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) => {
137 match scalar.primitive() {
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 match self.cx.sess().target.arch.as_ref() {
333 "avr" | "msp430" => self.icmp(IntPredicate::IntEQ, cmp, self.const_i16(0)),
334 _ => self.icmp(IntPredicate::IntEQ, cmp, self.const_i32(0)),
340 args[0].val.store(self, result);
342 // We need to "use" the argument in some way LLVM can't introspect, and on
343 // targets that support it we can typically leverage inline assembly to do
344 // this. LLVM's interpretation of inline assembly is that it's, well, a black
345 // box. This isn't the greatest implementation since it probably deoptimizes
346 // more than we want, but it's so far good enough.
347 crate::asm::inline_asm_call(
355 llvm::AsmDialect::Att,
360 .unwrap_or_else(|| bug!("failed to generate inline asm call for `black_box`"));
362 // We have copied the value to `result` already.
366 _ if name.as_str().starts_with("simd_") => {
367 match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
373 _ => bug!("unknown intrinsic '{}'", name),
376 if !fn_abi.ret.is_ignore() {
377 if let PassMode::Cast(ty) = fn_abi.ret.mode {
378 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
379 let ptr = self.pointercast(result.llval, ptr_llty);
380 self.store(llval, ptr, result.align);
382 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
384 .store(self, result);
389 fn abort(&mut self) {
390 self.call_intrinsic("llvm.trap", &[]);
393 fn assume(&mut self, val: Self::Value) {
394 self.call_intrinsic("llvm.assume", &[val]);
397 fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
398 self.call_intrinsic("llvm.expect.i1", &[cond, self.const_bool(expected)])
401 fn type_test(&mut self, pointer: Self::Value, typeid: Self::Value) -> Self::Value {
402 // Test the called operand using llvm.type.test intrinsic. The LowerTypeTests link-time
403 // optimization pass replaces calls to this intrinsic with code to test type membership.
404 let i8p_ty = self.type_i8p();
405 let bitcast = self.bitcast(pointer, i8p_ty);
406 self.call_intrinsic("llvm.type.test", &[bitcast, typeid])
409 fn type_checked_load(
411 llvtable: &'ll Value,
412 vtable_byte_offset: u64,
415 let vtable_byte_offset = self.const_i32(vtable_byte_offset as i32);
416 self.call_intrinsic("llvm.type.checked.load", &[llvtable, vtable_byte_offset, typeid])
419 fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
420 self.call_intrinsic("llvm.va_start", &[va_list])
423 fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
424 self.call_intrinsic("llvm.va_end", &[va_list])
428 fn try_intrinsic<'ll>(
429 bx: &mut Builder<'_, 'll, '_>,
430 try_func: &'ll Value,
432 catch_func: &'ll Value,
435 if bx.sess().panic_strategy() == PanicStrategy::Abort {
436 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
437 bx.call(try_func_ty, try_func, &[data], None);
438 // Return 0 unconditionally from the intrinsic call;
439 // we can never unwind.
440 let ret_align = bx.tcx().data_layout.i32_align.abi;
441 bx.store(bx.const_i32(0), dest, ret_align);
442 } else if wants_msvc_seh(bx.sess()) {
443 codegen_msvc_try(bx, try_func, data, catch_func, dest);
444 } else if bx.sess().target.os == "emscripten" {
445 codegen_emcc_try(bx, try_func, data, catch_func, dest);
447 codegen_gnu_try(bx, try_func, data, catch_func, dest);
451 // MSVC's definition of the `rust_try` function.
453 // This implementation uses the new exception handling instructions in LLVM
454 // which have support in LLVM for SEH on MSVC targets. Although these
455 // instructions are meant to work for all targets, as of the time of this
456 // writing, however, LLVM does not recommend the usage of these new instructions
457 // as the old ones are still more optimized.
