1 use crate::abi::{Abi, FnAbi, FnAbiLlvmExt, LlvmType, PassMode};
2 use crate::builder::Builder;
3 use crate::context::CodegenCx;
5 use crate::type_::Type;
6 use crate::type_of::LayoutLlvmExt;
7 use crate::va_arg::emit_va_arg;
8 use crate::value::Value;
10 use rustc_codegen_ssa::base::{compare_simd_types, wants_msvc_seh};
11 use rustc_codegen_ssa::common::span_invalid_monomorphization_error;
12 use rustc_codegen_ssa::common::{IntPredicate, TypeKind};
13 use rustc_codegen_ssa::mir::operand::OperandRef;
14 use rustc_codegen_ssa::mir::place::PlaceRef;
15 use rustc_codegen_ssa::traits::*;
17 use rustc_middle::ty::layout::{FnAbiOf, HasTyCtxt, LayoutOf};
18 use rustc_middle::ty::{self, Ty};
19 use rustc_middle::{bug, span_bug};
20 use rustc_span::{sym, symbol::kw, Span, Symbol};
21 use rustc_target::abi::{self, Align, HasDataLayout, Primitive};
22 use rustc_target::spec::{HasTargetSpec, PanicStrategy};
24 use std::cmp::Ordering;
27 fn get_simple_intrinsic<'ll>(
28 cx: &CodegenCx<'ll, '_>,
30 ) -> Option<(&'ll Type, &'ll Value)> {
31 let llvm_name = match name {
32 sym::sqrtf32 => "llvm.sqrt.f32",
33 sym::sqrtf64 => "llvm.sqrt.f64",
34 sym::powif32 => "llvm.powi.f32",
35 sym::powif64 => "llvm.powi.f64",
36 sym::sinf32 => "llvm.sin.f32",
37 sym::sinf64 => "llvm.sin.f64",
38 sym::cosf32 => "llvm.cos.f32",
39 sym::cosf64 => "llvm.cos.f64",
40 sym::powf32 => "llvm.pow.f32",
41 sym::powf64 => "llvm.pow.f64",
42 sym::expf32 => "llvm.exp.f32",
43 sym::expf64 => "llvm.exp.f64",
44 sym::exp2f32 => "llvm.exp2.f32",
45 sym::exp2f64 => "llvm.exp2.f64",
46 sym::logf32 => "llvm.log.f32",
47 sym::logf64 => "llvm.log.f64",
48 sym::log10f32 => "llvm.log10.f32",
49 sym::log10f64 => "llvm.log10.f64",
50 sym::log2f32 => "llvm.log2.f32",
51 sym::log2f64 => "llvm.log2.f64",
52 sym::fmaf32 => "llvm.fma.f32",
53 sym::fmaf64 => "llvm.fma.f64",
54 sym::fabsf32 => "llvm.fabs.f32",
55 sym::fabsf64 => "llvm.fabs.f64",
56 sym::minnumf32 => "llvm.minnum.f32",
57 sym::minnumf64 => "llvm.minnum.f64",
58 sym::maxnumf32 => "llvm.maxnum.f32",
59 sym::maxnumf64 => "llvm.maxnum.f64",
60 sym::copysignf32 => "llvm.copysign.f32",
61 sym::copysignf64 => "llvm.copysign.f64",
62 sym::floorf32 => "llvm.floor.f32",
63 sym::floorf64 => "llvm.floor.f64",
64 sym::ceilf32 => "llvm.ceil.f32",
65 sym::ceilf64 => "llvm.ceil.f64",
66 sym::truncf32 => "llvm.trunc.f32",
67 sym::truncf64 => "llvm.trunc.f64",
68 sym::rintf32 => "llvm.rint.f32",
69 sym::rintf64 => "llvm.rint.f64",
70 sym::nearbyintf32 => "llvm.nearbyint.f32",
71 sym::nearbyintf64 => "llvm.nearbyint.f64",
72 sym::roundf32 => "llvm.round.f32",
73 sym::roundf64 => "llvm.round.f64",
76 Some(cx.get_intrinsic(llvm_name))
79 impl<'ll, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'_, 'll, 'tcx> {
80 fn codegen_intrinsic_call(
82 instance: ty::Instance<'tcx>,
83 fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
84 args: &[OperandRef<'tcx, &'ll Value>],
89 let callee_ty = instance.ty(tcx, ty::ParamEnv::reveal_all());
91 let ty::FnDef(def_id, substs) = *callee_ty.kind() else {
92 bug!("expected fn item type, found {}", callee_ty);
95 let sig = callee_ty.fn_sig(tcx);
96 let sig = tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), sig);
97 let arg_tys = sig.inputs();
98 let ret_ty = sig.output();
99 let name = tcx.item_name(def_id);
101 let llret_ty = self.layout_of(ret_ty).llvm_type(self);
102 let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
104 let simple = get_simple_intrinsic(self, name);
105 let llval = match name {
106 _ if simple.is_some() => {
107 let (simple_ty, simple_fn) = simple.unwrap();
111 &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
116 self.call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(true)])
118 sym::unlikely => self
119 .call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(false)]),
130 sym::breakpoint => self.call_intrinsic("llvm.debugtrap", &[]),
132 self.call_intrinsic("llvm.va_copy", &[args[0].immediate(), args[1].immediate()])
135 match fn_abi.ret.layout.abi {
136 abi::Abi::Scalar(scalar) => {
138 Primitive::Int(..) => {
139 if self.cx().size_of(ret_ty).bytes() < 4 {
140 // `va_arg` should not be called on an integer type
141 // less than 4 bytes in length. If it is, promote
142 // the integer to an `i32` and truncate the result
143 // back to the smaller type.
144 let promoted_result = emit_va_arg(self, args[0], tcx.types.i32);
145 self.trunc(promoted_result, llret_ty)
147 emit_va_arg(self, args[0], ret_ty)
150 Primitive::F64 | Primitive::Pointer => {
151 emit_va_arg(self, args[0], ret_ty)
153 // `va_arg` should never be used with the return type f32.
