind: Option<&str>) -> RValue<'gcc> {
+ if let Some(global_value) = self.const_globals.borrow().get(&cv) {
+ // TODO(antoyo): upgrade alignment.
+ return *global_value;
+ }
+ let global_value = self.static_addr_of_mut(cv, align, kind);
+ // TODO(antoyo): set global constant.
+ self.const_globals.borrow_mut().insert(cv, global_value);
+ global_value
+ }
+
+ fn codegen_static(&self, def_id: DefId, is_mutable: bool) {
+ let attrs = self.tcx.codegen_fn_attrs(def_id);
+
+ let instance = Instance::mono(self.tcx, def_id);
+ let name = &*self.tcx.symbol_name(instance).name;
+
+ let (value, alloc) =
+ match codegen_static_initializer(&self, def_id) {
+ Ok(value) => value,
+ // Error has already been reported
+ Err(_) => return,
+ };
+
+ let is_tls = attrs.flags.contains(CodegenFnAttrFlags::THREAD_LOCAL);
+ let global = self.get_static(def_id);
+
+ // boolean SSA values are i1, but they have to be stored in i8 slots,
+ // otherwise some LLVM optimization passes don't work as expected
+ let val_llty = self.val_ty(value);
+ let value =
+ if val_llty == self.type_i1() {
+ unimplemented!();
+ }
+ else {
+ value
+ };
+
+ let instance = Instance::mono(self.tcx, def_id);
+ let ty = instance.ty(self.tcx, ty::ParamEnv::reveal_all());
+ let gcc_type = self.layout_of(ty).gcc_type(self, true);
+
+ let global =
+ if val_llty == gcc_type {
+ global
+ }
+ else {
+ // If we created the global with the wrong type,
+ // correct the type.
+ // TODO(antoyo): set value name, linkage and visibility.
+
+ let new_global = self.get_or_insert_global(&name, val_llty, is_tls, attrs.link_section);
+
+ // To avoid breaking any invariants, we leave around the old
+ // global for the moment; we'll replace all references to it
+ // with the new global later. (See base::codegen_backend.)
+ //self.statics_to_rauw.borrow_mut().push((global, new_global));
+ new_global
+ };
+ // TODO(antoyo): set alignment and initializer.
+ let value = self.rvalue_as_lvalue(value);
+ let value = value.get_address(None);
+ let dest_typ = global.get_type();
+ let value = self.context.new_cast(None, value, dest_typ);
+
+ // NOTE: do not init the variables related to argc/argv because it seems we cannot
+ // overwrite those variables.
+ // FIXME(antoyo): correctly support global variable initialization.
+ let skip_init = [
+ ARGV_INIT_ARRAY,
+ ARGC,
+ ARGV,
+ ];
+ if !skip_init.iter().any(|symbol_name| name.starts_with(symbol_name)) {
+ // TODO(antoyo): switch to set_initializer when libgccjit supports that.
+ let memcpy = self.context.get_builtin_function("memcpy");
+ let dst = self.context.new_cast(None, global, self.type_i8p());
+ let src = self.context.new_cast(None, value, self.type_ptr_to(self.type_void()));
+ let size = self.context.new_rvalue_from_long(self.sizet_type, alloc.size().bytes() as i64);
+ self.global_init_block.add_eval(None, self.context.new_call(None, memcpy, &[dst, src, size]));
+ }
+
+ // As an optimization, all shared statics which do not have interior
+ // mutability are placed into read-only memory.
+ if !is_mutable {
+ if self.type_is_freeze(ty) {
+ // TODO(antoyo): set global constant.
+ }
+ }
+
+ if attrs.flags.contains(CodegenFnAttrFlags::THREAD_LOCAL) {
+ // Do not allow LLVM to change the alignment of a TLS on macOS.
+ //
+ // By default a global's alignment can be freely increased.
+ // This allows LLVM to generate more performant instructions
+ // e.g., using load-aligned into a SIMD register.
+ //
+ // However, on macOS 10.10 or below, the dynamic linker does not
+ // respect any alignment given on the TLS (radar 24221680).
+ // This will violate the alignment assumption, and causing segfault at runtime.
+ //
+ // This bug is very easy to trigger. In `println!` and `panic!`,
+ // the `LOCAL_STDOUT`/`LOCAL_STDERR` handles are stored in a TLS,
+ // which the values would be `mem::replace`d on initialization.
+ // The implementation of `mem::replace` will use SIMD
+ // whenever the size is 32 bytes or higher. LLVM notices SIMD is used
+ // and tries to align `LOCAL_STDOUT`/`LOCAL_STDERR` to a 32-byte boundary,
+ // which macOS's dyld disregarded and causing crashes
+ // (see issues #51794, #51758, #50867, #48866 and #44056).
+ //
+ // To workaround the bug, we trick LLVM into not increasing
+ // the global's alignment by explicitly assigning a section to it
+ // (equivalent to automatically generating a `#[link_section]` attribute).
+ // See the comment in the `GlobalValue::canIncreaseAlignment()` function
+ // of `lib/IR/Globals.cpp` for why this works.
+ //
+ // When the alignment is not increased, the optimized `mem::replace`
+ // will use load-unaligned instructions instead, and thus avoiding the crash.
+ //
+ // We could remove this hack whenever we decide to drop macOS 10.10 support.
+ if self.tcx.sess.target.options.is_like_osx {
+ // The `inspect` method is okay here because we checked relocations, and
+ // because we are doing this access to inspect the final interpreter state
+ // (not as part of the interpreter execution).
+ //
+ // FIXME: This check requires that the (arbitrary) value of undefined bytes
+ // happens to be zero. Instead, we should only check the value of defined bytes
+ // and set all undefined bytes to zero if this allocation is headed for the
+ // BSS.
+ unimplemented!();
+ }
+ }
+
+ // Wasm statics with custom link sections get special treatment as they
+ // go into custom sections of the wasm executable.
+ if self.tcx.sess.opts.target_triple.triple().starts_with("wasm32") {
+ if let Some(_section) = attrs.link_section {
+ unimplemented!();
+ }
+ } else {
+ // TODO(antoyo): set link section.
+ }
+
+ if attrs.flags.contains(CodegenFnAttrFlags::USED) {
+ self.add_used_global(global);
+ }
+ }
+
+ /// Add a global value to a list to be stored in the `llvm.used` variable, an array of i8*.
+ fn add_used_global(&self, _global: RValue<'gcc>) {
+ // TODO(antoyo)
+ }
++
++ fn add_compiler_used_global(&self, _global: RValue<'gcc>) {
++ // TODO(antoyo)
++ }
+}
+
+impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
+ pub fn static_addr_of_mut(&self, cv: RValue<'gcc>, align: Align, kind: Option<&str>) -> RValue<'gcc> {
+ let (name, gv) =
+ match kind {
+ Some(kind) if !self.tcx.sess.fewer_names() => {
+ let name = self.generate_local_symbol_name(kind);
+ // TODO(antoyo): check if it's okay that TLS is off here.
+ // TODO(antoyo): check if it's okay that link_section is None here.
+ // TODO(antoyo): set alignment here as well.
+ let gv = self.define_global(&name[..], self.val_ty(cv), false, None).unwrap_or_else(|| {
+ bug!("symbol `{}` is already defined", name);
+ });
+ // TODO(antoyo): set linkage.
+ (name, gv)
+ }
+ _ => {
+ let index = self.global_gen_sym_counter.get();
+ let name = format!("global_{}_{}", index, self.codegen_unit.name());
+ let typ = self.val_ty(cv).get_aligned(align.bytes());
+ let global = self.define_private_global(typ);
+ (name, global)
+ },
+ };
+ // FIXME(antoyo): I think the name coming from generate_local_symbol_name() above cannot be used
+ // globally.
+ // NOTE: global seems to only be global in a module. So save the name instead of the value
+ // to import it later.
+ self.global_names.borrow_mut().insert(cv, name);
+ self.global_init_block.add_assignment(None, gv.dereference(None), cv);
+ // TODO(antoyo): set unnamed address.
+ gv
+ }
+
+ pub fn get_static(&self, def_id: DefId) -> RValue<'gcc> {
+ let instance = Instance::mono(self.tcx, def_id);
+ let fn_attrs = self.tcx.codegen_fn_attrs(def_id);
+ if let Some(&global) = self.instances.borrow().get(&instance) {
+ return global;
+ }
+
+ let defined_in_current_codegen_unit =
+ self.codegen_unit.items().contains_key(&MonoItem::Static(def_id));
+ assert!(
+ !defined_in_current_codegen_unit,
+ "consts::get_static() should always hit the cache for \
+ statics defined in the same CGU, but did not for `{:?}`",
+ def_id
+ );
+
+ let ty = instance.ty(self.tcx, ty::ParamEnv::reveal_all());
+ let sym = self.tcx.symbol_name(instance).name;
+
+ let global =
+ if let Some(def_id) = def_id.as_local() {
+ let id = self.tcx.hir().local_def_id_to_hir_id(def_id);
+ let llty = self.layout_of(ty).gcc_type(self, true);
+ // FIXME: refactor this to work without accessing the HIR
+ let global = match self.tcx.hir().get(id) {
+ Node::Item(&hir::Item { span, kind: hir::ItemKind::Static(..), .. }) => {
+ if let Some(global) = self.get_declared_value(&sym) {
+ if self.val_ty(global) != self.type_ptr_to(llty) {
+ span_bug!(span, "Conflicting types for static");
+ }
+ }
+
+ let is_tls = fn_attrs.flags.contains(CodegenFnAttrFlags::THREAD_LOCAL);
+ let global = self.declare_global(&sym, llty, is_tls, fn_attrs.link_section);
+
+ if !self.tcx.is_reachable_non_generic(def_id) {
+ // TODO(antoyo): set visibility.
+ }
+
+ global
+ }
+
+ Node::ForeignItem(&hir::ForeignItem {
+ span,
+ kind: hir::ForeignItemKind::Static(..),
+ ..
+ }) => {
+ let fn_attrs = self.tcx.codegen_fn_attrs(def_id);
+ check_and_apply_linkage(&self, &fn_attrs, ty, sym, span)
+ }
+
+ item => bug!("get_static: expected static, found {:?}", item),
+ };
+
+ global
+ }
+ else {
+ // FIXME(nagisa): perhaps the map of externs could be offloaded to llvm somehow?
+ //debug!("get_static: sym={} item_attr={:?}", sym, self.tcx.item_attrs(def_id));
+
+ let attrs = self.tcx.codegen_fn_attrs(def_id);
+ let span = self.tcx.def_span(def_id);
+ let global = check_and_apply_linkage(&self, &attrs, ty, sym, span);
+
+ let needs_dll_storage_attr = false; // TODO(antoyo)
+
+ // If this assertion triggers, there's something wrong with commandline
+ // argument validation.
+ debug_assert!(
+ !(self.tcx.sess.opts.cg.linker_plugin_lto.enabled()
+ && self.tcx.sess.target.options.is_like_msvc
+ && self.tcx.sess.opts.cg.prefer_dynamic)
+ );
+
+ if needs_dll_storage_attr {
+ // This item is external but not foreign, i.e., it originates from an external Rust
+ // crate. Since we don't know whether this crate will be linked dynamically or
+ // statically in the final application, we always mark such symbols as 'dllimport'.
+ // If final linkage happens to be static, we rely on compiler-emitted __imp_ stubs
+ // to make things work.
+ //
+ // However, in some scenarios we defer emission of statics to downstream
+ // crates, so there are cases where a static with an upstream DefId
+ // is actually present in the current crate. We can find out via the
+ // is_codegened_item query.
+ if !self.tcx.is_codegened_item(def_id) {
+ unimplemented!();
+ }
+ }
+ global
+ };
+
+ // TODO(antoyo): set dll storage class.
+
+ self.instances.borrow_mut().insert(instance, global);
+ global
+ }
+}
+
+pub fn const_alloc_to_gcc<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, alloc: &Allocation) -> RValue<'gcc> {
+ let mut llvals = Vec::with_capacity(alloc.relocations().len() + 1);
+ let dl = cx.data_layout();
+ let pointer_size = dl.pointer_size.bytes() as usize;
+
+ let mut next_offset = 0;
+ for &(offset, alloc_id) in alloc.relocations().iter() {
+ let offset = offset.bytes();
+ assert_eq!(offset as usize as u64, offset);
+ let offset = offset as usize;
+ if offset > next_offset {
+ // This `inspect` is okay since we have checked that it is not within a relocation, it
+ // is within the bounds of the allocation, and it doesn't affect interpreter execution
+ // (we inspect the result after interpreter execution). Any undef byte is replaced with
+ // some arbitrary byte value.
+ //
+ // FIXME: relay undef bytes to codegen as undef const bytes
+ let bytes = alloc.inspect_with_uninit_and_ptr_outside_interpreter(next_offset..offset);
+ llvals.push(cx.const_bytes(bytes));
+ }
+ let ptr_offset =
+ read_target_uint( dl.endian,
+ // This `inspect` is okay since it is within the bounds of the allocation, it doesn't
+ // affect interpreter execution (we inspect the result after interpreter execution),
+ // and we properly interpret the relocation as a relocation pointer offset.
+ alloc.inspect_with_uninit_and_ptr_outside_interpreter(offset..(offset + pointer_size)),
+ )
+ .expect("const_alloc_to_llvm: could not read relocation pointer")
+ as u64;
+ llvals.push(cx.scalar_to_backend(
+ InterpScalar::from_pointer(
+ interpret::Pointer::new(alloc_id, Size::from_bytes(ptr_offset)),
+ &cx.tcx,
+ ),
++ abi::Scalar { value: Primitive::Pointer, valid_range: WrappingRange { start: 0, end: !0 } },
+ cx.type_i8p(),
+ ));
+ next_offset = offset + pointer_size;
+ }
+ if alloc.len() >= next_offset {
+ let range = next_offset..alloc.len();
+ // This `inspect` is okay since we have check that it is after all relocations, it is
+ // within the bounds of the allocation, and it doesn't affect interpreter execution (we
+ // inspect the result after interpreter execution). Any undef byte is replaced with some
+ // arbitrary byte value.
+ //
+ // FIXME: relay undef bytes to codegen as undef const bytes
+ let bytes = alloc.inspect_with_uninit_and_ptr_outside_interpreter(range);
+ llvals.push(cx.const_bytes(bytes));
+ }
+
+ cx.const_struct(&llvals, true)
+}
+
+pub fn codegen_static_initializer<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, def_id: DefId) -> Result<(RValue<'gcc>, &'tcx Allocation), ErrorHandled> {
+ let alloc = cx.tcx.eval_static_initializer(def_id)?;
+ Ok((const_alloc_to_gcc(cx, alloc), alloc))
+}
+
+fn check_and_apply_linkage<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, attrs: &CodegenFnAttrs, ty: Ty<'tcx>, sym: &str, span: Span) -> RValue<'gcc> {
+ let is_tls = attrs.flags.contains(CodegenFnAttrFlags::THREAD_LOCAL);
+ let llty = cx.layout_of(ty).gcc_type(cx, true);
+ if let Some(linkage) = attrs.linkage {
+ // If this is a static with a linkage specified, then we need to handle
+ // it a little specially. The typesystem prevents things like &T and
+ // extern "C" fn() from being non-null, so we can't just declare a
+ // static and call it a day. Some linkages (like weak) will make it such
+ // that the static actually has a null value.
+ let llty2 =
+ if let ty::RawPtr(ref mt) = ty.kind() {
+ cx.layout_of(mt.ty).gcc_type(cx, true)
+ }
+ else {
+ cx.sess().span_fatal(
+ span,
+ "must have type `*const T` or `*mut T` due to `#[linkage]` attribute",
+ )
+ };
+ // Declare a symbol `foo` with the desired linkage.
+ let global1 = cx.declare_global_with_linkage(&sym, llty2, base::global_linkage_to_gcc(linkage));
+
+ // Declare an internal global `extern_with_linkage_foo` which
+ // is initialized with the address of `foo`. If `foo` is
+ // discarded during linking (for example, if `foo` has weak
+ // linkage and there are no definitions), then
+ // `extern_with_linkage_foo` will instead be initialized to
+ // zero.