458 fn codegen_msvc_try<'ll>(
459 bx: &mut Builder<'_, 'll, '_>,
460 try_func: &'ll Value,
462 catch_func: &'ll Value,
465 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
466 bx.set_personality_fn(bx.eh_personality());
468 let normal = bx.append_sibling_block("normal");
469 let catchswitch = bx.append_sibling_block("catchswitch");
470 let catchpad_rust = bx.append_sibling_block("catchpad_rust");
471 let catchpad_foreign = bx.append_sibling_block("catchpad_foreign");
472 let caught = bx.append_sibling_block("caught");
474 let try_func = llvm::get_param(bx.llfn(), 0);
475 let data = llvm::get_param(bx.llfn(), 1);
476 let catch_func = llvm::get_param(bx.llfn(), 2);
478 // We're generating an IR snippet that looks like:
480 // declare i32 @rust_try(%try_func, %data, %catch_func) {
481 // %slot = alloca i8*
482 // invoke %try_func(%data) to label %normal unwind label %catchswitch
488 // %cs = catchswitch within none [%catchpad_rust, %catchpad_foreign] unwind to caller
491 // %tok = catchpad within %cs [%type_descriptor, 8, %slot]
493 // call %catch_func(%data, %ptr)
494 // catchret from %tok to label %caught
497 // %tok = catchpad within %cs [null, 64, null]
498 // call %catch_func(%data, null)
499 // catchret from %tok to label %caught
505 // This structure follows the basic usage of throw/try/catch in LLVM.
506 // For example, compile this C++ snippet to see what LLVM generates:
508 // struct rust_panic {
509 // rust_panic(const rust_panic&);
516 // void (*try_func)(void*),
518 // void (*catch_func)(void*, void*) noexcept
523 // } catch(rust_panic& a) {
524 // catch_func(data, &a);
527 // catch_func(data, NULL);
532 // More information can be found in libstd's seh.rs implementation.
533 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
534 let slot = bx.alloca(bx.type_i8p(), ptr_align);
535 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
536 bx.invoke(try_func_ty, try_func, &[data], normal, catchswitch, None);
538 bx.switch_to_block(normal);
539 bx.ret(bx.const_i32(0));
541 bx.switch_to_block(catchswitch);
542 let cs = bx.catch_switch(None, None, &[catchpad_rust, catchpad_foreign]);
544 // We can't use the TypeDescriptor defined in libpanic_unwind because it
545 // might be in another DLL and the SEH encoding only supports specifying
546 // a TypeDescriptor from the current module.
548 // However this isn't an issue since the MSVC runtime uses string
549 // comparison on the type name to match TypeDescriptors rather than
552 // So instead we generate a new TypeDescriptor in each module that uses
553 // `try` and let the linker merge duplicate definitions in the same
556 // When modifying, make sure that the type_name string exactly matches
557 // the one used in src/libpanic_unwind/seh.rs.
558 let type_info_vtable = bx.declare_global("??_7type_info@@6B@", bx.type_i8p());
559 let type_name = bx.const_bytes(b"rust_panic\0");
561 bx.const_struct(&[type_info_vtable, bx.const_null(bx.type_i8p()), type_name], false);
562 let tydesc = bx.declare_global("__rust_panic_type_info", bx.val_ty(type_info));
564 llvm::LLVMRustSetLinkage(tydesc, llvm::Linkage::LinkOnceODRLinkage);
565 llvm::SetUniqueComdat(bx.llmod, tydesc);
566 llvm::LLVMSetInitializer(tydesc, type_info);
569 // The flag value of 8 indicates that we are catching the exception by
570 // reference instead of by value. We can't use catch by value because
571 // that requires copying the exception object, which we don't support
572 // since our exception object effectively contains a Box.
574 // Source: MicrosoftCXXABI::getAddrOfCXXCatchHandlerType in clang
575 bx.switch_to_block(catchpad_rust);
576 let flags = bx.const_i32(8);
577 let funclet = bx.catch_pad(cs, &[tydesc, flags, slot]);
578 let ptr = bx.load(bx.type_i8p(), slot, ptr_align);
579 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
580 bx.call(catch_ty, catch_func, &[data, ptr], Some(&funclet));
581 bx.catch_ret(&funclet, caught);
583 // The flag value of 64 indicates a "catch-all".
584 bx.switch_to_block(catchpad_foreign);
585 let flags = bx.const_i32(64);
586 let null = bx.const_null(bx.type_i8p());
587 let funclet = bx.catch_pad(cs, &[null, flags, null]);
588 bx.call(catch_ty, catch_func, &[data, null], Some(&funclet));
589 bx.catch_ret(&funclet, caught);
591 bx.switch_to_block(caught);
592 bx.ret(bx.const_i32(1));
595 // Note that no invoke is used here because by definition this function
596 // can't panic (that's what it's catching).