154 Primitive::F32 => bug!("the va_arg intrinsic does not work with `f32`"),
157 _ => bug!("the va_arg intrinsic does not work with non-scalar types"),
161 sym::volatile_load | sym::unaligned_volatile_load => {
162 let tp_ty = substs.type_at(0);
163 let ptr = args[0].immediate();
164 let load = if let PassMode::Cast(ty) = fn_abi.ret.mode {
165 let llty = ty.llvm_type(self);
166 let ptr = self.pointercast(ptr, self.type_ptr_to(llty));
167 self.volatile_load(llty, ptr)
169 self.volatile_load(self.layout_of(tp_ty).llvm_type(self), ptr)
171 let align = if name == sym::unaligned_volatile_load {
174 self.align_of(tp_ty).bytes() as u32
177 llvm::LLVMSetAlignment(load, align);
179 self.to_immediate(load, self.layout_of(tp_ty))
181 sym::volatile_store => {
182 let dst = args[0].deref(self.cx());
183 args[1].val.volatile_store(self, dst);
186 sym::unaligned_volatile_store => {
187 let dst = args[0].deref(self.cx());
188 args[1].val.unaligned_volatile_store(self, dst);
191 sym::prefetch_read_data
192 | sym::prefetch_write_data
193 | sym::prefetch_read_instruction
194 | sym::prefetch_write_instruction => {
195 let (rw, cache_type) = match name {
196 sym::prefetch_read_data => (0, 1),
197 sym::prefetch_write_data => (1, 1),
198 sym::prefetch_read_instruction => (0, 0),
199 sym::prefetch_write_instruction => (1, 0),
208 self.const_i32(cache_type),
221 | sym::saturating_add
222 | sym::saturating_sub => {
224 match int_type_width_signed(ty, self) {
225 Some((width, signed)) => match name {
226 sym::ctlz | sym::cttz => {
227 let y = self.const_bool(false);
229 &format!("llvm.{}.i{}", name, width),
230 &[args[0].immediate(), y],
233 sym::ctlz_nonzero => {
234 let y = self.const_bool(true);
235 let llvm_name = &format!("llvm.ctlz.i{}", width);
236 self.call_intrinsic(llvm_name, &[args[0].immediate(), y])
238 sym::cttz_nonzero => {
239 let y = self.const_bool(true);
240 let llvm_name = &format!("llvm.cttz.i{}", width);
241 self.call_intrinsic(llvm_name, &[args[0].immediate(), y])
243 sym::ctpop => self.call_intrinsic(
244 &format!("llvm.ctpop.i{}", width),
245 &[args[0].immediate()],
249 args[0].immediate() // byte swap a u8/i8 is just a no-op
252 &format!("llvm.bswap.i{}", width),
253 &[args[0].immediate()],
257 sym::bitreverse => self.call_intrinsic(
258 &format!("llvm.bitreverse.i{}", width),
259 &[args[0].immediate()],
261 sym::rotate_left | sym::rotate_right => {
262 let is_left = name == sym::rotate_left;
263 let val = args[0].immediate();
264 let raw_shift = args[1].immediate();
265 // rotate = funnel shift with first two args the same
267 &format!("llvm.fsh{}.i{}", if is_left { 'l' } else { 'r' }, width);
268 self.call_intrinsic(llvm_name, &[val, val, raw_shift])
270 sym::saturating_add | sym::saturating_sub => {
271 let is_add = name == sym::saturating_add;
272 let lhs = args[0].immediate();
273 let rhs = args[1].immediate();
274 let llvm_name = &format!(
276 if signed { 's' } else { 'u' },
277 if is_add { "add" } else { "sub" },
280 self.call_intrinsic(llvm_name, &[lhs, rhs])
285 span_invalid_monomorphization_error(
289 "invalid monomorphization of `{}` intrinsic: \
290 expected basic integer type, found `{}`",
301 let tp_ty = substs.type_at(0);
302 let layout = self.layout_of(tp_ty).layout;
303 let use_integer_compare = match layout.abi {
304 Scalar(_) | ScalarPair(_, _) => true,
305 Uninhabited | Vector { .. } => false,
306 Aggregate { .. } => {
307 // For rusty ABIs, small aggregates are actually passed
308 // as `RegKind::Integer` (see `FnAbi::adjust_for_abi`),
309 // so we re-use that same threshold here.
310 layout.size <= self.data_layout().pointer_size * 2
314 let a = args[0].immediate();
315 let b = args[1].immediate();
316 if layout.size.bytes() == 0 {
317 self.const_bool(true)
318 } else if use_integer_compare {
319 let integer_ty = self.type_ix(layout.size.bits());
320 let ptr_ty = self.type_ptr_to(integer_ty);
321 let a_ptr = self.bitcast(a, ptr_ty);
322 let a_val = self.load(integer_ty, a_ptr, layout.align.abi);
323 let b_ptr = self.bitcast(b, ptr_ty);
324 let b_val = self.load(integer_ty, b_ptr, layout.align.abi);
325 self.icmp(IntPredicate::IntEQ, a_val, b_val)
327 let i8p_ty = self.type_i8p();
328 let a_ptr = self.bitcast(a, i8p_ty);
329 let b_ptr = self.bitcast(b, i8p_ty);
330 let n = self.const_usize(layout.size.bytes());
331 let cmp = self.call_intrinsic("memcmp", &[a_ptr, b_ptr, n]);
332 self.icmp(IntPredicate::IntEQ, cmp, self.const_i32(0))
337 args[0].val.store(self, result);
339 // We need to "use" the argument in some way LLVM can't introspect, and on
340 // targets that support it we can typically leverage inline assembly to do
341 // this. LLVM's interpretation of inline assembly is that it's, well, a black
342 // box. This isn't the greatest implementation since it probably deoptimizes
343 // more than we want, but it's so far good enough.
344 crate::asm::inline_asm_call(
352 llvm::AsmDialect::Att,
357 .unwrap_or_else(|| bug!("failed to generate inline asm call for `black_box`"));
359 // We have copied the value to `result` already.
363 _ if name.as_str().starts_with("simd_") => {
364 match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
370 _ => bug!("unknown intrinsic '{}'", name),
373 if !fn_abi.ret.is_ignore() {
374 if let PassMode::Cast(ty) = fn_abi.ret.mode {
375 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
376 let ptr = self.pointercast(result.llval, ptr_llty);
377 self.store(llval, ptr, result.align);
379 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
381 .store(self, result);
386 fn abort(&mut self) {
387 self.call_intrinsic("llvm.trap", &[]);
390 fn assume(&mut self, val: Self::Value) {
391 self.call_intrinsic("llvm.assume", &[val]);
394 fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
395 self.call_intrinsic("llvm.expect.i1", &[cond, self.const_bool(expected)])
398 fn type_test(&mut self, pointer: Self::Value, typeid: Self::Value) -> Self::Value {
399 // Test the called operand using llvm.type.test intrinsic. The LowerTypeTests link-time
400 // optimization pass replaces calls to this intrinsic with code to test type membership.