+ let mut real_name = "_rust_extern_with_linkage_".to_string();
+ real_name.push_str(&sym);
+ let global2 =
+ cx.define_global(&real_name, llty, is_tls, attrs.link_section).unwrap_or_else(|| {
+ cx.sess().span_fatal(span, &format!("symbol `{}` is already defined", &sym))
+ });
+ // TODO(antoyo): set linkage.
+ let lvalue = global2.dereference(None);
+ cx.global_init_block.add_assignment(None, lvalue, global1);
+ // TODO(antoyo): use global_set_initializer() when it will work.
+ global2
+ }
+ else {
+ // Generate an external declaration.
+ // FIXME(nagisa): investigate whether it can be changed into define_global
+
+ // Thread-local statics in some other crate need to *always* be linked
+ // against in a thread-local fashion, so we need to be sure to apply the
+ // thread-local attribute locally if it was present remotely. If we
+ // don't do this then linker errors can be generated where the linker
+ // complains that one object files has a thread local version of the
+ // symbol and another one doesn't.
+ cx.declare_global(&sym, llty, is_tls, attrs.link_section)
+ }
+}
--- /dev/null
- use rustc_middle::bug;
+use std::cell::{Cell, RefCell};
+
+use gccjit::{
+ Block,
+ Context,
+ CType,
+ Function,
+ FunctionType,
+ LValue,
+ RValue,
+ Struct,
+ Type,
+};
+use rustc_codegen_ssa::base::wants_msvc_seh;
+use rustc_codegen_ssa::traits::{
+ BackendTypes,
+ MiscMethods,
+};
+use rustc_data_structures::base_n;
+use rustc_data_structures::fx::{FxHashMap, FxHashSet};
- use rustc_middle::ty::layout::{HasParamEnv, HasTyCtxt, LayoutError, TyAndLayout};
++use rustc_middle::span_bug;
+use rustc_middle::mir::mono::CodegenUnit;
+use rustc_middle::ty::{self, Instance, ParamEnv, PolyExistentialTraitRef, Ty, TyCtxt};
- use rustc_span::{Span, Symbol, DUMMY_SP};
- use rustc_target::abi::{HasDataLayout, LayoutOf, PointeeInfo, Size, TargetDataLayout, VariantIdx};
++use rustc_middle::ty::layout::{HasParamEnv, HasTyCtxt, LayoutError, TyAndLayout, LayoutOfHelpers};
+use rustc_session::Session;
- impl<'gcc, 'tcx> LayoutOf for CodegenCx<'gcc, 'tcx> {
- type Ty = Ty<'tcx>;
- type TyAndLayout = TyAndLayout<'tcx>;
++use rustc_span::{Span, Symbol};
++use rustc_target::abi::{HasDataLayout, PointeeInfo, Size, TargetDataLayout, VariantIdx};
+use rustc_target::spec::{HasTargetSpec, Target, TlsModel};
+
+use crate::callee::get_fn;
+use crate::declare::mangle_name;
+
+#[derive(Clone)]
+pub struct FuncSig<'gcc> {
+ pub params: Vec<Type<'gcc>>,
+ pub return_type: Type<'gcc>,
+}
+
+pub struct CodegenCx<'gcc, 'tcx> {
+ pub check_overflow: bool,
+ pub codegen_unit: &'tcx CodegenUnit<'tcx>,
+ pub context: &'gcc Context<'gcc>,
+
+ // TODO(antoyo): First set it to a dummy block to avoid using Option?
+ pub current_block: RefCell<Option<Block<'gcc>>>,
+ pub current_func: RefCell<Option<Function<'gcc>>>,
+ pub normal_function_addresses: RefCell<FxHashSet<RValue<'gcc>>>,
+
+ /// The function where globals are initialized.
+ pub global_init_func: Function<'gcc>,
+ pub global_init_block: Block<'gcc>,
+
+ pub functions: RefCell<FxHashMap<String, Function<'gcc>>>,
+
+ pub tls_model: gccjit::TlsModel,
+
+ pub bool_type: Type<'gcc>,
+ pub i8_type: Type<'gcc>,
+ pub i16_type: Type<'gcc>,
+ pub i32_type: Type<'gcc>,
+ pub i64_type: Type<'gcc>,
+ pub i128_type: Type<'gcc>,
+ pub isize_type: Type<'gcc>,
+
+ pub u8_type: Type<'gcc>,
+ pub u16_type: Type<'gcc>,
+ pub u32_type: Type<'gcc>,
+ pub u64_type: Type<'gcc>,
+ pub u128_type: Type<'gcc>,
+ pub usize_type: Type<'gcc>,
+
+ pub int_type: Type<'gcc>,
+ pub uint_type: Type<'gcc>,
+ pub long_type: Type<'gcc>,
+ pub ulong_type: Type<'gcc>,
+ pub ulonglong_type: Type<'gcc>,
+ pub sizet_type: Type<'gcc>,
+
+ pub float_type: Type<'gcc>,
+ pub double_type: Type<'gcc>,
+
+ pub linkage: Cell<FunctionType>,
+ pub scalar_types: RefCell<FxHashMap<Ty<'tcx>, Type<'gcc>>>,
+ pub types: RefCell<FxHashMap<(Ty<'tcx>, Option<VariantIdx>), Type<'gcc>>>,
+ pub tcx: TyCtxt<'tcx>,
+
+ pub struct_types: RefCell<FxHashMap<Vec<Type<'gcc>>, Type<'gcc>>>,
+
+ pub types_with_fields_to_set: RefCell<FxHashMap<Type<'gcc>, (Struct<'gcc>, TyAndLayout<'tcx>)>>,
+
+ /// Cache instances of monomorphic and polymorphic items
+ pub instances: RefCell<FxHashMap<Instance<'tcx>, RValue<'gcc>>>,
+ /// Cache generated vtables
+ pub vtables: RefCell<FxHashMap<(Ty<'tcx>, Option<ty::PolyExistentialTraitRef<'tcx>>), RValue<'gcc>>>,
+
+ /// Cache of emitted const globals (value -> global)
+ pub const_globals: RefCell<FxHashMap<RValue<'gcc>, RValue<'gcc>>>,
+
+ pub init_argv_var: RefCell<String>,
+ pub argv_initialized: Cell<bool>,
+
+ /// Cache of constant strings,
+ pub const_cstr_cache: RefCell<FxHashMap<Symbol, LValue<'gcc>>>,
+
+ /// Cache of globals.
+ pub globals: RefCell<FxHashMap<String, RValue<'gcc>>>,
+ // TODO(antoyo): remove global_names.
+ pub global_names: RefCell<FxHashMap<RValue<'gcc>, String>>,
+
+ /// A counter that is used for generating local symbol names
+ local_gen_sym_counter: Cell<usize>,
+ pub global_gen_sym_counter: Cell<usize>,
+
+ eh_personality: Cell<Option<RValue<'gcc>>>,
+
+ pub pointee_infos: RefCell<FxHashMap<(Ty<'tcx>, Size), Option<PointeeInfo>>>,
+
+ /// NOTE: a hack is used because the rustc API is not suitable to libgccjit and as such,
+ /// `const_undef()` returns struct as pointer so that they can later be assigned a value.
+ /// As such, this set remembers which of these pointers were returned by this function so that
+ /// they can be derefered later.
+ /// FIXME(antoyo): fix the rustc API to avoid having this hack.
+ pub structs_as_pointer: RefCell<FxHashSet<RValue<'gcc>>>,
+
+ /// Store the pointer of different types for safety.
+ /// When casting the values back to their original types, check that they are indeed that type
+ /// with these sets.
+ /// FIXME(antoyo): remove when the API supports more types.
+ #[cfg(debug_assertions)]
+ lvalues: RefCell<FxHashSet<LValue<'gcc>>>,
+}
+
+impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
+ pub fn new(context: &'gcc Context<'gcc>, codegen_unit: &'tcx CodegenUnit<'tcx>, tcx: TyCtxt<'tcx>) -> Self {
+ let check_overflow = tcx.sess.overflow_checks();
+ // TODO(antoyo): fix this mess. libgccjit seems to return random type when using new_int_type().
+ let isize_type = context.new_c_type(CType::LongLong);
+ let usize_type = context.new_c_type(CType::ULongLong);
+ let bool_type = context.new_type::<bool>();
+ let i8_type = context.new_type::<i8>();
+ let i16_type = context.new_type::<i16>();
+ let i32_type = context.new_type::<i32>();
+ let i64_type = context.new_c_type(CType::LongLong);
+ let i128_type = context.new_c_type(CType::Int128t).get_aligned(8); // TODO(antoyo): should the alignment be hard-coded?
+ let u8_type = context.new_type::<u8>();
+ let u16_type = context.new_type::<u16>();
+ let u32_type = context.new_type::<u32>();
+ let u64_type = context.new_c_type(CType::ULongLong);
+ let u128_type = context.new_c_type(CType::UInt128t).get_aligned(8); // TODO(antoyo): should the alignment be hard-coded?
+
+ let tls_model = to_gcc_tls_mode(tcx.sess.tls_model());
+
+ let float_type = context.new_type::<f32>();
+ let double_type = context.new_type::<f64>();
+
+ let int_type = context.new_c_type(CType::Int);
+ let uint_type = context.new_c_type(CType::UInt);
+ let long_type = context.new_c_type(CType::Long);
+ let ulong_type = context.new_c_type(CType::ULong);
+ let ulonglong_type = context.new_c_type(CType::ULongLong);
+ let sizet_type = context.new_c_type(CType::SizeT);
+
+ assert_eq!(isize_type, i64_type);
+ assert_eq!(usize_type, u64_type);
+
+ let mut functions = FxHashMap::default();
+ let builtins = [
+ "__builtin_unreachable", "abort", "__builtin_expect", "__builtin_add_overflow", "__builtin_mul_overflow",
+ "__builtin_saddll_overflow", /*"__builtin_sadd_overflow",*/ "__builtin_smulll_overflow", /*"__builtin_smul_overflow",*/
+ "__builtin_ssubll_overflow", /*"__builtin_ssub_overflow",*/ "__builtin_sub_overflow", "__builtin_uaddll_overflow",
+ "__builtin_uadd_overflow", "__builtin_umulll_overflow", "__builtin_umul_overflow", "__builtin_usubll_overflow",
+ "__builtin_usub_overflow", "sqrtf", "sqrt", "__builtin_powif", "__builtin_powi", "sinf", "sin", "cosf", "cos",
+ "powf", "pow", "expf", "exp", "exp2f", "exp2", "logf", "log", "log10f", "log10", "log2f", "log2", "fmaf",
+ "fma", "fabsf", "fabs", "fminf", "fmin", "fmaxf", "fmax", "copysignf", "copysign", "floorf", "floor", "ceilf",
+ "ceil", "truncf", "trunc", "rintf", "rint", "nearbyintf", "nearbyint", "roundf", "round",
+ "__builtin_expect_with_probability",
+ ];
+
+ for builtin in builtins.iter() {
+ functions.insert(builtin.to_string(), context.get_builtin_function(builtin));
+ }
+
+ let global_init_func = context.new_function(None, FunctionType::Exported, context.new_type::<()>(), &[],
+ &format!("__gccGlobalInit{}", unit_name(&codegen_unit)), false);
+ let global_init_block = global_init_func.new_block("initial");
+
+ Self {
+ check_overflow,
+ codegen_unit,
+ context,
+ current_block: RefCell::new(None),
+ current_func: RefCell::new(None),
+ normal_function_addresses: Default::default(),
+ functions: RefCell::new(functions),
+ global_init_func,
+ global_init_block,
+
+ tls_model,
+
+ bool_type,
+ i8_type,
+ i16_type,
+ i32_type,
+ i64_type,
+ i128_type,
+ isize_type,
+ usize_type,
+ u8_type,
+ u16_type,
+ u32_type,
+ u64_type,
+ u128_type,
+ int_type,
+ uint_type,
+ long_type,
+ ulong_type,
+ ulonglong_type,
+ sizet_type,
+
+ float_type,
+ double_type,
+
+ linkage: Cell::new(FunctionType::Internal),
+ #[cfg(debug_assertions)]
+ lvalues: Default::default(),
+ instances: Default::default(),
+ vtables: Default::default(),
+ const_globals: Default::default(),
+ init_argv_var: RefCell::new(String::new()),
+ argv_initialized: Cell::new(false),
+ const_cstr_cache: Default::default(),
+ global_names: Default::default(),
+ globals: Default::default(),
+ scalar_types: Default::default(),
+ types: Default::default(),
+ tcx,
+ struct_types: Default::default(),
+ types_with_fields_to_set: Default::default(),
+ local_gen_sym_counter: Cell::new(0),
+ global_gen_sym_counter: Cell::new(0),
+ eh_personality: Cell::new(None),
+ pointee_infos: Default::default(),
+ structs_as_pointer: Default::default(),
+ }
+ }
+
+ pub fn lvalue_to_rvalue(&self, value: LValue<'gcc>) -> RValue<'gcc> {
+ #[cfg(debug_assertions)]
+ self.lvalues.borrow_mut().insert(value);
+ unsafe { std::mem::transmute(value) }
+ }
+
+ pub fn rvalue_as_function(&self, value: RValue<'gcc>) -> Function<'gcc> {
+ let function: Function<'gcc> = unsafe { std::mem::transmute(value) };
+ debug_assert!(self.functions.borrow().values().find(|value| **value == function).is_some(),
+ "{:?} ({:?}) is not a function", value, value.get_type());
+ function
+ }
+
+ pub fn rvalue_as_lvalue(&self, value: RValue<'gcc>) -> LValue<'gcc> {
+ let lvalue: LValue<'gcc> = unsafe { std::mem::transmute(value) };
+ lvalue
+ }
+
+ pub fn sess(&self) -> &Session {
+ &self.tcx.sess
+ }
+}
+
+impl<'gcc, 'tcx> BackendTypes for CodegenCx<'gcc, 'tcx> {
+ type Value = RValue<'gcc>;
+ type Function = RValue<'gcc>;
+
+ type BasicBlock = Block<'gcc>;
+ type Type = Type<'gcc>;
+ type Funclet = (); // TODO(antoyo)
+
+ type DIScope = (); // TODO(antoyo)
+ type DILocation = (); // TODO(antoyo)
+ type DIVariable = (); // TODO(antoyo)
+}
+
+impl<'gcc, 'tcx> MiscMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
+ fn vtables(&self) -> &RefCell<FxHashMap<(Ty<'tcx>, Option<PolyExistentialTraitRef<'tcx>>), RValue<'gcc>>> {
+ &self.vtables
+ }
+
+ fn get_fn(&self, instance: Instance<'tcx>) -> RValue<'gcc> {
+ let func = get_fn(self, instance);
+ *self.current_func.borrow_mut() = Some(self.rvalue_as_function(func));
+ func
+ }
+
+ fn get_fn_addr(&self, instance: Instance<'tcx>) -> RValue<'gcc> {
+ let func = get_fn(self, instance);
+ let func = self.rvalue_as_function(func);
+ let ptr = func.get_address(None);
+
+ // TODO(antoyo): don't do this twice: i.e. in declare_fn and here.
+ // FIXME(antoyo): the rustc API seems to call get_fn_addr() when not needed (e.g. for FFI).
+
+ self.normal_function_addresses.borrow_mut().insert(ptr);
+
+ ptr
+ }
+
+ fn eh_personality(&self) -> RValue<'gcc> {
+ // The exception handling personality function.
+ //
+ // If our compilation unit has the `eh_personality` lang item somewhere
+ // within it, then we just need to codegen that. Otherwise, we're
+ // building an rlib which will depend on some upstream implementation of
+ // this function, so we just codegen a generic reference to it. We don't
+ // specify any of the types for the function, we just make it a symbol
+ // that LLVM can later use.
+ //
+ // Note that MSVC is a little special here in that we don't use the
+ // `eh_personality` lang item at all. Currently LLVM has support for
+ // both Dwarf and SEH unwind mechanisms for MSVC targets and uses the
+ // *name of the personality function* to decide what kind of unwind side
+ // tables/landing pads to emit. It looks like Dwarf is used by default,
+ // injecting a dependency on the `_Unwind_Resume` symbol for resuming
+ // an "exception", but for MSVC we want to force SEH. This means that we
+ // can't actually have the personality function be our standard
+ // `rust_eh_personality` function, but rather we wired it up to the
+ // CRT's custom personality function, which forces LLVM to consider
+ // landing pads as "landing pads for SEH".