597 let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
598 let i32_align = bx.tcx().data_layout.i32_align.abi;
599 bx.store(ret, dest, i32_align);
602 // Definition of the standard `try` function for Rust using the GNU-like model
603 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
606 // This codegen is a little surprising because we always call a shim
607 // function instead of inlining the call to `invoke` manually here. This is done
608 // because in LLVM we're only allowed to have one personality per function
609 // definition. The call to the `try` intrinsic is being inlined into the
610 // function calling it, and that function may already have other personality
611 // functions in play. By calling a shim we're guaranteed that our shim will have
612 // the right personality function.
613 fn codegen_gnu_try<'ll>(
614 bx: &mut Builder<'_, 'll, '_>,
615 try_func: &'ll Value,
617 catch_func: &'ll Value,
620 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
621 // Codegens the shims described above:
624 // invoke %try_func(%data) normal %normal unwind %catch
630 // (%ptr, _) = landingpad
631 // call %catch_func(%data, %ptr)
633 let then = bx.append_sibling_block("then");
634 let catch = bx.append_sibling_block("catch");
636 let try_func = llvm::get_param(bx.llfn(), 0);
637 let data = llvm::get_param(bx.llfn(), 1);
638 let catch_func = llvm::get_param(bx.llfn(), 2);
639 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
640 bx.invoke(try_func_ty, try_func, &[data], then, catch, None);
642 bx.switch_to_block(then);
643 bx.ret(bx.const_i32(0));
645 // Type indicator for the exception being thrown.
647 // The first value in this tuple is a pointer to the exception object
648 // being thrown. The second value is a "selector" indicating which of
649 // the landing pad clauses the exception's type had been matched to.
650 // rust_try ignores the selector.
651 bx.switch_to_block(catch);
652 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
653 let vals = bx.landing_pad(lpad_ty, bx.eh_personality(), 1);
654 let tydesc = bx.const_null(bx.type_i8p());
655 bx.add_clause(vals, tydesc);
656 let ptr = bx.extract_value(vals, 0);
657 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
658 bx.call(catch_ty, catch_func, &[data, ptr], None);
659 bx.ret(bx.const_i32(1));
662 // Note that no invoke is used here because by definition this function
663 // can't panic (that's what it's catching).
664 let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
665 let i32_align = bx.tcx().data_layout.i32_align.abi;
666 bx.store(ret, dest, i32_align);
669 // Variant of codegen_gnu_try used for emscripten where Rust panics are
670 // implemented using C++ exceptions. Here we use exceptions of a specific type
671 // (`struct rust_panic`) to represent Rust panics.
672 fn codegen_emcc_try<'ll>(
673 bx: &mut Builder<'_, 'll, '_>,
674 try_func: &'ll Value,
676 catch_func: &'ll Value,
679 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
680 // Codegens the shims described above:
683 // invoke %try_func(%data) normal %normal unwind %catch
689 // (%ptr, %selector) = landingpad
690 // %rust_typeid = @llvm.eh.typeid.for(@_ZTI10rust_panic)
691 // %is_rust_panic = %selector == %rust_typeid
692 // %catch_data = alloca { i8*, i8 }
693 // %catch_data[0] = %ptr
694 // %catch_data[1] = %is_rust_panic
695 // call %catch_func(%data, %catch_data)
697 let then = bx.append_sibling_block("then");
698 let catch = bx.append_sibling_block("catch");
700 let try_func = llvm::get_param(bx.llfn(), 0);
701 let data = llvm::get_param(bx.llfn(), 1);
702 let catch_func = llvm::get_param(bx.llfn(), 2);
703 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
704 bx.invoke(try_func_ty, try_func, &[data], then, catch, None);
706 bx.switch_to_block(then);
707 bx.ret(bx.const_i32(0));
709 // Type indicator for the exception being thrown.
711 // The first value in this tuple is a pointer to the exception object
712 // being thrown. The second value is a "selector" indicating which of
713 // the landing pad clauses the exception's type had been matched to.