401 let i8p_ty = self.type_i8p();
402 let bitcast = self.bitcast(pointer, i8p_ty);
403 self.call_intrinsic("llvm.type.test", &[bitcast, typeid])
406 fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
407 self.call_intrinsic("llvm.va_start", &[va_list])
410 fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
411 self.call_intrinsic("llvm.va_end", &[va_list])
415 fn try_intrinsic<'ll>(
416 bx: &mut Builder<'_, 'll, '_>,
417 try_func: &'ll Value,
419 catch_func: &'ll Value,
422 if bx.sess().panic_strategy() == PanicStrategy::Abort {
423 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
424 bx.call(try_func_ty, try_func, &[data], None);
425 // Return 0 unconditionally from the intrinsic call;
426 // we can never unwind.
427 let ret_align = bx.tcx().data_layout.i32_align.abi;
428 bx.store(bx.const_i32(0), dest, ret_align);
429 } else if wants_msvc_seh(bx.sess()) {
430 codegen_msvc_try(bx, try_func, data, catch_func, dest);
431 } else if bx.sess().target.is_like_emscripten {
432 codegen_emcc_try(bx, try_func, data, catch_func, dest);
434 codegen_gnu_try(bx, try_func, data, catch_func, dest);
438 // MSVC's definition of the `rust_try` function.
440 // This implementation uses the new exception handling instructions in LLVM
441 // which have support in LLVM for SEH on MSVC targets. Although these
442 // instructions are meant to work for all targets, as of the time of this
443 // writing, however, LLVM does not recommend the usage of these new instructions
444 // as the old ones are still more optimized.
445 fn codegen_msvc_try<'ll>(
446 bx: &mut Builder<'_, 'll, '_>,
447 try_func: &'ll Value,
449 catch_func: &'ll Value,
452 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
453 bx.set_personality_fn(bx.eh_personality());
455 let mut normal = bx.build_sibling_block("normal");
456 let mut catchswitch = bx.build_sibling_block("catchswitch");
457 let mut catchpad_rust = bx.build_sibling_block("catchpad_rust");
458 let mut catchpad_foreign = bx.build_sibling_block("catchpad_foreign");
459 let mut caught = bx.build_sibling_block("caught");
461 let try_func = llvm::get_param(bx.llfn(), 0);
462 let data = llvm::get_param(bx.llfn(), 1);
463 let catch_func = llvm::get_param(bx.llfn(), 2);
465 // We're generating an IR snippet that looks like:
467 // declare i32 @rust_try(%try_func, %data, %catch_func) {
468 // %slot = alloca i8*
469 // invoke %try_func(%data) to label %normal unwind label %catchswitch
475 // %cs = catchswitch within none [%catchpad_rust, %catchpad_foreign] unwind to caller
478 // %tok = catchpad within %cs [%type_descriptor, 8, %slot]
480 // call %catch_func(%data, %ptr)
481 // catchret from %tok to label %caught
484 // %tok = catchpad within %cs [null, 64, null]
485 // call %catch_func(%data, null)
486 // catchret from %tok to label %caught
492 // This structure follows the basic usage of throw/try/catch in LLVM.
493 // For example, compile this C++ snippet to see what LLVM generates:
495 // struct rust_panic {
496 // rust_panic(const rust_panic&);
503 // void (*try_func)(void*),
505 // void (*catch_func)(void*, void*) noexcept
510 // } catch(rust_panic& a) {
511 // catch_func(data, &a);
514 // catch_func(data, NULL);
519 // More information can be found in libstd's seh.rs implementation.
520 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
521 let slot = bx.alloca(bx.type_i8p(), ptr_align);
522 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
523 bx.invoke(try_func_ty, try_func, &[data], normal.llbb(), catchswitch.llbb(), None);
525 normal.ret(bx.const_i32(0));
528 catchswitch.catch_switch(None, None, &[catchpad_rust.llbb(), catchpad_foreign.llbb()]);
530 // We can't use the TypeDescriptor defined in libpanic_unwind because it
531 // might be in another DLL and the SEH encoding only supports specifying
532 // a TypeDescriptor from the current module.
534 // However this isn't an issue since the MSVC runtime uses string
535 // comparison on the type name to match TypeDescriptors rather than
538 // So instead we generate a new TypeDescriptor in each module that uses
539 // `try` and let the linker merge duplicate definitions in the same
542 // When modifying, make sure that the type_name string exactly matches
543 // the one used in src/libpanic_unwind/seh.rs.
544 let type_info_vtable = bx.declare_global("??_7type_info@@6B@", bx.type_i8p());
545 let type_name = bx.const_bytes(b"rust_panic\0");
547 bx.const_struct(&[type_info_vtable, bx.const_null(bx.type_i8p()), type_name], false);
548 let tydesc = bx.declare_global("__rust_panic_type_info", bx.val_ty(type_info));
550 llvm::LLVMRustSetLinkage(tydesc, llvm::Linkage::LinkOnceODRLinkage);
551 llvm::SetUniqueComdat(bx.llmod, tydesc);
552 llvm::LLVMSetInitializer(tydesc, type_info);
555 // The flag value of 8 indicates that we are catching the exception by
556 // reference instead of by value. We can't use catch by value because
557 // that requires copying the exception object, which we don't support
558 // since our exception object effectively contains a Box.
560 // Source: MicrosoftCXXABI::getAddrOfCXXCatchHandlerType in clang
561 let flags = bx.const_i32(8);
562 let funclet = catchpad_rust.catch_pad(cs, &[tydesc, flags, slot]);
563 let ptr = catchpad_rust.load(bx.type_i8p(), slot, ptr_align);
564 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
565 catchpad_rust.call(catch_ty, catch_func, &[data, ptr], Some(&funclet));
566 catchpad_rust.catch_ret(&funclet, caught.llbb());
568 // The flag value of 64 indicates a "catch-all".
569 let flags = bx.const_i32(64);
570 let null = bx.const_null(bx.type_i8p());
571 let funclet = catchpad_foreign.catch_pad(cs, &[null, flags, null]);
572 catchpad_foreign.call(catch_ty, catch_func, &[data, null], Some(&funclet));
573 catchpad_foreign.catch_ret(&funclet, caught.llbb());
575 caught.ret(bx.const_i32(1));
578 // Note that no invoke is used here because by definition this function
579 // can't panic (that's what it's catching).
580 let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
581 let i32_align = bx.tcx().data_layout.i32_align.abi;
582 bx.store(ret, dest, i32_align);
585 // Definition of the standard `try` function for Rust using the GNU-like model
586 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
589 // This codegen is a little surprising because we always call a shim
590 // function instead of inlining the call to `invoke` manually here. This is done
591 // because in LLVM we're only allowed to have one personality per function
592 // definition. The call to the `try` intrinsic is being inlined into the
593 // function calling it, and that function may already have other personality
594 // functions in play. By calling a shim we're guaranteed that our shim will have
595 // the right personality function.