+ if let Some(llpersonality) = self.eh_personality.get() {
+ return llpersonality;
+ }
+ let tcx = self.tcx;
+ let llfn = match tcx.lang_items().eh_personality() {
+ Some(def_id) if !wants_msvc_seh(self.sess()) => self.get_fn_addr(
+ ty::Instance::resolve(
+ tcx,
+ ty::ParamEnv::reveal_all(),
+ def_id,
+ tcx.intern_substs(&[]),
+ )
+ .unwrap().unwrap(),
+ ),
+ _ => {
+ let _name = if wants_msvc_seh(self.sess()) {
+ "__CxxFrameHandler3"
+ } else {
+ "rust_eh_personality"
+ };
+ //let func = self.declare_func(name, self.type_i32(), &[], true);
+ // FIXME(antoyo): this hack should not be needed. That will probably be removed when
+ // unwinding support is added.
+ self.context.new_rvalue_from_int(self.int_type, 0)
+ }
+ };
+ // TODO(antoyo): apply target cpu attributes.
+ self.eh_personality.set(Some(llfn));
+ llfn
+ }
+
+ fn sess(&self) -> &Session {
+ &self.tcx.sess
+ }
+
+ fn check_overflow(&self) -> bool {
+ self.check_overflow
+ }
+
+ fn codegen_unit(&self) -> &'tcx CodegenUnit<'tcx> {
+ self.codegen_unit
+ }
+
+ fn used_statics(&self) -> &RefCell<Vec<RValue<'gcc>>> {
+ unimplemented!();
+ }
+
+ fn set_frame_pointer_type(&self, _llfn: RValue<'gcc>) {
+ // TODO(antoyo)
+ }
+
+ fn apply_target_cpu_attr(&self, _llfn: RValue<'gcc>) {
+ // TODO(antoyo)
+ }
+
+ fn create_used_variable(&self) {
+ unimplemented!();
+ }
+
+ fn declare_c_main(&self, fn_type: Self::Type) -> Option<Self::Function> {
+ if self.get_declared_value("main").is_none() {
+ Some(self.declare_cfn("main", fn_type))
+ }
+ else {
+ // If the symbol already exists, it is an error: for example, the user wrote
+ // #[no_mangle] extern "C" fn main(..) {..}
+ // instead of #[start]
+ None
+ }
+ }
++
++ fn compiler_used_statics(&self) -> &RefCell<Vec<RValue<'gcc>>> {
++ unimplemented!()
++ }
++
++ fn create_compiler_used_variable(&self) {
++ unimplemented!()
++ }
+}
+
+impl<'gcc, 'tcx> HasTyCtxt<'tcx> for CodegenCx<'gcc, 'tcx> {
+ fn tcx(&self) -> TyCtxt<'tcx> {
+ self.tcx
+ }
+}
+
+impl<'gcc, 'tcx> HasDataLayout for CodegenCx<'gcc, 'tcx> {
+ fn data_layout(&self) -> &TargetDataLayout {
+ &self.tcx.data_layout
+ }
+}
+
+impl<'gcc, 'tcx> HasTargetSpec for CodegenCx<'gcc, 'tcx> {
+ fn target_spec(&self) -> &Target {
+ &self.tcx.sess.target
+ }
+}
+
- fn layout_of(&self, ty: Ty<'tcx>) -> Self::TyAndLayout {
- self.spanned_layout_of(ty, DUMMY_SP)
- }
-
- fn spanned_layout_of(&self, ty: Ty<'tcx>, span: Span) -> Self::TyAndLayout {
- self.tcx.layout_of(ParamEnv::reveal_all().and(ty)).unwrap_or_else(|e| {
- if let LayoutError::SizeOverflow(_) = e {
- self.sess().span_fatal(span, &e.to_string())
- } else {
- bug!("failed to get layout for `{}`: {}", ty, e)
- }
- })
++impl<'gcc, 'tcx> LayoutOfHelpers<'tcx> for CodegenCx<'gcc, 'tcx> {
++ type LayoutOfResult = TyAndLayout<'tcx>;
+
++ #[inline]
++ fn handle_layout_err(&self, err: LayoutError<'tcx>, span: Span, ty: Ty<'tcx>) -> ! {
++ if let LayoutError::SizeOverflow(_) = err {
++ self.sess().span_fatal(span, &err.to_string())
++ } else {
++ span_bug!(span, "failed to get layout for `{}`: {}", ty, err)
++ }
+ }
+}
+
+impl<'tcx, 'gcc> HasParamEnv<'tcx> for CodegenCx<'gcc, 'tcx> {
+ fn param_env(&self) -> ParamEnv<'tcx> {
+ ParamEnv::reveal_all()
+ }
+}
+
+impl<'b, 'tcx> CodegenCx<'b, 'tcx> {
+ /// Generates a new symbol name with the given prefix. This symbol name must
+ /// only be used for definitions with `internal` or `private` linkage.
+ pub fn generate_local_symbol_name(&self, prefix: &str) -> String {
+ let idx = self.local_gen_sym_counter.get();
+ self.local_gen_sym_counter.set(idx + 1);
+ // Include a '.' character, so there can be no accidental conflicts with
+ // user defined names
+ let mut name = String::with_capacity(prefix.len() + 6);
+ name.push_str(prefix);
+ name.push_str(".");
+ base_n::push_str(idx as u128, base_n::ALPHANUMERIC_ONLY, &mut name);
+ name
+ }
+}
+
+pub fn unit_name<'tcx>(codegen_unit: &CodegenUnit<'tcx>) -> String {
+ let name = &codegen_unit.name().to_string();
+ mangle_name(&name.replace('-', "_"))
+}
+
+fn to_gcc_tls_mode(tls_model: TlsModel) -> gccjit::TlsModel {
+ match tls_model {
+ TlsModel::GeneralDynamic => gccjit::TlsModel::GlobalDynamic,
+ TlsModel::LocalDynamic => gccjit::TlsModel::LocalDynamic,
+ TlsModel::InitialExec => gccjit::TlsModel::InitialExec,
+ TlsModel::LocalExec => gccjit::TlsModel::LocalExec,
+ }
+}
--- /dev/null
- use rustc_target::abi::{HasDataLayout, LayoutOf};
+pub mod llvm;
+mod simd;
+
+use gccjit::{ComparisonOp, Function, RValue, ToRValue, Type, UnaryOp};
+use rustc_codegen_ssa::MemFlags;
+use rustc_codegen_ssa::base::wants_msvc_seh;
+use rustc_codegen_ssa::common::{IntPredicate, span_invalid_monomorphization_error};
+use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
+use rustc_codegen_ssa::mir::place::PlaceRef;
+use rustc_codegen_ssa::traits::{ArgAbiMethods, BaseTypeMethods, BuilderMethods, ConstMethods, IntrinsicCallMethods};
+use rustc_middle::bug;
+use rustc_middle::ty::{self, Instance, Ty};
++use rustc_middle::ty::layout::LayoutOf;
+use rustc_span::{Span, Symbol, symbol::kw, sym};
- self.block.expect("block").end_with_conditional(None, cond, then_block, else_block);
++use rustc_target::abi::HasDataLayout;
+use rustc_target::abi::call::{ArgAbi, FnAbi, PassMode};
+use rustc_target::spec::PanicStrategy;
+
+use crate::abi::GccType;
+use crate::builder::Builder;
+use crate::common::TypeReflection;
+use crate::context::CodegenCx;
+use crate::type_of::LayoutGccExt;
+use crate::intrinsic::simd::generic_simd_intrinsic;
+
+fn get_simple_intrinsic<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, name: Symbol) -> Option<Function<'gcc>> {
+ let gcc_name = match name {
+ sym::sqrtf32 => "sqrtf",
+ sym::sqrtf64 => "sqrt",
+ sym::powif32 => "__builtin_powif",
+ sym::powif64 => "__builtin_powi",
+ sym::sinf32 => "sinf",
+ sym::sinf64 => "sin",
+ sym::cosf32 => "cosf",
+ sym::cosf64 => "cos",
+ sym::powf32 => "powf",
+ sym::powf64 => "pow",
+ sym::expf32 => "expf",
+ sym::expf64 => "exp",
+ sym::exp2f32 => "exp2f",
+ sym::exp2f64 => "exp2",
+ sym::logf32 => "logf",
+ sym::logf64 => "log",
+ sym::log10f32 => "log10f",
+ sym::log10f64 => "log10",
+ sym::log2f32 => "log2f",
+ sym::log2f64 => "log2",
+ sym::fmaf32 => "fmaf",
+ sym::fmaf64 => "fma",
+ sym::fabsf32 => "fabsf",
+ sym::fabsf64 => "fabs",
+ sym::minnumf32 => "fminf",
+ sym::minnumf64 => "fmin",
+ sym::maxnumf32 => "fmaxf",
+ sym::maxnumf64 => "fmax",
+ sym::copysignf32 => "copysignf",
+ sym::copysignf64 => "copysign",
+ sym::floorf32 => "floorf",
+ sym::floorf64 => "floor",
+ sym::ceilf32 => "ceilf",
+ sym::ceilf64 => "ceil",
+ sym::truncf32 => "truncf",
+ sym::truncf64 => "trunc",
+ sym::rintf32 => "rintf",
+ sym::rintf64 => "rint",
+ sym::nearbyintf32 => "nearbyintf",
+ sym::nearbyintf64 => "nearbyint",
+ sym::roundf32 => "roundf",
+ sym::roundf64 => "round",
+ sym::abort => "abort",
+ _ => return None,
+ };
+ Some(cx.context.get_builtin_function(&gcc_name))
+}
+
+impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> {
+ fn codegen_intrinsic_call(&mut self, instance: Instance<'tcx>, fn_abi: &FnAbi<'tcx, Ty<'tcx>>, args: &[OperandRef<'tcx, RValue<'gcc>>], llresult: RValue<'gcc>, span: Span) {
+ let tcx = self.tcx;
+ let callee_ty = instance.ty(tcx, ty::ParamEnv::reveal_all());
+
+ let (def_id, substs) = match *callee_ty.kind() {
+ ty::FnDef(def_id, substs) => (def_id, substs),
+ _ => bug!("expected fn item type, found {}", callee_ty),
+ };
+
+ let sig = callee_ty.fn_sig(tcx);
+ let sig = tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), sig);
+ let arg_tys = sig.inputs();
+ let ret_ty = sig.output();
+ let name = tcx.item_name(def_id);
+ let name_str = &*name.as_str();
+
+ let llret_ty = self.layout_of(ret_ty).gcc_type(self, true);
+ let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
+
+ let simple = get_simple_intrinsic(self, name);
+ let llval =
+ match name {
+ _ if simple.is_some() => {
+ // FIXME(antoyo): remove this cast when the API supports function.
+ let func = unsafe { std::mem::transmute(simple.expect("simple")) };
+ self.call(self.type_void(), func, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None)
+ },
+ sym::likely => {
+ self.expect(args[0].immediate(), true)
+ }
+ sym::unlikely => {
+ self.expect(args[0].immediate(), false)
+ }
+ kw::Try => {
+ try_intrinsic(
+ self,
+ args[0].immediate(),
+ args[1].immediate(),
+ args[2].immediate(),
+ llresult,
+ );
+ return;
+ }
+ sym::breakpoint => {
+ unimplemented!();
+ }
+ sym::va_copy => {
+ unimplemented!();
+ }
+ sym::va_arg => {
+ unimplemented!();
+ }
+
+ sym::volatile_load | sym::unaligned_volatile_load => {
+ let tp_ty = substs.type_at(0);
+ let mut ptr = args[0].immediate();
+ if let PassMode::Cast(ty) = fn_abi.ret.mode {
+ ptr = self.pointercast(ptr, self.type_ptr_to(ty.gcc_type(self)));
+ }
+ let load = self.volatile_load(ptr.get_type(), ptr);
+ // TODO(antoyo): set alignment.
+ self.to_immediate(load, self.layout_of(tp_ty))
+ }
+ sym::volatile_store => {
+ let dst = args[0].deref(self.cx());
+ args[1].val.volatile_store(self, dst);
+ return;
+ }
+ sym::unaligned_volatile_store => {
+ let dst = args[0].deref(self.cx());
+ args[1].val.unaligned_volatile_store(self, dst);
+ return;
+ }
+ sym::prefetch_read_data
+ | sym::prefetch_write_data
+ | sym::prefetch_read_instruction
+ | sym::prefetch_write_instruction => {
+ unimplemented!();
+ }
+ sym::ctlz
+ | sym::ctlz_nonzero
+ | sym::cttz
+ | sym::cttz_nonzero
+ | sym::ctpop
+ | sym::bswap
+ | sym::bitreverse
+ | sym::rotate_left
+ | sym::rotate_right
+ | sym::saturating_add
+ | sym::saturating_sub => {
+ let ty = arg_tys[0];
+ match int_type_width_signed(ty, self) {
+ Some((width, signed)) => match name {
+ sym::ctlz | sym::cttz => {
+ let func = self.current_func.borrow().expect("func");
+ let then_block = func.new_block("then");
+ let else_block = func.new_block("else");
+ let after_block = func.new_block("after");
+
+ let arg = args[0].immediate();
+ let result = func.new_local(None, arg.get_type(), "zeros");
+ let zero = self.cx.context.new_rvalue_zero(arg.get_type());
+ let cond = self.cx.context.new_comparison(None, ComparisonOp::Equals, arg, zero);
- self.block.expect("block").end_with_conditional(None, overflow, then_block, after_block);
++ self.llbb().end_with_conditional(None, cond, then_block, else_block);
+
+ let zero_result = self.cx.context.new_rvalue_from_long(arg.get_type(), width as i64);
+ then_block.add_assignment(None, result, zero_result);
+ then_block.end_with_jump(None, after_block);
+
+ // NOTE: since jumps were added in a place
+ // count_leading_zeroes() does not expect, the current blocks
+ // in the state need to be updated.
+ *self.current_block.borrow_mut() = Some(else_block);
+ self.block = Some(else_block);
+
+ let zeros =
+ match name {
+ sym::ctlz => self.count_leading_zeroes(width, arg),
+ sym::cttz => self.count_trailing_zeroes(width, arg),
+ _ => unreachable!(),
+ };
+ else_block.add_assignment(None, result, zeros);
+ else_block.end_with_jump(None, after_block);
+
+ // NOTE: since jumps were added in a place rustc does not
+ // expect, the current blocks in the state need to be updated.
+ *self.current_block.borrow_mut() = Some(after_block);
+ self.block = Some(after_block);
+
+ result.to_rvalue()
+ }
+ sym::ctlz_nonzero => {
+ self.count_leading_zeroes(width, args[0].immediate())
+ },
+ sym::cttz_nonzero => {
+ self.count_trailing_zeroes(width, args[0].immediate())
+ }
+ sym::ctpop => self.pop_count(args[0].immediate()),
+ sym::bswap => {
+ if width == 8 {
+ args[0].immediate() // byte swap a u8/i8 is just a no-op
+ }
+ else {
+ // TODO(antoyo): check if it's faster to use string literals and a
+ // match instead of format!.
+ let bswap = self.cx.context.get_builtin_function(&format!("__builtin_bswap{}", width));
+ let mut arg = args[0].immediate();
+ // FIXME(antoyo): this cast should not be necessary. Remove
+ // when having proper sized integer types.
+ let param_type = bswap.get_param(0).to_rvalue().get_type();
+ if param_type != arg.get_type() {
+ arg = self.bitcast(arg, param_type);
+ }
+ self.cx.context.new_call(None, bswap, &[arg])
+ }
+ },
+ sym::bitreverse => self.bit_reverse(width, args[0].immediate()),
+ sym::rotate_left | sym::rotate_right => {
+ // TODO(antoyo): implement using algorithm from:
+ // https://blog.regehr.org/archives/1063
+ // for other platforms.