714 bx.switch_to_block(catch);
715 let tydesc = bx.eh_catch_typeinfo();
716 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
717 let vals = bx.landing_pad(lpad_ty, bx.eh_personality(), 2);
718 bx.add_clause(vals, tydesc);
719 bx.add_clause(vals, bx.const_null(bx.type_i8p()));
720 let ptr = bx.extract_value(vals, 0);
721 let selector = bx.extract_value(vals, 1);
723 // Check if the typeid we got is the one for a Rust panic.
724 let rust_typeid = bx.call_intrinsic("llvm.eh.typeid.for", &[tydesc]);
725 let is_rust_panic = bx.icmp(IntPredicate::IntEQ, selector, rust_typeid);
726 let is_rust_panic = bx.zext(is_rust_panic, bx.type_bool());
728 // We need to pass two values to catch_func (ptr and is_rust_panic), so
729 // create an alloca and pass a pointer to that.
730 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
731 let i8_align = bx.tcx().data_layout.i8_align.abi;
732 let catch_data_type = bx.type_struct(&[bx.type_i8p(), bx.type_bool()], false);
733 let catch_data = bx.alloca(catch_data_type, ptr_align);
735 bx.inbounds_gep(catch_data_type, catch_data, &[bx.const_usize(0), bx.const_usize(0)]);
736 bx.store(ptr, catch_data_0, ptr_align);
738 bx.inbounds_gep(catch_data_type, catch_data, &[bx.const_usize(0), bx.const_usize(1)]);
739 bx.store(is_rust_panic, catch_data_1, i8_align);
740 let catch_data = bx.bitcast(catch_data, bx.type_i8p());
742 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
743 bx.call(catch_ty, catch_func, &[data, catch_data], None);
744 bx.ret(bx.const_i32(1));
747 // Note that no invoke is used here because by definition this function
748 // can't panic (that's what it's catching).
749 let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
750 let i32_align = bx.tcx().data_layout.i32_align.abi;
751 bx.store(ret, dest, i32_align);
754 // Helper function to give a Block to a closure to codegen a shim function.
755 // This is currently primarily used for the `try` intrinsic functions above.
756 fn gen_fn<'ll, 'tcx>(
757 cx: &CodegenCx<'ll, 'tcx>,
759 rust_fn_sig: ty::PolyFnSig<'tcx>,
760 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
761 ) -> (&'ll Type, &'ll Value) {
762 let fn_abi = cx.fn_abi_of_fn_ptr(rust_fn_sig, ty::List::empty());
763 let llty = fn_abi.llvm_type(cx);
764 let llfn = cx.declare_fn(name, fn_abi);
765 cx.set_frame_pointer_type(llfn);
766 cx.apply_target_cpu_attr(llfn);
767 // FIXME(eddyb) find a nicer way to do this.
768 unsafe { llvm::LLVMRustSetLinkage(llfn, llvm::Linkage::InternalLinkage) };
769 let llbb = Builder::append_block(cx, llfn, "entry-block");
770 let bx = Builder::build(cx, llbb);
775 // Helper function used to get a handle to the `__rust_try` function used to
778 // This function is only generated once and is then cached.
779 fn get_rust_try_fn<'ll, 'tcx>(
780 cx: &CodegenCx<'ll, 'tcx>,
781 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
782 ) -> (&'ll Type, &'ll Value) {
783 if let Some(llfn) = cx.rust_try_fn.get() {
787 // Define the type up front for the signature of the rust_try function.