596 fn codegen_gnu_try<'ll>(
597 bx: &mut Builder<'_, 'll, '_>,
598 try_func: &'ll Value,
600 catch_func: &'ll Value,
603 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
604 // Codegens the shims described above:
607 // invoke %try_func(%data) normal %normal unwind %catch
613 // (%ptr, _) = landingpad
614 // call %catch_func(%data, %ptr)
616 let mut then = bx.build_sibling_block("then");
617 let mut catch = bx.build_sibling_block("catch");
619 let try_func = llvm::get_param(bx.llfn(), 0);
620 let data = llvm::get_param(bx.llfn(), 1);
621 let catch_func = llvm::get_param(bx.llfn(), 2);
622 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
623 bx.invoke(try_func_ty, try_func, &[data], then.llbb(), catch.llbb(), None);
624 then.ret(bx.const_i32(0));
626 // Type indicator for the exception being thrown.
628 // The first value in this tuple is a pointer to the exception object
629 // being thrown. The second value is a "selector" indicating which of
630 // the landing pad clauses the exception's type had been matched to.
631 // rust_try ignores the selector.
632 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
633 let vals = catch.landing_pad(lpad_ty, bx.eh_personality(), 1);
634 let tydesc = bx.const_null(bx.type_i8p());
635 catch.add_clause(vals, tydesc);
636 let ptr = catch.extract_value(vals, 0);
637 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
638 catch.call(catch_ty, catch_func, &[data, ptr], None);
639 catch.ret(bx.const_i32(1));
642 // Note that no invoke is used here because by definition this function
643 // can't panic (that's what it's catching).
644 let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
645 let i32_align = bx.tcx().data_layout.i32_align.abi;
646 bx.store(ret, dest, i32_align);
649 // Variant of codegen_gnu_try used for emscripten where Rust panics are
650 // implemented using C++ exceptions. Here we use exceptions of a specific type
651 // (`struct rust_panic`) to represent Rust panics.
652 fn codegen_emcc_try<'ll>(
653 bx: &mut Builder<'_, 'll, '_>,
654 try_func: &'ll Value,
656 catch_func: &'ll Value,
659 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
660 // Codegens the shims described above:
663 // invoke %try_func(%data) normal %normal unwind %catch
669 // (%ptr, %selector) = landingpad
670 // %rust_typeid = @llvm.eh.typeid.for(@_ZTI10rust_panic)
671 // %is_rust_panic = %selector == %rust_typeid
672 // %catch_data = alloca { i8*, i8 }
673 // %catch_data[0] = %ptr
674 // %catch_data[1] = %is_rust_panic
675 // call %catch_func(%data, %catch_data)
677 let mut then = bx.build_sibling_block("then");
678 let mut catch = bx.build_sibling_block("catch");
680 let try_func = llvm::get_param(bx.llfn(), 0);
681 let data = llvm::get_param(bx.llfn(), 1);
682 let catch_func = llvm::get_param(bx.llfn(), 2);
683 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
684 bx.invoke(try_func_ty, try_func, &[data], then.llbb(), catch.llbb(), None);
685 then.ret(bx.const_i32(0));
687 // Type indicator for the exception being thrown.
689 // The first value in this tuple is a pointer to the exception object
690 // being thrown. The second value is a "selector" indicating which of
691 // the landing pad clauses the exception's type had been matched to.
692 let tydesc = bx.eh_catch_typeinfo();
693 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
694 let vals = catch.landing_pad(lpad_ty, bx.eh_personality(), 2);
695 catch.add_clause(vals, tydesc);
696 catch.add_clause(vals, bx.const_null(bx.type_i8p()));
697 let ptr = catch.extract_value(vals, 0);
698 let selector = catch.extract_value(vals, 1);
700 // Check if the typeid we got is the one for a Rust panic.
701 let rust_typeid = catch.call_intrinsic("llvm.eh.typeid.for", &[tydesc]);
702 let is_rust_panic = catch.icmp(IntPredicate::IntEQ, selector, rust_typeid);
703 let is_rust_panic = catch.zext(is_rust_panic, bx.type_bool());
705 // We need to pass two values to catch_func (ptr and is_rust_panic), so
706 // create an alloca and pass a pointer to that.
707 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
708 let i8_align = bx.tcx().data_layout.i8_align.abi;
709 let catch_data_type = bx.type_struct(&[bx.type_i8p(), bx.type_bool()], false);
710 let catch_data = catch.alloca(catch_data_type, ptr_align);
711 let catch_data_0 = catch.inbounds_gep(
714 &[bx.const_usize(0), bx.const_usize(0)],
716 catch.store(ptr, catch_data_0, ptr_align);
717 let catch_data_1 = catch.inbounds_gep(
720 &[bx.const_usize(0), bx.const_usize(1)],
722 catch.store(is_rust_panic, catch_data_1, i8_align);
723 let catch_data = catch.bitcast(catch_data, bx.type_i8p());
725 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
726 catch.call(catch_ty, catch_func, &[data, catch_data], None);
727 catch.ret(bx.const_i32(1));
730 // Note that no invoke is used here because by definition this function
731 // can't panic (that's what it's catching).
732 let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
733 let i32_align = bx.tcx().data_layout.i32_align.abi;
734 bx.store(ret, dest, i32_align);
737 // Helper function to give a Block to a closure to codegen a shim function.
738 // This is currently primarily used for the `try` intrinsic functions above.
739 fn gen_fn<'ll, 'tcx>(
740 cx: &CodegenCx<'ll, 'tcx>,
742 rust_fn_sig: ty::PolyFnSig<'tcx>,
743 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
744 ) -> (&'ll Type, &'ll Value) {
745 let fn_abi = cx.fn_abi_of_fn_ptr(rust_fn_sig, ty::List::empty());
746 let llty = fn_abi.llvm_type(cx);
747 let llfn = cx.declare_fn(name, fn_abi);
748 cx.set_frame_pointer_type(llfn);
749 cx.apply_target_cpu_attr(llfn);
750 // FIXME(eddyb) find a nicer way to do this.
751 unsafe { llvm::LLVMRustSetLinkage(llfn, llvm::Linkage::InternalLinkage) };
752 let llbb = Builder::append_block(cx, llfn, "entry-block");
753 let bx = Builder::build(cx, llbb);
758 // Helper function used to get a handle to the `__rust_try` function used to
761 // This function is only generated once and is then cached.
762 fn get_rust_try_fn<'ll, 'tcx>(
763 cx: &CodegenCx<'ll, 'tcx>,
764 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
765 ) -> (&'ll Type, &'ll Value) {
766 if let Some(llfn) = cx.rust_try_fn.get() {
770 // Define the type up front for the signature of the rust_try function.