+ let is_left = name == sym::rotate_left;
+ let val = args[0].immediate();
+ let raw_shift = args[1].immediate();
+ if is_left {
+ self.rotate_left(val, raw_shift, width)
+ }
+ else {
+ self.rotate_right(val, raw_shift, width)
+ }
+ },
+ sym::saturating_add => {
+ self.saturating_add(args[0].immediate(), args[1].immediate(), signed, width)
+ },
+ sym::saturating_sub => {
+ self.saturating_sub(args[0].immediate(), args[1].immediate(), signed, width)
+ },
+ _ => bug!(),
+ },
+ None => {
+ span_invalid_monomorphization_error(
+ tcx.sess,
+ span,
+ &format!(
+ "invalid monomorphization of `{}` intrinsic: \
+ expected basic integer type, found `{}`",
+ name, ty
+ ),
+ );
+ return;
+ }
+ }
+ }
+
+ sym::raw_eq => {
+ use rustc_target::abi::Abi::*;
+ let tp_ty = substs.type_at(0);
+ let layout = self.layout_of(tp_ty).layout;
+ let use_integer_compare = match layout.abi {
+ Scalar(_) | ScalarPair(_, _) => true,
+ Uninhabited | Vector { .. } => false,
+ Aggregate { .. } => {
+ // For rusty ABIs, small aggregates are actually passed
+ // as `RegKind::Integer` (see `FnAbi::adjust_for_abi`),
+ // so we re-use that same threshold here.
+ layout.size <= self.data_layout().pointer_size * 2
+ }
+ };
+
+ let a = args[0].immediate();
+ let b = args[1].immediate();
+ if layout.size.bytes() == 0 {
+ self.const_bool(true)
+ }
+ /*else if use_integer_compare {
+ let integer_ty = self.type_ix(layout.size.bits()); // FIXME(antoyo): LLVM creates an integer of 96 bits for [i32; 3], but gcc doesn't support this, so it creates an integer of 128 bits.
+ let ptr_ty = self.type_ptr_to(integer_ty);
+ let a_ptr = self.bitcast(a, ptr_ty);
+ let a_val = self.load(integer_ty, a_ptr, layout.align.abi);
+ let b_ptr = self.bitcast(b, ptr_ty);
+ let b_val = self.load(integer_ty, b_ptr, layout.align.abi);
+ self.icmp(IntPredicate::IntEQ, a_val, b_val)
+ }*/
+ else {
+ let void_ptr_type = self.context.new_type::<*const ()>();
+ let a_ptr = self.bitcast(a, void_ptr_type);
+ let b_ptr = self.bitcast(b, void_ptr_type);
+ let n = self.context.new_cast(None, self.const_usize(layout.size.bytes()), self.sizet_type);
+ let builtin = self.context.get_builtin_function("memcmp");
+ let cmp = self.context.new_call(None, builtin, &[a_ptr, b_ptr, n]);
+ self.icmp(IntPredicate::IntEQ, cmp, self.const_i32(0))
+ }
+ }
+
++ sym::black_box => {
++ args[0].val.store(self, result);
++
++ let block = self.llbb();
++ let extended_asm = block.add_extended_asm(None, "");
++ extended_asm.add_input_operand(None, "r", result.llval);
++ extended_asm.add_clobber("memory");
++ extended_asm.set_volatile_flag(true);
++
++ // We have copied the value to `result` already.
++ return;
++ }
++
+ _ if name_str.starts_with("simd_") => {
+ match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
+ Ok(llval) => llval,
+ Err(()) => return,
+ }
+ }
+
+ _ => bug!("unknown intrinsic '{}'", name),
+ };
+
+ if !fn_abi.ret.is_ignore() {
+ if let PassMode::Cast(ty) = fn_abi.ret.mode {
+ let ptr_llty = self.type_ptr_to(ty.gcc_type(self));
+ let ptr = self.pointercast(result.llval, ptr_llty);
+ self.store(llval, ptr, result.align);
+ }
+ else {
+ OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
+ .val
+ .store(self, result);
+ }
+ }
+ }
+
+ fn abort(&mut self) {
+ let func = self.context.get_builtin_function("abort");
+ let func: RValue<'gcc> = unsafe { std::mem::transmute(func) };
+ self.call(self.type_void(), func, &[], None);
+ }
+
+ fn assume(&mut self, value: Self::Value) {
+ // TODO(antoyo): switch to asumme when it exists.
+ // Or use something like this:
+ // #define __assume(cond) do { if (!(cond)) __builtin_unreachable(); } while (0)
+ self.expect(value, true);
+ }
+
+ fn expect(&mut self, cond: Self::Value, _expected: bool) -> Self::Value {
+ // TODO(antoyo)
+ cond
+ }
+
+ fn sideeffect(&mut self) {
+ // TODO(antoyo)
+ }
+
+ fn va_start(&mut self, _va_list: RValue<'gcc>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+
+ fn va_end(&mut self, _va_list: RValue<'gcc>) -> RValue<'gcc> {
+ unimplemented!();
+ }
+}
+
+impl<'a, 'gcc, 'tcx> ArgAbiMethods<'tcx> for Builder<'a, 'gcc, 'tcx> {
+ fn store_fn_arg(&mut self, arg_abi: &ArgAbi<'tcx, Ty<'tcx>>, idx: &mut usize, dst: PlaceRef<'tcx, Self::Value>) {
+ arg_abi.store_fn_arg(self, idx, dst)
+ }
+
+ fn store_arg(&mut self, arg_abi: &ArgAbi<'tcx, Ty<'tcx>>, val: RValue<'gcc>, dst: PlaceRef<'tcx, RValue<'gcc>>) {
+ arg_abi.store(self, val, dst)
+ }
+
+ fn arg_memory_ty(&self, arg_abi: &ArgAbi<'tcx, Ty<'tcx>>) -> Type<'gcc> {
+ arg_abi.memory_ty(self)
+ }
+}
+
+pub trait ArgAbiExt<'gcc, 'tcx> {
+ fn memory_ty(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc>;
+ fn store(&self, bx: &mut Builder<'_, 'gcc, 'tcx>, val: RValue<'gcc>, dst: PlaceRef<'tcx, RValue<'gcc>>);
+ fn store_fn_arg(&self, bx: &mut Builder<'_, 'gcc, 'tcx>, idx: &mut usize, dst: PlaceRef<'tcx, RValue<'gcc>>);
+}
+
+impl<'gcc, 'tcx> ArgAbiExt<'gcc, 'tcx> for ArgAbi<'tcx, Ty<'tcx>> {
+ /// Gets the LLVM type for a place of the original Rust type of
+ /// this argument/return, i.e., the result of `type_of::type_of`.
+ fn memory_ty(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc> {
+ self.layout.gcc_type(cx, true)
+ }
+
+ /// Stores a direct/indirect value described by this ArgAbi into a
+ /// place for the original Rust type of this argument/return.
+ /// Can be used for both storing formal arguments into Rust variables
+ /// or results of call/invoke instructions into their destinations.
+ fn store(&self, bx: &mut Builder<'_, 'gcc, 'tcx>, val: RValue<'gcc>, dst: PlaceRef<'tcx, RValue<'gcc>>) {
+ if self.is_ignore() {
+ return;
+ }
+ if self.is_sized_indirect() {
+ OperandValue::Ref(val, None, self.layout.align.abi).store(bx, dst)
+ }
+ else if self.is_unsized_indirect() {
+ bug!("unsized `ArgAbi` must be handled through `store_fn_arg`");
+ }
+ else if let PassMode::Cast(cast) = self.mode {
+ // FIXME(eddyb): Figure out when the simpler Store is safe, clang
+ // uses it for i16 -> {i8, i8}, but not for i24 -> {i8, i8, i8}.
+ let can_store_through_cast_ptr = false;
+ if can_store_through_cast_ptr {
+ let cast_ptr_llty = bx.type_ptr_to(cast.gcc_type(bx));
+ let cast_dst = bx.pointercast(dst.llval, cast_ptr_llty);
+ bx.store(val, cast_dst, self.layout.align.abi);
+ }
+ else {
+ // The actual return type is a struct, but the ABI
+ // adaptation code has cast it into some scalar type. The
+ // code that follows is the only reliable way I have
+ // found to do a transform like i64 -> {i32,i32}.
+ // Basically we dump the data onto the stack then memcpy it.
+ //
+ // Other approaches I tried:
+ // - Casting rust ret pointer to the foreign type and using Store
+ // is (a) unsafe if size of foreign type > size of rust type and
+ // (b) runs afoul of strict aliasing rules, yielding invalid
+ // assembly under -O (specifically, the store gets removed).
+ // - Truncating foreign type to correct integral type and then
+ // bitcasting to the struct type yields invalid cast errors.
+
+ // We instead thus allocate some scratch space...
+ let scratch_size = cast.size(bx);
+ let scratch_align = cast.align(bx);
+ let llscratch = bx.alloca(cast.gcc_type(bx), scratch_align);
+ bx.lifetime_start(llscratch, scratch_size);
+
+ // ... where we first store the value...
+ bx.store(val, llscratch, scratch_align);
+
+ // ... and then memcpy it to the intended destination.
+ bx.memcpy(
+ dst.llval,
+ self.layout.align.abi,
+ llscratch,
+ scratch_align,
+ bx.const_usize(self.layout.size.bytes()),
+ MemFlags::empty(),
+ );
+
+ bx.lifetime_end(llscratch, scratch_size);
+ }
+ }
+ else {
+ OperandValue::Immediate(val).store(bx, dst);
+ }
+ }
+
+ fn store_fn_arg<'a>(&self, bx: &mut Builder<'a, 'gcc, 'tcx>, idx: &mut usize, dst: PlaceRef<'tcx, RValue<'gcc>>) {
+ let mut next = || {
+ let val = bx.current_func().get_param(*idx as i32);
+ *idx += 1;
+ val.to_rvalue()
+ };
+ match self.mode {
+ PassMode::Ignore => {}
+ PassMode::Pair(..) => {
+ OperandValue::Pair(next(), next()).store(bx, dst);
+ }
+ PassMode::Indirect { extra_attrs: Some(_), .. } => {
+ OperandValue::Ref(next(), Some(next()), self.layout.align.abi).store(bx, dst);
+ }
+ PassMode::Direct(_) | PassMode::Indirect { extra_attrs: None, .. } | PassMode::Cast(_) => {
+ let next_arg = next();
+ self.store(bx, next_arg.to_rvalue(), dst);
+ }
+ }
+ }
+}
+
+fn int_type_width_signed<'gcc, 'tcx>(ty: Ty<'tcx>, cx: &CodegenCx<'gcc, 'tcx>) -> Option<(u64, bool)> {
+ match ty.kind() {
+ ty::Int(t) => Some((
+ match t {
+ rustc_middle::ty::IntTy::Isize => u64::from(cx.tcx.sess.target.pointer_width),
+ rustc_middle::ty::IntTy::I8 => 8,
+ rustc_middle::ty::IntTy::I16 => 16,
+ rustc_middle::ty::IntTy::I32 => 32,
+ rustc_middle::ty::IntTy::I64 => 64,
+ rustc_middle::ty::IntTy::I128 => 128,
+ },
+ true,
+ )),
+ ty::Uint(t) => Some((
+ match t {
+ rustc_middle::ty::UintTy::Usize => u64::from(cx.tcx.sess.target.pointer_width),
+ rustc_middle::ty::UintTy::U8 => 8,
+ rustc_middle::ty::UintTy::U16 => 16,
+ rustc_middle::ty::UintTy::U32 => 32,
+ rustc_middle::ty::UintTy::U64 => 64,
+ rustc_middle::ty::UintTy::U128 => 128,
+ },
+ false,
+ )),
+ _ => None,
+ }
+}
+
+impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
+ fn bit_reverse(&mut self, width: u64, value: RValue<'gcc>) -> RValue<'gcc> {
+ let typ = value.get_type();
+ let context = &self.cx.context;
+ match width {
+ 8 => {
+ // First step.
+ let left = self.and(value, context.new_rvalue_from_int(typ, 0xF0));
+ let left = self.lshr(left, context.new_rvalue_from_int(typ, 4));
+ let right = self.and(value, context.new_rvalue_from_int(typ, 0x0F));
+ let right = self.shl(right, context.new_rvalue_from_int(typ, 4));
+ let step1 = self.or(left, right);
+
+ // Second step.
+ let left = self.and(step1, context.new_rvalue_from_int(typ, 0xCC));
+ let left = self.lshr(left, context.new_rvalue_from_int(typ, 2));
+ let right = self.and(step1, context.new_rvalue_from_int(typ, 0x33));
+ let right = self.shl(right, context.new_rvalue_from_int(typ, 2));
+ let step2 = self.or(left, right);
+
+ // Third step.
+ let left = self.and(step2, context.new_rvalue_from_int(typ, 0xAA));
+ let left = self.lshr(left, context.new_rvalue_from_int(typ, 1));
+ let right = self.and(step2, context.new_rvalue_from_int(typ, 0x55));
+ let right = self.shl(right, context.new_rvalue_from_int(typ, 1));
+ let step3 = self.or(left, right);
+
+ step3
+ },
+ 16 => {
+ // First step.
+ let left = self.and(value, context.new_rvalue_from_int(typ, 0x5555));
+ let left = self.shl(left, context.new_rvalue_from_int(typ, 1));
+ let right = self.and(value, context.new_rvalue_from_int(typ, 0xAAAA));
+ let right = self.lshr(right, context.new_rvalue_from_int(typ, 1));
+ let step1 = self.or(left, right);
+
+ // Second step.
+ let left = self.and(step1, context.new_rvalue_from_int(typ, 0x3333));
+ let left = self.shl(left, context.new_rvalue_from_int(typ, 2));
+ let right = self.and(step1, context.new_rvalue_from_int(typ, 0xCCCC));
+ let right = self.lshr(right, context.new_rvalue_from_int(typ, 2));
+ let step2 = self.or(left, right);
+
+ // Third step.
+ let left = self.and(step2, context.new_rvalue_from_int(typ, 0x0F0F));
+ let left = self.shl(left, context.new_rvalue_from_int(typ, 4));
+ let right = self.and(step2, context.new_rvalue_from_int(typ, 0xF0F0));
+ let right = self.lshr(right, context.new_rvalue_from_int(typ, 4));
+ let step3 = self.or(left, right);
+
+ // Fourth step.
+ let left = self.and(step3, context.new_rvalue_from_int(typ, 0x00FF));
+ let left = self.shl(left, context.new_rvalue_from_int(typ, 8));
+ let right = self.and(step3, context.new_rvalue_from_int(typ, 0xFF00));
+ let right = self.lshr(right, context.new_rvalue_from_int(typ, 8));
+ let step4 = self.or(left, right);
+
+ step4
+ },
+ 32 => {
+ // TODO(antoyo): Refactor with other implementations.
+ // First step.
+ let left = self.and(value, context.new_rvalue_from_long(typ, 0x55555555));
+ let left = self.shl(left, context.new_rvalue_from_long(typ, 1));
+ let right = self.and(value, context.new_rvalue_from_long(typ, 0xAAAAAAAA));
+ let right = self.lshr(right, context.new_rvalue_from_long(typ, 1));
+ let step1 = self.or(left, right);
+
+ // Second step.
+ let left = self.and(step1, context.new_rvalue_from_long(typ, 0x33333333));
+ let left = self.shl(left, context.new_rvalue_from_long(typ, 2));
+ let right = self.and(step1, context.new_rvalue_from_long(typ, 0xCCCCCCCC));
+ let right = self.lshr(right, context.new_rvalue_from_long(typ, 2));
+ let step2 = self.or(left, right);
+
+ // Third step.
+ let left = self.and(step2, context.new_rvalue_from_long(typ, 0x0F0F0F0F));
+ let left = self.shl(left, context.new_rvalue_from_long(typ, 4));
+ let right = self.and(step2, context.new_rvalue_from_long(typ, 0xF0F0F0F0));
+ let right = self.lshr(right, context.new_rvalue_from_long(typ, 4));
+ let step3 = self.or(left, right);
+
+ // Fourth step.
+ let left = self.and(step3, context.new_rvalue_from_long(typ, 0x00FF00FF));
+ let left = self.shl(left, context.new_rvalue_from_long(typ, 8));
+ let right = self.and(step3, context.new_rvalue_from_long(typ, 0xFF00FF00));
+ let right = self.lshr(right, context.new_rvalue_from_long(typ, 8));
+ let step4 = self.or(left, right);
+
+ // Fifth step.