789 let i8p = tcx.mk_mut_ptr(tcx.types.i8);
790 // `unsafe fn(*mut i8) -> ()`
791 let try_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
795 hir::Unsafety::Unsafe,
798 // `unsafe fn(*mut i8, *mut i8) -> ()`
799 let catch_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
800 [i8p, i8p].iter().cloned(),
803 hir::Unsafety::Unsafe,
806 // `unsafe fn(unsafe fn(*mut i8) -> (), *mut i8, unsafe fn(*mut i8, *mut i8) -> ()) -> i32`
807 let rust_fn_sig = ty::Binder::dummy(cx.tcx.mk_fn_sig(
808 [try_fn_ty, i8p, catch_fn_ty].into_iter(),
811 hir::Unsafety::Unsafe,
814 let rust_try = gen_fn(cx, "__rust_try", rust_fn_sig, codegen);
815 cx.rust_try_fn.set(Some(rust_try));
819 fn generic_simd_intrinsic<'ll, 'tcx>(
820 bx: &mut Builder<'_, 'll, 'tcx>,
823 args: &[OperandRef<'tcx, &'ll Value>],
827 ) -> Result<&'ll Value, ()> {
828 // macros for error handling:
829 #[allow(unused_macro_rules)]
830 macro_rules! emit_error {
834 ($msg: tt, $($fmt: tt)*) => {
835 span_invalid_monomorphization_error(
837 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
842 macro_rules! return_error {
845 emit_error!($($fmt)*);
851 macro_rules! require {
852 ($cond: expr, $($fmt: tt)*) => {
854 return_error!($($fmt)*);
859 macro_rules! require_simd {
860 ($ty: expr, $position: expr) => {
861 require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
867 tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), callee_ty.fn_sig(tcx));
868 let arg_tys = sig.inputs();
870 if name == sym::simd_select_bitmask {
871 require_simd!(arg_tys[1], "argument");
872 let (len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
874 let expected_int_bits = (len.max(8) - 1).next_power_of_two();
875 let expected_bytes = len / 8 + ((len % 8 > 0) as u64);
877 let mask_ty = arg_tys[0];
878 let mask = match mask_ty.kind() {
879 ty::Int(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
880 ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
882 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
883 && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
884 == Some(expected_bytes) =>
886 let place = PlaceRef::alloca(bx, args[0].layout);
887 args[0].val.store(bx, place);
888 let int_ty = bx.type_ix(expected_bytes * 8);
889 let ptr = bx.pointercast(place.llval, bx.cx.type_ptr_to(int_ty));
890 bx.load(int_ty, ptr, Align::ONE)
893 "invalid bitmask `{}`, expected `u{}` or `[u8; {}]`",
900 let i1 = bx.type_i1();
901 let im = bx.type_ix(len);
902 let i1xn = bx.type_vector(i1, len);
903 let m_im = bx.trunc(mask, im);
904 let m_i1s = bx.bitcast(m_im, i1xn);
905 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
908 // every intrinsic below takes a SIMD vector as its first argument
909 require_simd!(arg_tys[0], "input");
910 let in_ty = arg_tys[0];
912 let comparison = match name {
913 sym::simd_eq => Some(hir::BinOpKind::Eq),
914 sym::simd_ne => Some(hir::BinOpKind::Ne),
915 sym::simd_lt => Some(hir::BinOpKind::Lt),
916 sym::simd_le => Some(hir::BinOpKind::Le),
917 sym::simd_gt => Some(hir::BinOpKind::Gt),
918 sym::simd_ge => Some(hir::BinOpKind::Ge),
922 let (in_len, in_elem) = arg_tys[0].simd_size_and_type(bx.tcx());
923 if let Some(cmp_op) = comparison {
924 require_simd!(ret_ty, "return");
926 let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
929 "expected return type with length {} (same as input type `{}`), \
930 found `{}` with length {}",
937 bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
938 "expected return type with integer elements, found `{}` with non-integer `{}`",
943 return Ok(compare_simd_types(
953 if let Some(stripped) = name.as_str().strip_prefix("simd_shuffle") {
954 // If this intrinsic is the older "simd_shuffleN" form, simply parse the integer.
955 // If there is no suffix, use the index array length.
956 let n: u64 = if stripped.is_empty() {
957 // Make sure this is actually an array, since typeck only checks the length-suffixed
958 // version of this intrinsic.