772 let i8p = tcx.mk_mut_ptr(tcx.types.i8);
773 // `unsafe fn(*mut i8) -> ()`
774 let try_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
778 hir::Unsafety::Unsafe,
781 // `unsafe fn(*mut i8, *mut i8) -> ()`
782 let catch_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
783 [i8p, i8p].iter().cloned(),
786 hir::Unsafety::Unsafe,
789 // `unsafe fn(unsafe fn(*mut i8) -> (), *mut i8, unsafe fn(*mut i8, *mut i8) -> ()) -> i32`
790 let rust_fn_sig = ty::Binder::dummy(cx.tcx.mk_fn_sig(
791 [try_fn_ty, i8p, catch_fn_ty].into_iter(),
794 hir::Unsafety::Unsafe,
797 let rust_try = gen_fn(cx, "__rust_try", rust_fn_sig, codegen);
798 cx.rust_try_fn.set(Some(rust_try));
802 fn generic_simd_intrinsic<'ll, 'tcx>(
803 bx: &mut Builder<'_, 'll, 'tcx>,
806 args: &[OperandRef<'tcx, &'ll Value>],
810 ) -> Result<&'ll Value, ()> {
811 // macros for error handling:
812 macro_rules! emit_error {
816 ($msg: tt, $($fmt: tt)*) => {
817 span_invalid_monomorphization_error(
819 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
824 macro_rules! return_error {
827 emit_error!($($fmt)*);
833 macro_rules! require {
834 ($cond: expr, $($fmt: tt)*) => {
836 return_error!($($fmt)*);
841 macro_rules! require_simd {
842 ($ty: expr, $position: expr) => {
843 require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
849 tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), callee_ty.fn_sig(tcx));
850 let arg_tys = sig.inputs();
852 if name == sym::simd_select_bitmask {
853 require_simd!(arg_tys[1], "argument");
854 let (len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
856 let expected_int_bits = (len.max(8) - 1).next_power_of_two();
857 let expected_bytes = len / 8 + ((len % 8 > 0) as u64);
859 let mask_ty = arg_tys[0];
860 let mask = match mask_ty.kind() {
861 ty::Int(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
862 ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
864 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
865 && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
866 == Some(expected_bytes) =>
868 let place = PlaceRef::alloca(bx, args[0].layout);
869 args[0].val.store(bx, place);
870 let int_ty = bx.type_ix(expected_bytes * 8);
871 let ptr = bx.pointercast(place.llval, bx.cx.type_ptr_to(int_ty));
872 bx.load(int_ty, ptr, Align::ONE)
875 "invalid bitmask `{}`, expected `u{}` or `[u8; {}]`",
882 let i1 = bx.type_i1();
883 let im = bx.type_ix(len);
884 let i1xn = bx.type_vector(i1, len);
885 let m_im = bx.trunc(mask, im);
886 let m_i1s = bx.bitcast(m_im, i1xn);
887 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
890 // every intrinsic below takes a SIMD vector as its first argument
891 require_simd!(arg_tys[0], "input");
892 let in_ty = arg_tys[0];
894 let comparison = match name {
895 sym::simd_eq => Some(hir::BinOpKind::Eq),
896 sym::simd_ne => Some(hir::BinOpKind::Ne),
897 sym::simd_lt => Some(hir::BinOpKind::Lt),
898 sym::simd_le => Some(hir::BinOpKind::Le),
899 sym::simd_gt => Some(hir::BinOpKind::Gt),
900 sym::simd_ge => Some(hir::BinOpKind::Ge),
904 let (in_len, in_elem) = arg_tys[0].simd_size_and_type(bx.tcx());
905 if let Some(cmp_op) = comparison {
906 require_simd!(ret_ty, "return");
908 let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
911 "expected return type with length {} (same as input type `{}`), \
912 found `{}` with length {}",
919 bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
920 "expected return type with integer elements, found `{}` with non-integer `{}`",
925 return Ok(compare_simd_types(
935 if let Some(stripped) = name.as_str().strip_prefix("simd_shuffle") {
936 // If this intrinsic is the older "simd_shuffleN" form, simply parse the integer.
937 // If there is no suffix, use the index array length.
938 let n: u64 = if stripped.is_empty() {
939 // Make sure this is actually an array, since typeck only checks the length-suffixed
940 // version of this intrinsic.
941 match args[2].layout.ty.kind() {
942 ty::Array(ty, len) if matches!(ty.kind(), ty::Uint(ty::UintTy::U32)) => {
943 len.try_eval_usize(bx.cx.tcx, ty::ParamEnv::reveal_all()).unwrap_or_else(|| {
944 span_bug!(span, "could not evaluate shuffle index array length")
948 "simd_shuffle index must be an array of `u32`, got `{}`",
953 stripped.parse().unwrap_or_else(|_| {
954 span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
958 require_simd!(ret_ty, "return");
959 let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
962 "expected return type of length {}, found `{}` with length {}",
969 "expected return element type `{}` (element of input `{}`), \
970 found `{}` with element type `{}`",
977 let total_len = u128::from(in_len) * 2;
979 let vector = args[2].immediate();
981 let indices: Option<Vec<_>> = (0..n)
984 let val = bx.const_get_elt(vector, i as u64);
985 match bx.const_to_opt_u128(val, true) {
987 emit_error!("shuffle index #{} is not a constant", arg_idx);
990 Some(idx) if idx >= total_len => {
992 "shuffle index #{} is out of bounds (limit {})",
998 Some(idx) => Some(bx.const_i32(idx as i32)),
1002 let Some(indices) = indices else {
1003 return Ok(bx.const_null(llret_ty));
1006 return Ok(bx.shuffle_vector(
1007 args[0].immediate(),
1008 args[1].immediate(),
1009 bx.const_vector(&indices),
1013 if name == sym::simd_insert {
1015 in_elem == arg_tys[2],
1016 "expected inserted type `{}` (element of input `{}`), found `{}`",
1021 return Ok(bx.insert_element(
1022 args[0].immediate(),
1023 args[2].immediate(),
1024 args[1].immediate(),
1027 if name == sym::simd_extract {
1030 "expected return type `{}` (element of input `{}`), found `{}`",
1035 return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()));
1038 if name == sym::simd_select {
1039 let m_elem_ty = in_elem;
1041 require_simd!(arg_tys[1], "argument");
1042 let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1045 "mismatched lengths: mask length `{}` != other vector length `{}`",
1049 match m_elem_ty.kind() {
1051 _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
1053 // truncate the mask to a vector of i1s
1054 let i1 = bx.type_i1();
1055 let i1xn = bx.type_vector(i1, m_len as u64);
1056 let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1057 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1060 if name == sym::simd_bitmask {
1061 // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1062 // vector mask and returns the most significant bit (MSB) of each lane in the form
1064 // * an unsigned integer
1065 // * an array of `u8`
1066 // If the vector has less than 8 lanes, a u8 is returned with zeroed trailing bits.