+ let left = self.and(step4, context.new_rvalue_from_long(typ, 0x0000FFFF));
+ let left = self.shl(left, context.new_rvalue_from_long(typ, 16));
+ let right = self.and(step4, context.new_rvalue_from_long(typ, 0xFFFF0000));
+ let right = self.lshr(right, context.new_rvalue_from_long(typ, 16));
+ let step5 = self.or(left, right);
+
+ step5
+ },
+ 64 => {
+ // First step.
+ let left = self.shl(value, context.new_rvalue_from_long(typ, 32));
+ let right = self.lshr(value, context.new_rvalue_from_long(typ, 32));
+ let step1 = self.or(left, right);
+
+ // Second step.
+ let left = self.and(step1, context.new_rvalue_from_long(typ, 0x0001FFFF0001FFFF));
+ let left = self.shl(left, context.new_rvalue_from_long(typ, 15));
+ let right = self.and(step1, context.new_rvalue_from_long(typ, 0xFFFE0000FFFE0000u64 as i64)); // TODO(antoyo): transmute the number instead?
+ let right = self.lshr(right, context.new_rvalue_from_long(typ, 17));
+ let step2 = self.or(left, right);
+
+ // Third step.
+ let left = self.lshr(step2, context.new_rvalue_from_long(typ, 10));
+ let left = self.xor(step2, left);
+ let temp = self.and(left, context.new_rvalue_from_long(typ, 0x003F801F003F801F));
+
+ let left = self.shl(temp, context.new_rvalue_from_long(typ, 10));
+ let left = self.or(temp, left);
+ let step3 = self.xor(left, step2);
+
+ // Fourth step.
+ let left = self.lshr(step3, context.new_rvalue_from_long(typ, 4));
+ let left = self.xor(step3, left);
+ let temp = self.and(left, context.new_rvalue_from_long(typ, 0x0E0384210E038421));
+
+ let left = self.shl(temp, context.new_rvalue_from_long(typ, 4));
+ let left = self.or(temp, left);
+ let step4 = self.xor(left, step3);
+
+ // Fifth step.
+ let left = self.lshr(step4, context.new_rvalue_from_long(typ, 2));
+ let left = self.xor(step4, left);
+ let temp = self.and(left, context.new_rvalue_from_long(typ, 0x2248884222488842));
+
+ let left = self.shl(temp, context.new_rvalue_from_long(typ, 2));
+ let left = self.or(temp, left);
+ let step5 = self.xor(left, step4);
+
+ step5
+ },
+ 128 => {
+ // TODO(antoyo): find a more efficient implementation?
+ let sixty_four = self.context.new_rvalue_from_long(typ, 64);
+ let high = self.context.new_cast(None, value >> sixty_four, self.u64_type);
+ let low = self.context.new_cast(None, value, self.u64_type);
+
+ let reversed_high = self.bit_reverse(64, high);
+ let reversed_low = self.bit_reverse(64, low);
+
+ let new_low = self.context.new_cast(None, reversed_high, typ);
+ let new_high = self.context.new_cast(None, reversed_low, typ) << sixty_four;
+
+ new_low | new_high
+ },
+ _ => {
+ panic!("cannot bit reverse with width = {}", width);
+ },
+ }
+ }
+
+ fn count_leading_zeroes(&self, width: u64, arg: RValue<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): use width?
+ let arg_type = arg.get_type();
+ let count_leading_zeroes =
+ if arg_type.is_uint(&self.cx) {
+ "__builtin_clz"
+ }
+ else if arg_type.is_ulong(&self.cx) {
+ "__builtin_clzl"
+ }
+ else if arg_type.is_ulonglong(&self.cx) {
+ "__builtin_clzll"
+ }
+ else if width == 128 {
+ // Algorithm from: https://stackoverflow.com/a/28433850/389119
+ let array_type = self.context.new_array_type(None, arg_type, 3);
+ let result = self.current_func()
+ .new_local(None, array_type, "count_loading_zeroes_results");
+
+ let sixty_four = self.context.new_rvalue_from_long(arg_type, 64);
+ let high = self.context.new_cast(None, arg >> sixty_four, self.u64_type);
+ let low = self.context.new_cast(None, arg, self.u64_type);
+
+ let zero = self.context.new_rvalue_zero(self.usize_type);
+ let one = self.context.new_rvalue_one(self.usize_type);
+ let two = self.context.new_rvalue_from_long(self.usize_type, 2);
+
+ let clzll = self.context.get_builtin_function("__builtin_clzll");
+
+ let first_elem = self.context.new_array_access(None, result, zero);
+ let first_value = self.context.new_cast(None, self.context.new_call(None, clzll, &[high]), arg_type);
+ self.llbb()
+ .add_assignment(None, first_elem, first_value);
+
+ let second_elem = self.context.new_array_access(None, result, one);
+ let second_value = self.context.new_cast(None, self.context.new_call(None, clzll, &[low]), arg_type) + sixty_four;
+ self.llbb()
+ .add_assignment(None, second_elem, second_value);
+
+ let third_elem = self.context.new_array_access(None, result, two);
+ let third_value = self.context.new_rvalue_from_long(arg_type, 128);
+ self.llbb()
+ .add_assignment(None, third_elem, third_value);
+
+ let not_high = self.context.new_unary_op(None, UnaryOp::LogicalNegate, self.u64_type, high);
+ let not_low = self.context.new_unary_op(None, UnaryOp::LogicalNegate, self.u64_type, low);
+ let not_low_and_not_high = not_low & not_high;
+ let index = not_high + not_low_and_not_high;
+
+ let res = self.context.new_array_access(None, result, index);
+
+ return self.context.new_cast(None, res, arg_type);
+ }
+ else {
+ let count_leading_zeroes = self.context.get_builtin_function("__builtin_clz");
+ let arg = self.context.new_cast(None, arg, self.uint_type);
+ let diff = self.int_width(self.uint_type) - self.int_width(arg_type);
+ let diff = self.context.new_rvalue_from_long(self.int_type, diff);
+ let res = self.context.new_call(None, count_leading_zeroes, &[arg]) - diff;
+ return self.context.new_cast(None, res, arg_type);
+ };
+ let count_leading_zeroes = self.context.get_builtin_function(count_leading_zeroes);
+ let res = self.context.new_call(None, count_leading_zeroes, &[arg]);
+ self.context.new_cast(None, res, arg_type)
+ }
+
+ fn count_trailing_zeroes(&self, _width: u64, arg: RValue<'gcc>) -> RValue<'gcc> {
+ let arg_type = arg.get_type();
+ let (count_trailing_zeroes, expected_type) =
+ if arg_type.is_uchar(&self.cx) || arg_type.is_ushort(&self.cx) || arg_type.is_uint(&self.cx) {
+ // NOTE: we don't need to & 0xFF for uchar because the result is undefined on zero.
+ ("__builtin_ctz", self.cx.uint_type)
+ }
+ else if arg_type.is_ulong(&self.cx) {
+ ("__builtin_ctzl", self.cx.ulong_type)
+ }
+ else if arg_type.is_ulonglong(&self.cx) {
+ ("__builtin_ctzll", self.cx.ulonglong_type)
+ }
+ else if arg_type.is_u128(&self.cx) {
+ // Adapted from the algorithm to count leading zeroes from: https://stackoverflow.com/a/28433850/389119
+ let array_type = self.context.new_array_type(None, arg_type, 3);
+ let result = self.current_func()
+ .new_local(None, array_type, "count_loading_zeroes_results");
+
+ let sixty_four = self.context.new_rvalue_from_long(arg_type, 64);
+ let high = self.context.new_cast(None, arg >> sixty_four, self.u64_type);
+ let low = self.context.new_cast(None, arg, self.u64_type);
+
+ let zero = self.context.new_rvalue_zero(self.usize_type);
+ let one = self.context.new_rvalue_one(self.usize_type);
+ let two = self.context.new_rvalue_from_long(self.usize_type, 2);
+
+ let ctzll = self.context.get_builtin_function("__builtin_ctzll");
+
+ let first_elem = self.context.new_array_access(None, result, zero);
+ let first_value = self.context.new_cast(None, self.context.new_call(None, ctzll, &[low]), arg_type);
+ self.llbb()
+ .add_assignment(None, first_elem, first_value);
+
+ let second_elem = self.context.new_array_access(None, result, one);
+ let second_value = self.context.new_cast(None, self.context.new_call(None, ctzll, &[high]), arg_type) + sixty_four;
+ self.llbb()
+ .add_assignment(None, second_elem, second_value);
+
+ let third_elem = self.context.new_array_access(None, result, two);
+ let third_value = self.context.new_rvalue_from_long(arg_type, 128);
+ self.llbb()
+ .add_assignment(None, third_elem, third_value);
+
+ let not_low = self.context.new_unary_op(None, UnaryOp::LogicalNegate, self.u64_type, low);
+ let not_high = self.context.new_unary_op(None, UnaryOp::LogicalNegate, self.u64_type, high);
+ let not_low_and_not_high = not_low & not_high;
+ let index = not_low + not_low_and_not_high;
+
+ let res = self.context.new_array_access(None, result, index);
+
+ return self.context.new_cast(None, res, arg_type);
+ }
+ else {
+ unimplemented!("count_trailing_zeroes for {:?}", arg_type);
+ };
+ let count_trailing_zeroes = self.context.get_builtin_function(count_trailing_zeroes);
+ let arg =
+ if arg_type != expected_type {
+ self.context.new_cast(None, arg, expected_type)
+ }
+ else {
+ arg
+ };
+ let res = self.context.new_call(None, count_trailing_zeroes, &[arg]);
+ self.context.new_cast(None, res, arg_type)
+ }
+
+ fn int_width(&self, typ: Type<'gcc>) -> i64 {
+ self.cx.int_width(typ) as i64
+ }
+
+ fn pop_count(&self, value: RValue<'gcc>) -> RValue<'gcc> {
+ // TODO(antoyo): use the optimized version with fewer operations.
+ let value_type = value.get_type();
+
+ if value_type.is_u128(&self.cx) {
+ // TODO(antoyo): implement in the normal algorithm below to have a more efficient
+ // implementation (that does not require a call to __popcountdi2).
+ let popcount = self.context.get_builtin_function("__builtin_popcountll");
+ let sixty_four = self.context.new_rvalue_from_long(value_type, 64);
+ let high = self.context.new_cast(None, value >> sixty_four, self.cx.ulonglong_type);
+ let high = self.context.new_call(None, popcount, &[high]);
+ let low = self.context.new_cast(None, value, self.cx.ulonglong_type);
+ let low = self.context.new_call(None, popcount, &[low]);
+ return high + low;
+ }
+
+ // First step.
+ let mask = self.context.new_rvalue_from_long(value_type, 0x5555555555555555);
+ let left = value & mask;
+ let shifted = value >> self.context.new_rvalue_from_int(value_type, 1);
+ let right = shifted & mask;
+ let value = left + right;
+
+ // Second step.
+ let mask = self.context.new_rvalue_from_long(value_type, 0x3333333333333333);
+ let left = value & mask;
+ let shifted = value >> self.context.new_rvalue_from_int(value_type, 2);
+ let right = shifted & mask;
+ let value = left + right;
+
+ // Third step.
+ let mask = self.context.new_rvalue_from_long(value_type, 0x0F0F0F0F0F0F0F0F);
+ let left = value & mask;
+ let shifted = value >> self.context.new_rvalue_from_int(value_type, 4);
+ let right = shifted & mask;
+ let value = left + right;
+
+ if value_type.is_u8(&self.cx) {
+ return value;
+ }
+
+ // Fourth step.
+ let mask = self.context.new_rvalue_from_long(value_type, 0x00FF00FF00FF00FF);
+ let left = value & mask;
+ let shifted = value >> self.context.new_rvalue_from_int(value_type, 8);
+ let right = shifted & mask;
+ let value = left + right;
+
+ if value_type.is_u16(&self.cx) {
+ return value;
+ }
+
+ // Fifth step.
+ let mask = self.context.new_rvalue_from_long(value_type, 0x0000FFFF0000FFFF);
+ let left = value & mask;
+ let shifted = value >> self.context.new_rvalue_from_int(value_type, 16);
+ let right = shifted & mask;
+ let value = left + right;
+
+ if value_type.is_u32(&self.cx) {
+ return value;
+ }
+
+ // Sixth step.
+ let mask = self.context.new_rvalue_from_long(value_type, 0x00000000FFFFFFFF);
+ let left = value & mask;
+ let shifted = value >> self.context.new_rvalue_from_int(value_type, 32);
+ let right = shifted & mask;
+ let value = left + right;
+
+ value
+ }
+
+ // Algorithm from: https://blog.regehr.org/archives/1063
+ fn rotate_left(&mut self, value: RValue<'gcc>, shift: RValue<'gcc>, width: u64) -> RValue<'gcc> {
+ let max = self.context.new_rvalue_from_long(shift.get_type(), width as i64);
+ let shift = shift % max;
+ let lhs = self.shl(value, shift);
+ let result_and =
+ self.and(
+ self.context.new_unary_op(None, UnaryOp::Minus, shift.get_type(), shift),
+ self.context.new_rvalue_from_long(shift.get_type(), width as i64 - 1),
+ );
+ let rhs = self.lshr(value, result_and);
+ self.or(lhs, rhs)
+ }
+
+ // Algorithm from: https://blog.regehr.org/archives/1063
+ fn rotate_right(&mut self, value: RValue<'gcc>, shift: RValue<'gcc>, width: u64) -> RValue<'gcc> {
+ let max = self.context.new_rvalue_from_long(shift.get_type(), width as i64);
+ let shift = shift % max;
+ let lhs = self.lshr(value, shift);
+ let result_and =
+ self.and(
+ self.context.new_unary_op(None, UnaryOp::Minus, shift.get_type(), shift),
+ self.context.new_rvalue_from_long(shift.get_type(), width as i64 - 1),
+ );
+ let rhs = self.shl(value, result_and);
+ self.or(lhs, rhs)
+ }
+
+ fn saturating_add(&mut self, lhs: RValue<'gcc>, rhs: RValue<'gcc>, signed: bool, width: u64) -> RValue<'gcc> {
+ let func = self.current_func.borrow().expect("func");
+
+ if signed {
+ // Algorithm from: https://stackoverflow.com/a/56531252/389119
+ let after_block = func.new_block("after");
+ let func_name =
+ match width {
+ 8 => "__builtin_add_overflow",
+ 16 => "__builtin_add_overflow",
+ 32 => "__builtin_sadd_overflow",
+ 64 => "__builtin_saddll_overflow",
+ 128 => "__builtin_add_overflow",
+ _ => unreachable!(),
+ };
+ let overflow_func = self.context.get_builtin_function(func_name);
+ let result_type = lhs.get_type();
+ let res = func.new_local(None, result_type, "saturating_sum");
+ let overflow = self.overflow_call(overflow_func, &[lhs, rhs, res.get_address(None)], None);
+
+ let then_block = func.new_block("then");
+
+ let unsigned_type = self.context.new_int_type(width as i32 / 8, false);
+ let shifted = self.context.new_cast(None, lhs, unsigned_type) >> self.context.new_rvalue_from_int(unsigned_type, width as i32 - 1);
+ let uint_max = self.context.new_unary_op(None, UnaryOp::BitwiseNegate, unsigned_type,
+ self.context.new_rvalue_from_int(unsigned_type, 0)
+ );
+ let int_max = uint_max >> self.context.new_rvalue_one(unsigned_type);
+ then_block.add_assignment(None, res, self.context.new_cast(None, shifted + int_max, result_type));
+ then_block.end_with_jump(None, after_block);
+
- self.block.expect("block").end_with_conditional(None, overflow, then_block, after_block);
++ self.llbb().end_with_conditional(None, overflow, then_block, after_block);
+
+ // NOTE: since jumps were added in a place rustc does not
+ // expect, the current blocks in the state need to be updated.