959 match args[2].layout.ty.kind() {
960 ty::Array(ty, len) if matches!(ty.kind(), ty::Uint(ty::UintTy::U32)) => {
961 len.try_eval_usize(bx.cx.tcx, ty::ParamEnv::reveal_all()).unwrap_or_else(|| {
962 span_bug!(span, "could not evaluate shuffle index array length")
966 "simd_shuffle index must be an array of `u32`, got `{}`",
971 stripped.parse().unwrap_or_else(|_| {
972 span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
976 require_simd!(ret_ty, "return");
977 let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
980 "expected return type of length {}, found `{}` with length {}",
987 "expected return element type `{}` (element of input `{}`), \
988 found `{}` with element type `{}`",
995 let total_len = u128::from(in_len) * 2;
997 let vector = args[2].immediate();
999 let indices: Option<Vec<_>> = (0..n)
1002 let val = bx.const_get_elt(vector, i as u64);
1003 match bx.const_to_opt_u128(val, true) {
1005 emit_error!("shuffle index #{} is not a constant", arg_idx);
1008 Some(idx) if idx >= total_len => {
1010 "shuffle index #{} is out of bounds (limit {})",
1016 Some(idx) => Some(bx.const_i32(idx as i32)),
1020 let Some(indices) = indices else {
1021 return Ok(bx.const_null(llret_ty));
1024 return Ok(bx.shuffle_vector(
1025 args[0].immediate(),
1026 args[1].immediate(),
1027 bx.const_vector(&indices),
1031 if name == sym::simd_insert {
1033 in_elem == arg_tys[2],
1034 "expected inserted type `{}` (element of input `{}`), found `{}`",
1039 return Ok(bx.insert_element(
1040 args[0].immediate(),
1041 args[2].immediate(),
1042 args[1].immediate(),
1045 if name == sym::simd_extract {
1048 "expected return type `{}` (element of input `{}`), found `{}`",
1053 return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()));
1056 if name == sym::simd_select {
1057 let m_elem_ty = in_elem;
1059 require_simd!(arg_tys[1], "argument");
1060 let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1063 "mismatched lengths: mask length `{}` != other vector length `{}`",
1067 match m_elem_ty.kind() {
1069 _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
1071 // truncate the mask to a vector of i1s
1072 let i1 = bx.type_i1();
1073 let i1xn = bx.type_vector(i1, m_len as u64);
1074 let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1075 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1078 if name == sym::simd_bitmask {
1079 // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1080 // vector mask and returns the most significant bit (MSB) of each lane in the form
1082 // * an unsigned integer
1083 // * an array of `u8`
1084 // If the vector has less than 8 lanes, a u8 is returned with zeroed trailing bits.
1086 // The bit order of the result depends on the byte endianness, LSB-first for little
1087 // endian and MSB-first for big endian.
1088 let expected_int_bits = in_len.max(8);
1089 let expected_bytes = expected_int_bits / 8 + ((expected_int_bits % 8 > 0) as u64);
1091 // Integer vector <i{in_bitwidth} x in_len>:
1092 let (i_xn, in_elem_bitwidth) = match in_elem.kind() {
1094 args[0].immediate(),
1095 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1098 args[0].immediate(),
1099 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1102 "vector argument `{}`'s element type `{}`, expected integer element type",
1108 // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1111 bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
1114 let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1115 // Truncate vector to an <i1 x N>
1116 let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len));
1117 // Bitcast <i1 x N> to iN:
1118 let i_ = bx.bitcast(i1xn, bx.type_ix(in_len));
1120 match ret_ty.kind() {
1121 ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => {
1122 // Zero-extend iN to the bitmask type:
1123 return Ok(bx.zext(i_, bx.type_ix(expected_int_bits)));
1125 ty::Array(elem, len)
1126 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
1127 && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
1128 == Some(expected_bytes) =>
1130 // Zero-extend iN to the array length:
1131 let ze = bx.zext(i_, bx.type_ix(expected_bytes * 8));
1133 // Convert the integer to a byte array
1134 let ptr = bx.alloca(bx.type_ix(expected_bytes * 8), Align::ONE);
1135 bx.store(ze, ptr, Align::ONE);