1068 // The bit order of the result depends on the byte endianness, LSB-first for little
1069 // endian and MSB-first for big endian.
1070 let expected_int_bits = in_len.max(8);
1071 let expected_bytes = expected_int_bits / 8 + ((expected_int_bits % 8 > 0) as u64);
1073 // Integer vector <i{in_bitwidth} x in_len>:
1074 let (i_xn, in_elem_bitwidth) = match in_elem.kind() {
1076 args[0].immediate(),
1077 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1080 args[0].immediate(),
1081 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1084 "vector argument `{}`'s element type `{}`, expected integer element type",
1090 // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1093 bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
1096 let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1097 // Truncate vector to an <i1 x N>
1098 let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len));
1099 // Bitcast <i1 x N> to iN:
1100 let i_ = bx.bitcast(i1xn, bx.type_ix(in_len));
1102 match ret_ty.kind() {
1103 ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => {
1104 // Zero-extend iN to the bitmask type:
1105 return Ok(bx.zext(i_, bx.type_ix(expected_int_bits)));
1107 ty::Array(elem, len)
1108 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
1109 && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
1110 == Some(expected_bytes) =>
1112 // Zero-extend iN to the array lengh:
1113 let ze = bx.zext(i_, bx.type_ix(expected_bytes * 8));
1115 // Convert the integer to a byte array
1116 let ptr = bx.alloca(bx.type_ix(expected_bytes * 8), Align::ONE);
1117 bx.store(ze, ptr, Align::ONE);
1118 let array_ty = bx.type_array(bx.type_i8(), expected_bytes);
1119 let ptr = bx.pointercast(ptr, bx.cx.type_ptr_to(array_ty));
1120 return Ok(bx.load(array_ty, ptr, Align::ONE));
1123 "cannot return `{}`, expected `u{}` or `[u8; {}]`",
1131 fn simd_simple_float_intrinsic<'ll, 'tcx>(
1136 bx: &mut Builder<'_, 'll, 'tcx>,
1138 args: &[OperandRef<'tcx, &'ll Value>],
1139 ) -> Result<&'ll Value, ()> {
1140 macro_rules! emit_error {
1144 ($msg: tt, $($fmt: tt)*) => {
1145 span_invalid_monomorphization_error(
1147 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1151 macro_rules! return_error {
1154 emit_error!($($fmt)*);
1160 let (elem_ty_str, elem_ty) = if let ty::Float(f) = in_elem.kind() {
1161 let elem_ty = bx.cx.type_float_from_ty(*f);
1162 match f.bit_width() {
1163 32 => ("f32", elem_ty),
1164 64 => ("f64", elem_ty),
1167 "unsupported element type `{}` of floating-point vector `{}`",
1174 return_error!("`{}` is not a floating-point type", in_ty);
1177 let vec_ty = bx.type_vector(elem_ty, in_len);
1179 let (intr_name, fn_ty) = match name {
1180 sym::simd_ceil => ("ceil", bx.type_func(&[vec_ty], vec_ty)),
1181 sym::simd_fabs => ("fabs", bx.type_func(&[vec_ty], vec_ty)),
1182 sym::simd_fcos => ("cos", bx.type_func(&[vec_ty], vec_ty)),
1183 sym::simd_fexp2 => ("exp2", bx.type_func(&[vec_ty], vec_ty)),
1184 sym::simd_fexp => ("exp", bx.type_func(&[vec_ty], vec_ty)),
1185 sym::simd_flog10 => ("log10", bx.type_func(&[vec_ty], vec_ty)),
1186 sym::simd_flog2 => ("log2", bx.type_func(&[vec_ty], vec_ty)),
1187 sym::simd_flog => ("log", bx.type_func(&[vec_ty], vec_ty)),
1188 sym::simd_floor => ("floor", bx.type_func(&[vec_ty], vec_ty)),
1189 sym::simd_fma => ("fma", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)),
1190 sym::simd_fpowi => ("powi", bx.type_func(&[vec_ty, bx.type_i32()], vec_ty)),
1191 sym::simd_fpow => ("pow", bx.type_func(&[vec_ty, vec_ty], vec_ty)),
1192 sym::simd_fsin => ("sin", bx.type_func(&[vec_ty], vec_ty)),
1193 sym::simd_fsqrt => ("sqrt", bx.type_func(&[vec_ty], vec_ty)),
1194 sym::simd_round => ("round", bx.type_func(&[vec_ty], vec_ty)),
1195 sym::simd_trunc => ("trunc", bx.type_func(&[vec_ty], vec_ty)),
1196 _ => return_error!("unrecognized intrinsic `{}`", name),
1198 let llvm_name = &format!("llvm.{0}.v{1}{2}", intr_name, in_len, elem_ty_str);
1199 let f = bx.declare_cfn(llvm_name, llvm::UnnamedAddr::No, fn_ty);
1201 bx.call(fn_ty, f, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None);
1224 return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args);
1228 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1229 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1234 bx: &Builder<'_, '_, '_>,
1236 let p0s: String = "p0".repeat(no_pointers);
1237 match *elem_ty.kind() {
1238 ty::Int(v) => format!(
1242 // Normalize to prevent crash if v: IntTy::Isize
1243 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1245 ty::Uint(v) => format!(
1249 // Normalize to prevent crash if v: UIntTy::Usize
1250 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1252 ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1253 _ => unreachable!(),
1257 fn llvm_vector_ty<'ll>(
1258 cx: &CodegenCx<'ll, '_>,
1261 mut no_pointers: usize,
1263 // FIXME: use cx.layout_of(ty).llvm_type() ?