+ *self.current_block.borrow_mut() = Some(after_block);
+ self.block = Some(after_block);
+
+ res.to_rvalue()
+ }
+ else {
+ // Algorithm from: http://locklessinc.com/articles/sat_arithmetic/
+ let res = lhs + rhs;
+ let res_type = res.get_type();
+ let cond = self.context.new_comparison(None, ComparisonOp::LessThan, res, lhs);
+ let value = self.context.new_unary_op(None, UnaryOp::Minus, res_type, self.context.new_cast(None, cond, res_type));
+ res | value
+ }
+ }
+
+ // Algorithm from: https://locklessinc.com/articles/sat_arithmetic/
+ fn saturating_sub(&mut self, lhs: RValue<'gcc>, rhs: RValue<'gcc>, signed: bool, width: u64) -> RValue<'gcc> {
+ if signed {
+ // Also based on algorithm from: https://stackoverflow.com/a/56531252/389119
+ let func_name =
+ match width {
+ 8 => "__builtin_sub_overflow",
+ 16 => "__builtin_sub_overflow",
+ 32 => "__builtin_ssub_overflow",
+ 64 => "__builtin_ssubll_overflow",
+ 128 => "__builtin_sub_overflow",
+ _ => unreachable!(),
+ };
+ let overflow_func = self.context.get_builtin_function(func_name);
+ let result_type = lhs.get_type();
+ let func = self.current_func.borrow().expect("func");
+ let res = func.new_local(None, result_type, "saturating_diff");
+ let overflow = self.overflow_call(overflow_func, &[lhs, rhs, res.get_address(None)], None);
+
+ let then_block = func.new_block("then");
+ let after_block = func.new_block("after");
+
+ let unsigned_type = self.context.new_int_type(width as i32 / 8, false);
+ let shifted = self.context.new_cast(None, lhs, unsigned_type) >> self.context.new_rvalue_from_int(unsigned_type, width as i32 - 1);
+ let uint_max = self.context.new_unary_op(None, UnaryOp::BitwiseNegate, unsigned_type,
+ self.context.new_rvalue_from_int(unsigned_type, 0)
+ );
+ let int_max = uint_max >> self.context.new_rvalue_one(unsigned_type);
+ then_block.add_assignment(None, res, self.context.new_cast(None, shifted + int_max, result_type));
+ then_block.end_with_jump(None, after_block);
+
++ self.llbb().end_with_conditional(None, overflow, then_block, after_block);
+
+ // NOTE: since jumps were added in a place rustc does not
+ // expect, the current blocks in the state need to be updated.
+ *self.current_block.borrow_mut() = Some(after_block);
+ self.block = Some(after_block);
+
+ res.to_rvalue()
+ }
+ else {
+ let res = lhs - rhs;
+ let comparison = self.context.new_comparison(None, ComparisonOp::LessThanEquals, res, lhs);
+ let comparison = self.context.new_cast(None, comparison, lhs.get_type());
+ let unary_op = self.context.new_unary_op(None, UnaryOp::Minus, comparison.get_type(), comparison);
+ self.and(res, unary_op)
+ }
+ }
+}
+
+fn try_intrinsic<'gcc, 'tcx>(bx: &mut Builder<'_, 'gcc, 'tcx>, try_func: RValue<'gcc>, data: RValue<'gcc>, _catch_func: RValue<'gcc>, dest: RValue<'gcc>) {
+ if bx.sess().panic_strategy() == PanicStrategy::Abort {
+ bx.call(bx.type_void(), try_func, &[data], None);
+ // Return 0 unconditionally from the intrinsic call;
+ // we can never unwind.
+ let ret_align = bx.tcx.data_layout.i32_align.abi;
+ bx.store(bx.const_i32(0), dest, ret_align);
+ }
+ else if wants_msvc_seh(bx.sess()) {
+ unimplemented!();
+ }
+ else {
+ unimplemented!();
+ }
+}
--- /dev/null
- extern crate rustc_mir;
+/*
+ * TODO(antoyo): support #[inline] attributes.
+ * TODO(antoyo): support LTO.
+ *
+ * TODO(antoyo): remove the patches.
+ */
+
+#![feature(rustc_private, decl_macro, associated_type_bounds, never_type, trusted_len)]
+#![allow(broken_intra_doc_links)]
+#![recursion_limit="256"]
+#![warn(rust_2018_idioms)]
+#![warn(unused_lifetimes)]
+
+extern crate rustc_ast;
+extern crate rustc_codegen_ssa;
+extern crate rustc_data_structures;
+extern crate rustc_errors;
+extern crate rustc_hir;
+extern crate rustc_metadata;
+extern crate rustc_middle;
- fn codegen_allocator<'tcx>(&self, tcx: TyCtxt<'tcx>, mods: &mut Self::Module, kind: AllocatorKind, has_alloc_error_handler: bool) {
- unsafe { allocator::codegen(tcx, mods, kind, has_alloc_error_handler) }
+extern crate rustc_session;
+extern crate rustc_span;
+extern crate rustc_symbol_mangling;
+extern crate rustc_target;
+extern crate snap;
+
+// This prevents duplicating functions and statics that are already part of the host rustc process.
+#[allow(unused_extern_crates)]
+extern crate rustc_driver;
+
+mod abi;
+mod allocator;
+mod archive;
+mod asm;
+mod back;
+mod base;
+mod builder;
+mod callee;
+mod common;
+mod consts;
+mod context;
+mod coverageinfo;
+mod debuginfo;
+mod declare;
+mod intrinsic;
+mod mangled_std_symbols;
+mod mono_item;
+mod type_;
+mod type_of;
+
+use std::any::Any;
+use std::sync::Arc;
+
+use gccjit::{Block, Context, FunctionType, OptimizationLevel};
+use rustc_ast::expand::allocator::AllocatorKind;
+use rustc_codegen_ssa::{CodegenResults, CompiledModule, ModuleCodegen};
+use rustc_codegen_ssa::base::codegen_crate;
+use rustc_codegen_ssa::back::write::{CodegenContext, FatLTOInput, ModuleConfig, TargetMachineFactoryFn};
+use rustc_codegen_ssa::back::lto::{LtoModuleCodegen, SerializedModule, ThinModule};
+use rustc_codegen_ssa::target_features::supported_target_features;
+use rustc_codegen_ssa::traits::{CodegenBackend, ExtraBackendMethods, ModuleBufferMethods, ThinBufferMethods, WriteBackendMethods};
+use rustc_data_structures::fx::FxHashMap;
+use rustc_errors::{ErrorReported, Handler};
+use rustc_middle::dep_graph::{WorkProduct, WorkProductId};
+use rustc_middle::middle::cstore::EncodedMetadata;
+use rustc_middle::ty::TyCtxt;
+use rustc_session::config::{CrateType, Lto, OptLevel, OutputFilenames};
+use rustc_session::Session;
+use rustc_span::Symbol;
+use rustc_span::fatal_error::FatalError;
+
+use crate::context::unit_name;
+
+pub struct PrintOnPanic<F: Fn() -> String>(pub F);
+
+impl<F: Fn() -> String> Drop for PrintOnPanic<F> {
+ fn drop(&mut self) {
+ if ::std::thread::panicking() {
+ println!("{}", (self.0)());
+ }
+ }
+}
+
+#[derive(Clone)]
+pub struct GccCodegenBackend;
+
+impl CodegenBackend for GccCodegenBackend {
+ fn init(&self, sess: &Session) {
+ if sess.lto() != Lto::No {
+ sess.warn("LTO is not supported. You may get a linker error.");
+ }
+ }
+
+ fn codegen_crate<'tcx>(&self, tcx: TyCtxt<'tcx>, metadata: EncodedMetadata, need_metadata_module: bool) -> Box<dyn Any> {
+ let target_cpu = target_cpu(tcx.sess);
+ let res = codegen_crate(self.clone(), tcx, target_cpu.to_string(), metadata, need_metadata_module);
+
+ rustc_symbol_mangling::test::report_symbol_names(tcx);
+
+ Box::new(res)
+ }
+
+ fn join_codegen(&self, ongoing_codegen: Box<dyn Any>, sess: &Session) -> Result<(CodegenResults, FxHashMap<WorkProductId, WorkProduct>), ErrorReported> {
+ let (codegen_results, work_products) = ongoing_codegen
+ .downcast::<rustc_codegen_ssa::back::write::OngoingCodegen<GccCodegenBackend>>()
+ .expect("Expected GccCodegenBackend's OngoingCodegen, found Box<Any>")
+ .join(sess);
+
+ Ok((codegen_results, work_products))
+ }
+
+ fn link(&self, sess: &Session, mut codegen_results: CodegenResults, outputs: &OutputFilenames) -> Result<(), ErrorReported> {
+ use rustc_codegen_ssa::back::link::link_binary;
+ if let Some(symbols) = codegen_results.crate_info.exported_symbols.get_mut(&CrateType::Dylib) {
+ // TODO:(antoyo): remove when global initializer work without calling a function at runtime.
+ // HACK: since this codegen add some symbols (e.g. __gccGlobalCrateInit) and the UI
+ // tests load libstd.so as a dynamic library, and rustc use a version-script to specify
+ // the symbols visibility, we add * to export all symbols.
+ // It seems other symbols from libstd/libcore are causing some issues here as well.
+ symbols.push("*".to_string());
+ }
+
+ link_binary::<crate::archive::ArArchiveBuilder<'_>>(
+ sess,
+ &codegen_results,
+ outputs,
+ )
+ }
+
+ fn target_features(&self, sess: &Session) -> Vec<Symbol> {
+ target_features(sess)
+ }
+}
+
+impl ExtraBackendMethods for GccCodegenBackend {
+ fn new_metadata<'tcx>(&self, _tcx: TyCtxt<'tcx>, _mod_name: &str) -> Self::Module {
+ GccContext {
+ context: Context::default(),
+ }
+ }
+
+ fn write_compressed_metadata<'tcx>(&self, tcx: TyCtxt<'tcx>, metadata: &EncodedMetadata, gcc_module: &mut Self::Module) {
+ base::write_compressed_metadata(tcx, metadata, gcc_module)
+ }
+
++ fn codegen_allocator<'tcx>(&self, tcx: TyCtxt<'tcx>, mods: &mut Self::Module, module_name: &str, kind: AllocatorKind, has_alloc_error_handler: bool) {
++ unsafe { allocator::codegen(tcx, mods, module_name, kind, has_alloc_error_handler) }
+ }
+
+ fn compile_codegen_unit<'tcx>(&self, tcx: TyCtxt<'tcx>, cgu_name: Symbol) -> (ModuleCodegen<Self::Module>, u64) {
+ base::compile_codegen_unit(tcx, cgu_name)
+ }
+
+ fn target_machine_factory(&self, _sess: &Session, _opt_level: OptLevel) -> TargetMachineFactoryFn<Self> {
+ // TODO(antoyo): set opt level.
+ Arc::new(|_| {
+ Ok(())
+ })
+ }
+
+ fn target_cpu<'b>(&self, _sess: &'b Session) -> &'b str {
+ unimplemented!();
+ }
+
+ fn tune_cpu<'b>(&self, _sess: &'b Session) -> Option<&'b str> {
+ None
+ // TODO(antoyo)
+ }
+}
+
+pub struct ModuleBuffer;
+
+impl ModuleBufferMethods for ModuleBuffer {
+ fn data(&self) -> &[u8] {
+ unimplemented!();
+ }
+}
+
+pub struct ThinBuffer;
+
+impl ThinBufferMethods for ThinBuffer {
+ fn data(&self) -> &[u8] {
+ unimplemented!();
+ }
+}
+
+pub struct GccContext {
+ context: Context<'static>,
+}
+
+unsafe impl Send for GccContext {}
+// FIXME(antoyo): that shouldn't be Sync. Parallel compilation is currently disabled with "-Zno-parallel-llvm". Try to disable it here.
+unsafe impl Sync for GccContext {}
+
+impl WriteBackendMethods for GccCodegenBackend {
+ type Module = GccContext;
+ type TargetMachine = ();
+ type ModuleBuffer = ModuleBuffer;
+ type Context = ();
+ type ThinData = ();
+ type ThinBuffer = ThinBuffer;
+
+ fn run_fat_lto(_cgcx: &CodegenContext<Self>, mut modules: Vec<FatLTOInput<Self>>, _cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>) -> Result<LtoModuleCodegen<Self>, FatalError> {
+ // TODO(antoyo): implement LTO by sending -flto to libgccjit and adding the appropriate gcc linker plugins.
+ // NOTE: implemented elsewhere.
+ let module =
+ match modules.remove(0) {
+ FatLTOInput::InMemory(module) => module,
+ FatLTOInput::Serialized { .. } => {
+ unimplemented!();
+ }
+ };
+ Ok(LtoModuleCodegen::Fat { module: Some(module), _serialized_bitcode: vec![] })
+ }
+
+ fn run_thin_lto(_cgcx: &CodegenContext<Self>, _modules: Vec<(String, Self::ThinBuffer)>, _cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>) -> Result<(Vec<LtoModuleCodegen<Self>>, Vec<WorkProduct>), FatalError> {
+ unimplemented!();
+ }
+
+ fn print_pass_timings(&self) {
+ unimplemented!();
+ }
+
+ unsafe fn optimize(_cgcx: &CodegenContext<Self>, _diag_handler: &Handler, module: &ModuleCodegen<Self::Module>, config: &ModuleConfig) -> Result<(), FatalError> {
+ module.module_llvm.context.set_optimization_level(to_gcc_opt_level(config.opt_level));
+ Ok(())
+ }
+
+ unsafe fn optimize_thin(_cgcx: &CodegenContext<Self>, _thin: &mut ThinModule<Self>) -> Result<ModuleCodegen<Self::Module>, FatalError> {
+ unimplemented!();
+ }
+
+ unsafe fn codegen(cgcx: &CodegenContext<Self>, diag_handler: &Handler, module: ModuleCodegen<Self::Module>, config: &ModuleConfig) -> Result<CompiledModule, FatalError> {
+ back::write::codegen(cgcx, diag_handler, module, config)
+ }
+
+ fn prepare_thin(_module: ModuleCodegen<Self::Module>) -> (String, Self::ThinBuffer) {
+ unimplemented!();
+ }
+
+ fn serialize_module(_module: ModuleCodegen<Self::Module>) -> (String, Self::ModuleBuffer) {
+ unimplemented!();
+ }
+
+ fn run_lto_pass_manager(_cgcx: &CodegenContext<Self>, _module: &ModuleCodegen<Self::Module>, _config: &ModuleConfig, _thin: bool) -> Result<(), FatalError> {
+ // TODO(antoyo)
+ Ok(())
+ }
+
+ fn run_link(cgcx: &CodegenContext<Self>, diag_handler: &Handler, modules: Vec<ModuleCodegen<Self::Module>>) -> Result<ModuleCodegen<Self::Module>, FatalError> {
+ back::write::link(cgcx, diag_handler, modules)
+ }
+}
+
+/// This is the entrypoint for a hot plugged rustc_codegen_gccjit
+#[no_mangle]
+pub fn __rustc_codegen_backend() -> Box<dyn CodegenBackend> {
+ Box::new(GccCodegenBackend)
+}
+
+fn to_gcc_opt_level(optlevel: Option<OptLevel>) -> OptimizationLevel {
+ match optlevel {
+ None => OptimizationLevel::None,
+ Some(level) => {
+ match level {
+ OptLevel::No => OptimizationLevel::None,
+ OptLevel::Less => OptimizationLevel::Limited,
+ OptLevel::Default => OptimizationLevel::Standard,
+ OptLevel::Aggressive => OptimizationLevel::Aggressive,
+ OptLevel::Size | OptLevel::SizeMin => OptimizationLevel::Limited,
+ }
+ },
+ }
+}
+
+fn create_function_calling_initializers<'gcc, 'tcx>(tcx: TyCtxt<'tcx>, context: &Context<'gcc>, block: Block<'gcc>) {
+ let codegen_units = tcx.collect_and_partition_mono_items(()).1;
+ for codegen_unit in codegen_units {
+ let codegen_init_func = context.new_function(None, FunctionType::Extern, context.new_type::<()>(), &[],
+ &format!("__gccGlobalInit{}", unit_name(&codegen_unit)), false);
+ block.add_eval(None, context.new_call(None, codegen_init_func, &[]));
+ }
+}
+
+fn handle_native(name: &str) -> &str {
+ if name != "native" {
+ return name;
+ }
+
+ unimplemented!();
+}
+
+pub fn target_cpu(sess: &Session) -> &str {
+ let name = sess.opts.cg.target_cpu.as_ref().unwrap_or(&sess.target.cpu);
+ handle_native(name)
+}
+
+pub fn target_features(sess: &Session) -> Vec<Symbol> {
+ supported_target_features(sess)
+ .iter()
+ .filter_map(
+ |&(feature, gate)| {
+ if sess.is_nightly_build() || gate.is_none() { Some(feature) } else { None }
+ },
+ )
+ .filter(|_feature| {
+ // TODO(antoyo): implement a way to get enabled feature in libgccjit.