1264 let mut elem_ty = match *elem_ty.kind() {
1265 ty::Int(v) => cx.type_int_from_ty(v),
1266 ty::Uint(v) => cx.type_uint_from_ty(v),
1267 ty::Float(v) => cx.type_float_from_ty(v),
1268 _ => unreachable!(),
1270 while no_pointers > 0 {
1271 elem_ty = cx.type_ptr_to(elem_ty);
1274 cx.type_vector(elem_ty, vec_len)
1277 if name == sym::simd_gather {
1278 // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1279 // mask: <N x i{M}>) -> <N x T>
1280 // * N: number of elements in the input vectors
1281 // * T: type of the element to load
1282 // * M: any integer width is supported, will be truncated to i1
1284 // All types must be simd vector types
1285 require_simd!(in_ty, "first");
1286 require_simd!(arg_tys[1], "second");
1287 require_simd!(arg_tys[2], "third");
1288 require_simd!(ret_ty, "return");
1290 // Of the same length:
1291 let (out_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1292 let (out_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1295 "expected {} argument with length {} (same as input type `{}`), \
1296 found `{}` with length {}",
1305 "expected {} argument with length {} (same as input type `{}`), \
1306 found `{}` with length {}",
1314 // The return type must match the first argument type
1315 require!(ret_ty == in_ty, "expected return type `{}`, found `{}`", in_ty, ret_ty);
1317 // This counts how many pointers
1318 fn ptr_count(t: Ty<'_>) -> usize {
1320 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1326 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1328 ty::RawPtr(p) => non_ptr(p.ty),
1333 // The second argument must be a simd vector with an element type that's a pointer
1334 // to the element type of the first argument
1335 let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1336 let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1337 let (pointer_count, underlying_ty) = match element_ty1.kind() {
1338 ty::RawPtr(p) if p.ty == in_elem => (ptr_count(element_ty1), non_ptr(element_ty1)),
1342 "expected element type `{}` of second argument `{}` \
1343 to be a pointer to the element type `{}` of the first \
1344 argument `{}`, found `{}` != `*_ {}`",
1355 assert!(pointer_count > 0);
1356 assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1357 assert_eq!(underlying_ty, non_ptr(element_ty0));
1359 // The element type of the third argument must be a signed integer type of any width:
1360 let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1361 match element_ty2.kind() {
1366 "expected element type `{}` of third argument `{}` \
1367 to be a signed integer type",
1374 // Alignment of T, must be a constant integer value:
1375 let alignment_ty = bx.type_i32();
1376 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1378 // Truncate the mask vector to a vector of i1s:
1379 let (mask, mask_ty) = {
1380 let i1 = bx.type_i1();
1381 let i1xn = bx.type_vector(i1, in_len);
1382 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1385 // Type of the vector of pointers:
1386 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1387 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1389 // Type of the vector of elements:
1390 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1391 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1393 let llvm_intrinsic =
1394 format!("llvm.masked.gather.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1395 let fn_ty = bx.type_func(
1396 &[llvm_pointer_vec_ty, alignment_ty, mask_ty, llvm_elem_vec_ty],
1399 let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1401 bx.call(fn_ty, f, &[args[1].immediate(), alignment, mask, args[0].immediate()], None);
1405 if name == sym::simd_scatter {
1406 // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1407 // mask: <N x i{M}>) -> ()
1408 // * N: number of elements in the input vectors
1409 // * T: type of the element to load
1410 // * M: any integer width is supported, will be truncated to i1
1412 // All types must be simd vector types
1413 require_simd!(in_ty, "first");
1414 require_simd!(arg_tys[1], "second");
1415 require_simd!(arg_tys[2], "third");
1417 // Of the same length:
1418 let (element_len1, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1419 let (element_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1421 in_len == element_len1,
1422 "expected {} argument with length {} (same as input type `{}`), \
1423 found `{}` with length {}",
1431 in_len == element_len2,
1432 "expected {} argument with length {} (same as input type `{}`), \
1433 found `{}` with length {}",
1441 // This counts how many pointers
1442 fn ptr_count(t: Ty<'_>) -> usize {
1444 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1450 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1452 ty::RawPtr(p) => non_ptr(p.ty),
1457 // The second argument must be a simd vector with an element type that's a pointer
1458 // to the element type of the first argument
1459 let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1460 let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1461 let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1462 let (pointer_count, underlying_ty) = match element_ty1.kind() {
1463 ty::RawPtr(p) if p.ty == in_elem && p.mutbl == hir::Mutability::Mut => {
1464 (ptr_count(element_ty1), non_ptr(element_ty1))
1469 "expected element type `{}` of second argument `{}` \
1470 to be a pointer to the element type `{}` of the first \
1471 argument `{}`, found `{}` != `*mut {}`",
1482 assert!(pointer_count > 0);
1483 assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1484 assert_eq!(underlying_ty, non_ptr(element_ty0));
1486 // The element type of the third argument must be a signed integer type of any width:
1487 match element_ty2.kind() {
1492 "expected element type `{}` of third argument `{}` \
1493 be a signed integer type",
1500 // Alignment of T, must be a constant integer value:
1501 let alignment_ty = bx.type_i32();
1502 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1504 // Truncate the mask vector to a vector of i1s:
1505 let (mask, mask_ty) = {
1506 let i1 = bx.type_i1();
1507 let i1xn = bx.type_vector(i1, in_len);
1508 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1511 let ret_t = bx.type_void();
1513 // Type of the vector of pointers:
1514 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1515 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1517 // Type of the vector of elements:
1518 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1519 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1521 let llvm_intrinsic =
1522 format!("llvm.masked.scatter.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1524 bx.type_func(&[llvm_elem_vec_ty, llvm_pointer_vec_ty, alignment_ty, mask_ty], ret_t);
1525 let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1527 bx.call(fn_ty, f, &[args[0].immediate(), args[1].immediate(), alignment, mask], None);
1531 macro_rules! arith_red {
1532 ($name:ident : $integer_reduce:ident, $float_reduce:ident, $ordered:expr, $op:ident,
1533 $identity:expr) => {
1534 if name == sym::$name {
1537 "expected return type `{}` (element of input `{}`), found `{}`",
1542 return match in_elem.kind() {
1543 ty::Int(_) | ty::Uint(_) => {
1544 let r = bx.$integer_reduce(args[0].immediate());
1546 // if overflow occurs, the result is the
1547 // mathematical result modulo 2^n:
1548 Ok(bx.$op(args[1].immediate(), r))
1550 Ok(bx.$integer_reduce(args[0].immediate()))
1554 let acc = if $ordered {
1555 // ordered arithmetic reductions take an accumulator
1558 // unordered arithmetic reductions use the identity accumulator
1559 match f.bit_width() {
1560 32 => bx.const_real(bx.type_f32(), $identity),
1561 64 => bx.const_real(bx.type_f64(), $identity),
1564 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1573 Ok(bx.$float_reduce(acc, args[0].immediate()))
1576 "unsupported {} from `{}` with element `{}` to `{}`",
1587 arith_red!(simd_reduce_add_ordered: vector_reduce_add, vector_reduce_fadd, true, add, 0.0);
1588 arith_red!(simd_reduce_mul_ordered: vector_reduce_mul, vector_reduce_fmul, true, mul, 1.0);
1590 simd_reduce_add_unordered: vector_reduce_add,
1591 vector_reduce_fadd_fast,
1597 simd_reduce_mul_unordered: vector_reduce_mul,
1598 vector_reduce_fmul_fast,
1604 macro_rules! minmax_red {
1605 ($name:ident: $int_red:ident, $float_red:ident) => {
1606 if name == sym::$name {
1609 "expected return type `{}` (element of input `{}`), found `{}`",
1614 return match in_elem.kind() {
1615 ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
1616 ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
1617 ty::Float(_f) => Ok(bx.$float_red(args[0].immediate())),
1619 "unsupported {} from `{}` with element `{}` to `{}`",
1630 minmax_red!(simd_reduce_min: vector_reduce_min, vector_reduce_fmin);
1631 minmax_red!(simd_reduce_max: vector_reduce_max, vector_reduce_fmax);
1633 minmax_red!(simd_reduce_min_nanless: vector_reduce_min, vector_reduce_fmin_fast);
1634 minmax_red!(simd_reduce_max_nanless: vector_reduce_max, vector_reduce_fmax_fast);
1636 macro_rules! bitwise_red {
1637 ($name:ident : $red:ident, $boolean:expr) => {
1638 if name == sym::$name {
1639 let input = if !$boolean {
1642 "expected return type `{}` (element of input `{}`), found `{}`",
1649 match in_elem.kind() {
1650 ty::Int(_) | ty::Uint(_) => {}
1652 "unsupported {} from `{}` with element `{}` to `{}`",
1660 // boolean reductions operate on vectors of i1s:
1661 let i1 = bx.type_i1();
1662 let i1xn = bx.type_vector(i1, in_len as u64);
1663 bx.trunc(args[0].immediate(), i1xn)
1665 return match in_elem.kind() {
1666 ty::Int(_) | ty::Uint(_) => {
1667 let r = bx.$red(input);
1668 Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
1671 "unsupported {} from `{}` with element `{}` to `{}`",
1682 bitwise_red!(simd_reduce_and: vector_reduce_and, false);
1683 bitwise_red!(simd_reduce_or: vector_reduce_or, false);
1684 bitwise_red!(simd_reduce_xor: vector_reduce_xor, false);
1685 bitwise_red!(simd_reduce_all: vector_reduce_and, true);
1686 bitwise_red!(simd_reduce_any: vector_reduce_or, true);
1688 if name == sym::simd_cast || name == sym::simd_as {
1689 require_simd!(ret_ty, "return");
1690 let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1693 "expected return type with length {} (same as input type `{}`), \
1694 found `{}` with length {}",
1700 // casting cares about nominal type, not just structural type
1701 if in_elem == out_elem {
1702 return Ok(args[0].immediate());
1707 Int(/* is signed? */ bool),
1711 let (in_style, in_width) = match in_elem.kind() {
1712 // vectors of pointer-sized integers should've been
1713 // disallowed before here, so this unwrap is safe.
1716 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1720 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1722 ty::Float(f) => (Style::Float, f.bit_width()),
1723 _ => (Style::Unsupported, 0),
1725 let (out_style, out_width) = match out_elem.kind() {
1728 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1732 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1734 ty::Float(f) => (Style::Float, f.bit_width()),
1735 _ => (Style::Unsupported, 0),
1738 match (in_style, out_style) {
1739 (Style::Int(in_is_signed), Style::Int(_)) => {
1740 return Ok(match in_width.cmp(&out_width) {
1741 Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1742 Ordering::Equal => args[0].immediate(),
1745 bx.sext(args[0].immediate(), llret_ty)
1747 bx.zext(args[0].immediate(), llret_ty)
1752 (Style::Int(in_is_signed), Style::Float) => {
1753 return Ok(if in_is_signed {
1754 bx.sitofp(args[0].immediate(), llret_ty)
1756 bx.uitofp(args[0].immediate(), llret_ty)
1759 (Style::Float, Style::Int(out_is_signed)) => {
1760 return Ok(match (out_is_signed, name == sym::simd_as) {
1761 (false, false) => bx.fptoui(args[0].immediate(), llret_ty),
1762 (true, false) => bx.fptosi(args[0].immediate(), llret_ty),
1763 (_, true) => bx.cast_float_to_int(out_is_signed, args[0].immediate(), llret_ty),
1766 (Style::Float, Style::Float) => {
1767 return Ok(match in_width.cmp(&out_width) {
1768 Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1769 Ordering::Equal => args[0].immediate(),
1770 Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
1773 _ => { /* Unsupported. Fallthrough. */ }
1777 "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
1784 macro_rules! arith_binary {
1785 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1786 $(if name == sym::$name {
1787 match in_elem.kind() {
1788 $($(ty::$p(_))|* => {
1789 return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
1794 "unsupported operation on `{}` with element `{}`",
1801 simd_add: Uint, Int => add, Float => fadd;
1802 simd_sub: Uint, Int => sub, Float => fsub;
1803 simd_mul: Uint, Int => mul, Float => fmul;
1804 simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
1805 simd_rem: Uint => urem, Int => srem, Float => frem;
1806 simd_shl: Uint, Int => shl;
1807 simd_shr: Uint => lshr, Int => ashr;
1808 simd_and: Uint, Int => and;
1809 simd_or: Uint, Int => or;
1810 simd_xor: Uint, Int => xor;
1811 simd_fmax: Float => maxnum;
1812 simd_fmin: Float => minnum;
1815 macro_rules! arith_unary {
1816 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1817 $(if name == sym::$name {
1818 match in_elem.kind() {
1819 $($(ty::$p(_))|* => {
1820 return Ok(bx.$call(args[0].immediate()))
1825 "unsupported operation on `{}` with element `{}`",
1832 simd_neg: Int => neg, Float => fneg;
1835 if name == sym::simd_saturating_add || name == sym::simd_saturating_sub {
1836 let lhs = args[0].immediate();
1837 let rhs = args[1].immediate();
1838 let is_add = name == sym::simd_saturating_add;
1839 let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
1840 let (signed, elem_width, elem_ty) = match *in_elem.kind() {
1841 ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
1842 ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
1845 "expected element type `{}` of vector type `{}` \
1846 to be a signed or unsigned integer type",
1847 arg_tys[0].simd_size_and_type(bx.tcx()).1,
1852 let llvm_intrinsic = &format!(
1853 "llvm.{}{}.sat.v{}i{}",
1854 if signed { 's' } else { 'u' },
1855 if is_add { "add" } else { "sub" },
1859 let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
1861 let fn_ty = bx.type_func(&[vec_ty, vec_ty], vec_ty);
1862 let f = bx.declare_cfn(llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1863 let v = bx.call(fn_ty, f, &[lhs, rhs], None);
1867 span_bug!(span, "unknown SIMD intrinsic");
1870 // Returns the width of an int Ty, and if it's signed or not
1871 // Returns None if the type is not an integer
1872 // FIXME: there’s multiple of this functions, investigate using some of the already existing
1874 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
1877 Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), true))
1880 Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), false))