+ false
+ })
+ .map(|feature| Symbol::intern(feature))
+ .collect()
+}
--- /dev/null
- use rustc_middle::ty::layout::FnAbiExt;
+use rustc_codegen_ssa::traits::PreDefineMethods;
+use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrFlags;
+use rustc_middle::mir::mono::{Linkage, Visibility};
+use rustc_middle::ty::{self, Instance, TypeFoldable};
- use rustc_target::abi::LayoutOf;
++use rustc_middle::ty::layout::{FnAbiExt, LayoutOf};
+use rustc_span::def_id::DefId;
- assert!(!instance.substs.needs_infer() && !instance.substs.has_param_types_or_consts());
+use rustc_target::abi::call::FnAbi;
+
+use crate::base;
+use crate::context::CodegenCx;
+use crate::type_of::LayoutGccExt;
+
+impl<'gcc, 'tcx> PreDefineMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
+ fn predefine_static(&self, def_id: DefId, _linkage: Linkage, _visibility: Visibility, symbol_name: &str) {
+ let attrs = self.tcx.codegen_fn_attrs(def_id);
+ let instance = Instance::mono(self.tcx, def_id);
+ let ty = instance.ty(self.tcx, ty::ParamEnv::reveal_all());
+ let gcc_type = self.layout_of(ty).gcc_type(self, true);
+
+ let is_tls = attrs.flags.contains(CodegenFnAttrFlags::THREAD_LOCAL);
+ let global = self.define_global(symbol_name, gcc_type, is_tls, attrs.link_section).unwrap_or_else(|| {
+ self.sess().span_fatal(
+ self.tcx.def_span(def_id),
+ &format!("symbol `{}` is already defined", symbol_name),
+ )
+ });
+
+ // TODO(antoyo): set linkage and visibility.
+ self.instances.borrow_mut().insert(instance, global);
+ }
+
+ fn predefine_fn(&self, instance: Instance<'tcx>, linkage: Linkage, _visibility: Visibility, symbol_name: &str) {
++ assert!(!instance.substs.needs_infer());
+
+ let fn_abi = FnAbi::of_instance(self, instance, &[]);
+ self.linkage.set(base::linkage_to_gcc(linkage));
+ let _decl = self.declare_fn(symbol_name, &fn_abi);
+ //let attrs = self.tcx.codegen_fn_attrs(instance.def_id());
+
+ // TODO(antoyo): call set_link_section() to allow initializing argc/argv.
+ // TODO(antoyo): set unique comdat.
+ // TODO(antoyo): use inline attribute from there in linkage.set() above.
+ }
+}
--- /dev/null
- use rustc_middle::ty::layout::{FnAbiExt, TyAndLayout};
+use std::fmt::Write;
+
+use gccjit::{Struct, Type};
+use crate::rustc_codegen_ssa::traits::{BaseTypeMethods, DerivedTypeMethods, LayoutTypeMethods};
+use rustc_middle::bug;
+use rustc_middle::ty::{self, Ty, TypeFoldable};
- use rustc_target::abi::{self, Abi, F32, F64, FieldsShape, Int, Integer, LayoutOf, Pointer, PointeeInfo, Size, TyAndLayoutMethods, Variants};
++use rustc_middle::ty::layout::{FnAbiExt, LayoutOf, TyAndLayout};
+use rustc_middle::ty::print::with_no_trimmed_paths;
- let result = Ty::pointee_info_at(*self, cx, offset);
++use rustc_target::abi::{self, Abi, F32, F64, FieldsShape, Int, Integer, Pointer, PointeeInfo, Size, TyAbiInterface, Variants};
+use rustc_target::abi::call::{CastTarget, FnAbi, Reg};
+
+use crate::abi::{FnAbiGccExt, GccType};
+use crate::context::CodegenCx;
+use crate::type_::struct_fields;
+
+impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
+ fn type_from_unsigned_integer(&self, i: Integer) -> Type<'gcc> {
+ use Integer::*;
+ match i {
+ I8 => self.type_u8(),
+ I16 => self.type_u16(),
+ I32 => self.type_u32(),
+ I64 => self.type_u64(),
+ I128 => self.type_u128(),
+ }
+ }
+}
+
+pub fn uncached_gcc_type<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, layout: TyAndLayout<'tcx>, defer: &mut Option<(Struct<'gcc>, TyAndLayout<'tcx>)>) -> Type<'gcc> {
+ match layout.abi {
+ Abi::Scalar(_) => bug!("handled elsewhere"),
+ Abi::Vector { ref element, count } => {
+ let element = layout.scalar_gcc_type_at(cx, element, Size::ZERO);
+ return cx.context.new_vector_type(element, count);
+ },
+ Abi::ScalarPair(..) => {
+ return cx.type_struct(
+ &[
+ layout.scalar_pair_element_gcc_type(cx, 0, false),
+ layout.scalar_pair_element_gcc_type(cx, 1, false),
+ ],
+ false,
+ );
+ }
+ Abi::Uninhabited | Abi::Aggregate { .. } => {}
+ }
+
+ let name = match layout.ty.kind() {
+ // FIXME(eddyb) producing readable type names for trait objects can result
+ // in problematically distinct types due to HRTB and subtyping (see #47638).
+ // ty::Dynamic(..) |
+ ty::Adt(..) | ty::Closure(..) | ty::Foreign(..) | ty::Generator(..) | ty::Str
+ if !cx.sess().fewer_names() =>
+ {
+ let mut name = with_no_trimmed_paths(|| layout.ty.to_string());
+ if let (&ty::Adt(def, _), &Variants::Single { index }) =
+ (layout.ty.kind(), &layout.variants)
+ {
+ if def.is_enum() && !def.variants.is_empty() {
+ write!(&mut name, "::{}", def.variants[index].ident).unwrap();
+ }
+ }
+ if let (&ty::Generator(_, _, _), &Variants::Single { index }) =
+ (layout.ty.kind(), &layout.variants)
+ {
+ write!(&mut name, "::{}", ty::GeneratorSubsts::variant_name(index)).unwrap();
+ }
+ Some(name)
+ }
+ ty::Adt(..) => {
+ // If `Some` is returned then a named struct is created in LLVM. Name collisions are
+ // avoided by LLVM (with increasing suffixes). If rustc doesn't generate names then that
+ // can improve perf.
+ // FIXME(antoyo): I don't think that's true for libgccjit.
+ Some(String::new())
+ }
+ _ => None,
+ };
+
+ match layout.fields {
+ FieldsShape::Primitive | FieldsShape::Union(_) => {
+ let fill = cx.type_padding_filler(layout.size, layout.align.abi);
+ let packed = false;
+ match name {
+ None => cx.type_struct(&[fill], packed),
+ Some(ref name) => {
+ let gcc_type = cx.type_named_struct(name);
+ cx.set_struct_body(gcc_type, &[fill], packed);
+ gcc_type.as_type()
+ },
+ }
+ }
+ FieldsShape::Array { count, .. } => cx.type_array(layout.field(cx, 0).gcc_type(cx, true), count),
+ FieldsShape::Arbitrary { .. } =>
+ match name {
+ None => {
+ let (gcc_fields, packed) = struct_fields(cx, layout);
+ cx.type_struct(&gcc_fields, packed)
+ },
+ Some(ref name) => {
+ let gcc_type = cx.type_named_struct(name);
+ *defer = Some((gcc_type, layout));
+ gcc_type.as_type()
+ },
+ },
+ }
+}
+
+pub trait LayoutGccExt<'tcx> {
+ fn is_gcc_immediate(&self) -> bool;
+ fn is_gcc_scalar_pair(&self) -> bool;
+ fn gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, set_fields: bool) -> Type<'gcc>;
+ fn immediate_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc>;
+ fn scalar_gcc_type_at<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, scalar: &abi::Scalar, offset: Size) -> Type<'gcc>;
+ fn scalar_pair_element_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, index: usize, immediate: bool) -> Type<'gcc>;
+ fn gcc_field_index(&self, index: usize) -> u64;
+ fn pointee_info_at<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, offset: Size) -> Option<PointeeInfo>;
+}
+
+impl<'tcx> LayoutGccExt<'tcx> for TyAndLayout<'tcx> {
+ fn is_gcc_immediate(&self) -> bool {
+ match self.abi {
+ Abi::Scalar(_) | Abi::Vector { .. } => true,
+ Abi::ScalarPair(..) => false,
+ Abi::Uninhabited | Abi::Aggregate { .. } => self.is_zst(),
+ }
+ }
+
+ fn is_gcc_scalar_pair(&self) -> bool {
+ match self.abi {
+ Abi::ScalarPair(..) => true,
+ Abi::Uninhabited | Abi::Scalar(_) | Abi::Vector { .. } | Abi::Aggregate { .. } => false,
+ }
+ }
+
+ /// Gets the GCC type corresponding to a Rust type, i.e., `rustc_middle::ty::Ty`.
+ /// The pointee type of the pointer in `PlaceRef` is always this type.
+ /// For sized types, it is also the right LLVM type for an `alloca`
+ /// containing a value of that type, and most immediates (except `bool`).
+ /// Unsized types, however, are represented by a "minimal unit", e.g.
+ /// `[T]` becomes `T`, while `str` and `Trait` turn into `i8` - this
+ /// is useful for indexing slices, as `&[T]`'s data pointer is `T*`.
+ /// If the type is an unsized struct, the regular layout is generated,
+ /// with the inner-most trailing unsized field using the "minimal unit"
+ /// of that field's type - this is useful for taking the address of
+ /// that field and ensuring the struct has the right alignment.
+ //TODO(antoyo): do we still need the set_fields parameter?
+ fn gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, set_fields: bool) -> Type<'gcc> {
+ if let Abi::Scalar(ref scalar) = self.abi {
+ // Use a different cache for scalars because pointers to DSTs
+ // can be either fat or thin (data pointers of fat pointers).
+ if let Some(&ty) = cx.scalar_types.borrow().get(&self.ty) {
+ return ty;
+ }
+ let ty =
+ match *self.ty.kind() {
+ ty::Ref(_, ty, _) | ty::RawPtr(ty::TypeAndMut { ty, .. }) => {
+ cx.type_ptr_to(cx.layout_of(ty).gcc_type(cx, set_fields))
+ }
+ ty::Adt(def, _) if def.is_box() => {
+ cx.type_ptr_to(cx.layout_of(self.ty.boxed_ty()).gcc_type(cx, true))
+ }
+ ty::FnPtr(sig) => cx.fn_ptr_backend_type(&FnAbi::of_fn_ptr(cx, sig, &[])),
+ _ => self.scalar_gcc_type_at(cx, scalar, Size::ZERO),
+ };
+ cx.scalar_types.borrow_mut().insert(self.ty, ty);
+ return ty;
+ }
+
+ // Check the cache.
+ let variant_index =
+ match self.variants {
+ Variants::Single { index } => Some(index),
+ _ => None,
+ };
+ let cached_type = cx.types.borrow().get(&(self.ty, variant_index)).cloned();
+ if let Some(ty) = cached_type {
+ let type_to_set_fields = cx.types_with_fields_to_set.borrow_mut().remove(&ty);
+ if let Some((struct_type, layout)) = type_to_set_fields {
+ // Since we might be trying to generate a type containing another type which is not
+ // completely generated yet, we deferred setting the fields until now.
+ let (fields, packed) = struct_fields(cx, layout);
+ cx.set_struct_body(struct_type, &fields, packed);
+ }
+ return ty;
+ }
+
+ assert!(!self.ty.has_escaping_bound_vars(), "{:?} has escaping bound vars", self.ty);
+
+ // Make sure lifetimes are erased, to avoid generating distinct LLVM
+ // types for Rust types that only differ in the choice of lifetimes.
+ let normal_ty = cx.tcx.erase_regions(self.ty);
+
+ let mut defer = None;
+ let ty =
+ if self.ty != normal_ty {
+ let mut layout = cx.layout_of(normal_ty);
+ if let Some(v) = variant_index {
+ layout = layout.for_variant(cx, v);
+ }
+ layout.gcc_type(cx, true)
+ }
+ else {
+ uncached_gcc_type(cx, *self, &mut defer)
+ };
+
+ cx.types.borrow_mut().insert((self.ty, variant_index), ty);
+
+ if let Some((ty, layout)) = defer {
+ let (fields, packed) = struct_fields(cx, layout);
+ cx.set_struct_body(ty, &fields, packed);
+ }
+
+ ty
+ }
+
+ fn immediate_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc> {
+ if let Abi::Scalar(ref scalar) = self.abi {
+ if scalar.is_bool() {
+ return cx.type_i1();
+ }
+ }
+ self.gcc_type(cx, true)
+ }
+
+ fn scalar_gcc_type_at<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, scalar: &abi::Scalar, offset: Size) -> Type<'gcc> {
+ match scalar.value {
+ Int(i, true) => cx.type_from_integer(i),
+ Int(i, false) => cx.type_from_unsigned_integer(i),
+ F32 => cx.type_f32(),
+ F64 => cx.type_f64(),
+ Pointer => {
+ // If we know the alignment, pick something better than i8.
+ let pointee =
+ if let Some(pointee) = self.pointee_info_at(cx, offset) {
+ cx.type_pointee_for_align(pointee.align)
+ }
+ else {
+ cx.type_i8()
+ };
+ cx.type_ptr_to(pointee)
+ }
+ }
+ }
+
+ fn scalar_pair_element_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, index: usize, immediate: bool) -> Type<'gcc> {
+ // TODO(antoyo): remove llvm hack:
+ // HACK(eddyb) special-case fat pointers until LLVM removes
+ // pointee types, to avoid bitcasting every `OperandRef::deref`.
+ match self.ty.kind() {
+ ty::Ref(..) | ty::RawPtr(_) => {
+ return self.field(cx, index).gcc_type(cx, true);
+ }
+ ty::Adt(def, _) if def.is_box() => {
+ let ptr_ty = cx.tcx.mk_mut_ptr(self.ty.boxed_ty());
+ return cx.layout_of(ptr_ty).scalar_pair_element_gcc_type(cx, index, immediate);
+ }
+ _ => {}
+ }
+
+ let (a, b) = match self.abi {
+ Abi::ScalarPair(ref a, ref b) => (a, b),
+ _ => bug!("TyAndLayout::scalar_pair_element_llty({:?}): not applicable", self),
+ };
+ let scalar = [a, b][index];
+
+ // Make sure to return the same type `immediate_gcc_type` would when
+ // dealing with an immediate pair. This means that `(bool, bool)` is
+ // effectively represented as `{i8, i8}` in memory and two `i1`s as an
+ // immediate, just like `bool` is typically `i8` in memory and only `i1`
+ // when immediate. We need to load/store `bool` as `i8` to avoid
+ // crippling LLVM optimizations or triggering other LLVM bugs with `i1`.
+ // TODO(antoyo): this bugs certainly don't happen in this case since the bool type is used instead of i1.
+ if scalar.is_bool() {
+ return cx.type_i1();
+ }
+
+ let offset =
+ if index == 0 {
+ Size::ZERO
+ }
+ else {
+ a.value.size(cx).align_to(b.value.align(cx).abi)
+ };
+ self.scalar_gcc_type_at(cx, scalar, offset)
+ }
+
+ fn gcc_field_index(&self, index: usize) -> u64 {
+ match self.abi {
+ Abi::Scalar(_) | Abi::ScalarPair(..) => {
+ bug!("TyAndLayout::gcc_field_index({:?}): not applicable", self)
+ }
+ _ => {}
+ }
+ match self.fields {
+ FieldsShape::Primitive | FieldsShape::Union(_) => {
+ bug!("TyAndLayout::gcc_field_index({:?}): not applicable", self)
+ }
+
+ FieldsShape::Array { .. } => index as u64,
+
+ FieldsShape::Arbitrary { .. } => 1 + (self.fields.memory_index(index) as u64) * 2,
+ }
+ }
+
+ fn pointee_info_at<'a>(&self, cx: &CodegenCx<'a, 'tcx>, offset: Size) -> Option<PointeeInfo> {
+ if let Some(&pointee) = cx.pointee_infos.borrow().get(&(self.ty, offset)) {
+ return pointee;
+ }
+
++ let result = Ty::ty_and_layout_pointee_info_at(*self, cx, offset);
+
+ cx.pointee_infos.borrow_mut().insert((self.ty, offset), result);
+ result
+ }
+}
+
+impl<'gcc, 'tcx> LayoutTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
+ fn backend_type(&self, layout: TyAndLayout<'tcx>) -> Type<'gcc> {
+ layout.gcc_type(self, true)
+ }
+
+ fn immediate_backend_type(&self, layout: TyAndLayout<'tcx>) -> Type<'gcc> {
+ layout.immediate_gcc_type(self)
+ }
+
+ fn is_backend_immediate(&self, layout: TyAndLayout<'tcx>) -> bool {
+ layout.is_gcc_immediate()
+ }
+
+ fn is_backend_scalar_pair(&self, layout: TyAndLayout<'tcx>) -> bool {
+ layout.is_gcc_scalar_pair()
+ }
+
+ fn backend_field_index(&self, layout: TyAndLayout<'tcx>, index: usize) -> u64 {
+ layout.gcc_field_index(index)
+ }
+
+ fn scalar_pair_element_backend_type(&self, layout: TyAndLayout<'tcx>, index: usize, immediate: bool) -> Type<'gcc> {
+ layout.scalar_pair_element_gcc_type(self, index, immediate)
+ }
+
+ fn cast_backend_type(&self, ty: &CastTarget) -> Type<'gcc> {
+ ty.gcc_type(self)
+ }
+
+ fn fn_ptr_backend_type(&self, fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> Type<'gcc> {
+ fn_abi.ptr_to_gcc_type(self)
+ }
+
+ fn reg_backend_type(&self, _ty: &Reg) -> Type<'gcc> {
+ unimplemented!();
+ }
+
+ fn fn_decl_backend_type(&self, _fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> Type<'gcc> {
+ // FIXME(antoyo): return correct type.
+ self.type_void()
+ }
+}
--- /dev/null
- export GCC_PATH=$(cat gcc_path)
+#!/bin/bash
+
+# TODO(antoyo): rewrite to cargo-make (or just) or something like that to only rebuild the sysroot when needed?
+
+set -e
+
- rm -r src/test/ui/{abi*,extern/,panic-runtime/,panics/,unsized-locals/,proc-macro/,threads-sendsync/,thinlto/,simd*,borrowck/,test*,*lto*.rs} || true
++if [ -f ./gcc_path ]; then
++ export GCC_PATH=$(cat gcc_path)
++else
++ echo 'Please put the path to your custom build of libgccjit in the file `gcc_path`, see Readme.md for details'
++ exit 1
++fi
+
+export LD_LIBRARY_PATH="$GCC_PATH"
+export LIBRARY_PATH="$GCC_PATH"
+
+if [[ "$1" == "--release" ]]; then
+ export CHANNEL='release'
+ CARGO_INCREMENTAL=1 cargo rustc --release
+else
+ echo $LD_LIBRARY_PATH
+ export CHANNEL='debug'
+ cargo rustc
+fi
+
+source config.sh
+
+rm -r target/out || true
+mkdir -p target/out/gccjit
+
+echo "[BUILD] mini_core"
+$RUSTC example/mini_core.rs --crate-name mini_core --crate-type lib,dylib --target $TARGET_TRIPLE
+
+echo "[BUILD] example"
+$RUSTC example/example.rs --crate-type lib --target $TARGET_TRIPLE
+
+echo "[AOT] mini_core_hello_world"
+$RUSTC example/mini_core_hello_world.rs --crate-name mini_core_hello_world --crate-type bin -g --target $TARGET_TRIPLE
+$RUN_WRAPPER ./target/out/mini_core_hello_world abc bcd
+
+echo "[BUILD] sysroot"
+time ./build_sysroot/build_sysroot.sh
+
+echo "[AOT] arbitrary_self_types_pointers_and_wrappers"
+$RUSTC example/arbitrary_self_types_pointers_and_wrappers.rs --crate-name arbitrary_self_types_pointers_and_wrappers --crate-type bin --target $TARGET_TRIPLE
+$RUN_WRAPPER ./target/out/arbitrary_self_types_pointers_and_wrappers
+
+echo "[AOT] alloc_system"
+$RUSTC example/alloc_system.rs --crate-type lib --target "$TARGET_TRIPLE"
+
+echo "[AOT] alloc_example"
+$RUSTC example/alloc_example.rs --crate-type bin --target $TARGET_TRIPLE
+$RUN_WRAPPER ./target/out/alloc_example
+
+echo "[AOT] dst_field_align"
+# FIXME(antoyo): Re-add -Zmir-opt-level=2 once rust-lang/rust#67529 is fixed.
+$RUSTC example/dst-field-align.rs --crate-name dst_field_align --crate-type bin --target $TARGET_TRIPLE
+$RUN_WRAPPER ./target/out/dst_field_align || (echo $?; false)
+
+echo "[AOT] std_example"
+$RUSTC example/std_example.rs --crate-type bin --target $TARGET_TRIPLE
+$RUN_WRAPPER ./target/out/std_example --target $TARGET_TRIPLE
+
+echo "[AOT] subslice-patterns-const-eval"
+$RUSTC example/subslice-patterns-const-eval.rs --crate-type bin -Cpanic=abort --target $TARGET_TRIPLE
+$RUN_WRAPPER ./target/out/subslice-patterns-const-eval
+
+echo "[AOT] track-caller-attribute"
+$RUSTC example/track-caller-attribute.rs --crate-type bin -Cpanic=abort --target $TARGET_TRIPLE
+$RUN_WRAPPER ./target/out/track-caller-attribute
+
+echo "[BUILD] mod_bench"
+$RUSTC example/mod_bench.rs --crate-type bin --target $TARGET_TRIPLE
+
+# FIXME(antoyo): linker gives multiple definitions error on Linux
+#echo "[BUILD] sysroot in release mode"
+#./build_sysroot/build_sysroot.sh --release
+
+# TODO(antoyo): uncomment when it works.
+#pushd simple-raytracer
+#if [[ "$HOST_TRIPLE" = "$TARGET_TRIPLE" ]]; then
+ #echo "[BENCH COMPILE] ebobby/simple-raytracer"
+ #hyperfine --runs ${RUN_RUNS:-10} --warmup 1 --prepare "rm -r target/*/debug || true" \
+ #"RUSTFLAGS='' cargo build --target $TARGET_TRIPLE" \
+ #"../cargo.sh build"
+
+ #echo "[BENCH RUN] ebobby/simple-raytracer"
+ #cp ./target/*/debug/main ./raytracer_cg_gccjit
+ #hyperfine --runs ${RUN_RUNS:-10} ./raytracer_cg_llvm ./raytracer_cg_gccjit
+#else
+ #echo "[BENCH COMPILE] ebobby/simple-raytracer (skipped)"
+ #echo "[COMPILE] ebobby/simple-raytracer"
+ #../cargo.sh build
+ #echo "[BENCH RUN] ebobby/simple-raytracer (skipped)"
+#fi
+#popd
+
+pushd build_sysroot/sysroot_src/library/core/tests
+echo "[TEST] libcore"
+rm -r ./target || true
+../../../../../cargo.sh test
+popd
+
+# TODO(antoyo): uncomment when it works.
+#pushd regex
+#echo "[TEST] rust-lang/regex example shootout-regex-dna"
+#../cargo.sh clean
+## Make sure `[codegen mono items] start` doesn't poison the diff
+#../cargo.sh build --example shootout-regex-dna
+#cat examples/regexdna-input.txt | ../cargo.sh run --example shootout-regex-dna | grep -v "Spawned thread" > res.txt
+#diff -u res.txt examples/regexdna-output.txt
+
+#echo "[TEST] rust-lang/regex tests"
+#../cargo.sh test --tests -- --exclude-should-panic --test-threads 1 -Zunstable-options
+#popd
+
+#echo
+#echo "[BENCH COMPILE] mod_bench"
+
+#COMPILE_MOD_BENCH_INLINE="$RUSTC example/mod_bench.rs --crate-type bin -Zmir-opt-level=3 -O --crate-name mod_bench_inline"
+#COMPILE_MOD_BENCH_LLVM_0="rustc example/mod_bench.rs --crate-type bin -Copt-level=0 -o target/out/mod_bench_llvm_0 -Cpanic=abort"
+#COMPILE_MOD_BENCH_LLVM_1="rustc example/mod_bench.rs --crate-type bin -Copt-level=1 -o target/out/mod_bench_llvm_1 -Cpanic=abort"
+#COMPILE_MOD_BENCH_LLVM_2="rustc example/mod_bench.rs --crate-type bin -Copt-level=2 -o target/out/mod_bench_llvm_2 -Cpanic=abort"
+#COMPILE_MOD_BENCH_LLVM_3="rustc example/mod_bench.rs --crate-type bin -Copt-level=3 -o target/out/mod_bench_llvm_3 -Cpanic=abort"
+
+## Use 100 runs, because a single compilations doesn't take more than ~150ms, so it isn't very slow
+#hyperfine --runs ${COMPILE_RUNS:-100} "$COMPILE_MOD_BENCH_INLINE" "$COMPILE_MOD_BENCH_LLVM_0" "$COMPILE_MOD_BENCH_LLVM_1" "$COMPILE_MOD_BENCH_LLVM_2" "$COMPILE_MOD_BENCH_LLVM_3"
+
+#echo
+#echo "[BENCH RUN] mod_bench"
+#hyperfine --runs ${RUN_RUNS:-10} ./target/out/mod_bench{,_inline} ./target/out/mod_bench_llvm_*
+
+echo
+echo "[TEST] rust-lang/rust"
+
+rust_toolchain=$(cat rust-toolchain)
+
+git clone https://github.com/rust-lang/rust.git || true
+cd rust
+git fetch
+git checkout $(rustc -V | cut -d' ' -f3 | tr -d '(')
+export RUSTFLAGS=
+
+rm config.toml || true
+
+cat > config.toml <<EOF
+[rust]
+codegen-backends = []
++deny-warnings = false
+
+[build]
+cargo = "$(which cargo)"
+local-rebuild = true
+rustc = "$HOME/.rustup/toolchains/$rust_toolchain-$TARGET_TRIPLE/bin/rustc"
+EOF
+
+rustc -V | cut -d' ' -f3 | tr -d '('
+git checkout $(rustc -V | cut -d' ' -f3 | tr -d '(') src/test
+
+for test in $(rg -i --files-with-matches "//(\[\w+\])?~|// error-pattern:|// build-fail|// run-fail|-Cllvm-args" src/test/ui); do
+ rm $test
+done
+
+git checkout -- src/test/ui/issues/auxiliary/issue-3136-a.rs # contains //~ERROR, but shouldn't be removed
+
++rm -r src/test/ui/{abi*,extern/,llvm-asm/,panic-runtime/,panics/,unsized-locals/,proc-macro/,threads-sendsync/,thinlto/,simd*,borrowck/,test*,*lto*.rs} || true
+for test in $(rg --files-with-matches "catch_unwind|should_panic|thread|lto" src/test/ui); do
+ rm $test
+done
+git checkout src/test/ui/type-alias-impl-trait/auxiliary/cross_crate_ice.rs
+git checkout src/test/ui/type-alias-impl-trait/auxiliary/cross_crate_ice2.rs
+rm src/test/ui/llvm-asm/llvm-asm-in-out-operand.rs || true # TODO(antoyo): Enable back this test if I ever implement the llvm_asm! macro.
+
+RUSTC_ARGS="-Zpanic-abort-tests -Zcodegen-backend="$(pwd)"/../target/"$CHANNEL"/librustc_codegen_gcc."$dylib_ext" --sysroot "$(pwd)"/../build_sysroot/sysroot -Cpanic=abort"
+
+echo "[TEST] rustc test suite"
+COMPILETEST_FORCE_STAGE0=1 ./x.py test --run always --stage 0 src/test/ui/ --rustc-args "$RUSTC_ARGS"
--- /dev/null
+// Compiler:
+//
+// Run-time:
+// status: 0
+
+#![feature(asm, global_asm)]
+
+global_asm!("
+ .global add_asm
+add_asm:
+ mov rax, rdi
+ add rax, rsi
+ ret"
+);
+
+extern "C" {
+ fn add_asm(a: i64, b: i64) -> i64;
+}
+
+fn main() {
+ unsafe {
+ asm!("nop");
+ }
+
+ let x: u64;
+ unsafe {
+ asm!("mov $5, {}",
+ out(reg) x,
+ options(att_syntax)
+ );
+ }
+ assert_eq!(x, 5);
+
+ let x: u64;
+ let input: u64 = 42;
+ unsafe {
+ asm!("mov {input}, {output}",
+ "add $1, {output}",
+ input = in(reg) input,
+ output = out(reg) x,
+ options(att_syntax)
+ );
+ }
+ assert_eq!(x, 43);
+
+ let x: u64;
+ unsafe {
+ asm!("mov {}, 6",
+ out(reg) x,
+ );
+ }
+ assert_eq!(x, 6);
+
+ let x: u64;
+ let input: u64 = 42;
+ unsafe {
+ asm!("mov {output}, {input}",
+ "add {output}, 1",
+ input = in(reg) input,
+ output = out(reg) x,
+ );
+ }
+ assert_eq!(x, 43);
+
++ // check inout(reg_class) x
++ let mut x: u64 = 42;
++ unsafe {
++ asm!("add {0}, {0}",
++ inout(reg) x
++ );
++ }
++ assert_eq!(x, 84);
++
++ // check inout("reg") x
++ let mut x: u64 = 42;
++ unsafe {
++ asm!("add r11, r11",
++ inout("r11") x
++ );
++ }
++ assert_eq!(x, 84);
++
++ // check a mix of
++ // in("reg")
++ // inout(class) x => y
++ // inout (class) x
++ let x: u64 = 702;
++ let y: u64 = 100;
++ let res: u64;
++ let mut rem: u64 = 0;
++ unsafe {
++ asm!("div r11",
++ in("r11") y,
++ inout("eax") x => res,
++ inout("edx") rem,
++ );
++ }
++ assert_eq!(res, 7);
++ assert_eq!(rem, 2);
++
++ // check const
++ let mut x: u64 = 42;
++ unsafe {
++ asm!("add {}, {}",
++ inout(reg) x,
++ const 1
++ );
++ }
++ assert_eq!(x, 43);
++
++ // check const (ATT syntax)
++ let mut x: u64 = 42;
++ unsafe {
++ asm!("add {}, {}",
++ const 1,
++ inout(reg) x,
++ options(att_syntax)
++ );
++ }
++ assert_eq!(x, 43);
++
++ // check sym fn
++ extern "C" fn foo() -> u64 { 42 }
++ let x: u64;
++ unsafe {
++ asm!("call {}", sym foo, lateout("rax") x);
++ }
++ assert_eq!(x, 42);
++
++ // check sym fn (ATT syntax)
++ let x: u64;
++ unsafe {
++ asm!("call {}", sym foo, lateout("rax") x, options(att_syntax));
++ }
++ assert_eq!(x, 42);
++
++ // check sym static
++ static FOO: u64 = 42;
++ let x: u64;
++ unsafe {
++ asm!("mov {1}, qword ptr [rip + {0}]", sym FOO, lateout(reg) x);
++ }
++ assert_eq!(x, 42);
++
++ // check sym static (ATT syntax)
++ let x: u64;
++ unsafe {
++ asm!("movq {0}(%rip), {1}", sym FOO, lateout(reg) x, options(att_syntax));
++ }
++ assert_eq!(x, 42);
++
+ assert_eq!(unsafe { add_asm(40, 2) }, 42);
+}