1136 let array_ty = bx.type_array(bx.type_i8(), expected_bytes);
1137 let ptr = bx.pointercast(ptr, bx.cx.type_ptr_to(array_ty));
1138 return Ok(bx.load(array_ty, ptr, Align::ONE));
1141 "cannot return `{}`, expected `u{}` or `[u8; {}]`",
1149 fn simd_simple_float_intrinsic<'ll, 'tcx>(
1154 bx: &mut Builder<'_, 'll, 'tcx>,
1156 args: &[OperandRef<'tcx, &'ll Value>],
1157 ) -> Result<&'ll Value, ()> {
1158 #[allow(unused_macro_rules)]
1159 macro_rules! emit_error {
1163 ($msg: tt, $($fmt: tt)*) => {
1164 span_invalid_monomorphization_error(
1166 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1170 macro_rules! return_error {
1173 emit_error!($($fmt)*);
1179 let (elem_ty_str, elem_ty) = if let ty::Float(f) = in_elem.kind() {
1180 let elem_ty = bx.cx.type_float_from_ty(*f);
1181 match f.bit_width() {
1182 32 => ("f32", elem_ty),
1183 64 => ("f64", elem_ty),
1186 "unsupported element type `{}` of floating-point vector `{}`",
1193 return_error!("`{}` is not a floating-point type", in_ty);
1196 let vec_ty = bx.type_vector(elem_ty, in_len);
1198 let (intr_name, fn_ty) = match name {
1199 sym::simd_ceil => ("ceil", bx.type_func(&[vec_ty], vec_ty)),
1200 sym::simd_fabs => ("fabs", bx.type_func(&[vec_ty], vec_ty)),
1201 sym::simd_fcos => ("cos", bx.type_func(&[vec_ty], vec_ty)),
1202 sym::simd_fexp2 => ("exp2", bx.type_func(&[vec_ty], vec_ty)),
1203 sym::simd_fexp => ("exp", bx.type_func(&[vec_ty], vec_ty)),
1204 sym::simd_flog10 => ("log10", bx.type_func(&[vec_ty], vec_ty)),
1205 sym::simd_flog2 => ("log2", bx.type_func(&[vec_ty], vec_ty)),
1206 sym::simd_flog => ("log", bx.type_func(&[vec_ty], vec_ty)),
1207 sym::simd_floor => ("floor", bx.type_func(&[vec_ty], vec_ty)),
1208 sym::simd_fma => ("fma", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)),
1209 sym::simd_fpowi => ("powi", bx.type_func(&[vec_ty, bx.type_i32()], vec_ty)),
1210 sym::simd_fpow => ("pow", bx.type_func(&[vec_ty, vec_ty], vec_ty)),
1211 sym::simd_fsin => ("sin", bx.type_func(&[vec_ty], vec_ty)),
1212 sym::simd_fsqrt => ("sqrt", bx.type_func(&[vec_ty], vec_ty)),
1213 sym::simd_round => ("round", bx.type_func(&[vec_ty], vec_ty)),
1214 sym::simd_trunc => ("trunc", bx.type_func(&[vec_ty], vec_ty)),
1215 _ => return_error!("unrecognized intrinsic `{}`", name),
1217 let llvm_name = &format!("llvm.{0}.v{1}{2}", intr_name, in_len, elem_ty_str);
1218 let f = bx.declare_cfn(llvm_name, llvm::UnnamedAddr::No, fn_ty);
1220 bx.call(fn_ty, f, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None);
1243 return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args);
1247 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1248 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1253 bx: &Builder<'_, '_, '_>,
1255 let p0s: String = "p0".repeat(no_pointers);
1256 match *elem_ty.kind() {
1257 ty::Int(v) => format!(
1261 // Normalize to prevent crash if v: IntTy::Isize
1262 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1264 ty::Uint(v) => format!(
1268 // Normalize to prevent crash if v: UIntTy::Usize
1269 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1271 ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1272 _ => unreachable!(),
1276 fn llvm_vector_ty<'ll>(
1277 cx: &CodegenCx<'ll, '_>,
1280 mut no_pointers: usize,
1282 // FIXME: use cx.layout_of(ty).llvm_type() ?
1283 let mut elem_ty = match *elem_ty.kind() {
1284 ty::Int(v) => cx.type_int_from_ty(v),
1285 ty::Uint(v) => cx.type_uint_from_ty(v),
1286 ty::Float(v) => cx.type_float_from_ty(v),
1287 _ => unreachable!(),
1289 while no_pointers > 0 {
1290 elem_ty = cx.type_ptr_to(elem_ty);
1293 cx.type_vector(elem_ty, vec_len)
1296 if name == sym::simd_gather {
1297 // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1298 // mask: <N x i{M}>) -> <N x T>
1299 // * N: number of elements in the input vectors
1300 // * T: type of the element to load
1301 // * M: any integer width is supported, will be truncated to i1
1303 // All types must be simd vector types
1304 require_simd!(in_ty, "first");
1305 require_simd!(arg_tys[1], "second");
1306 require_simd!(arg_tys[2], "third");
1307 require_simd!(ret_ty, "return");
1309 // Of the same length:
1310 let (out_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1311 let (out_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1314 "expected {} argument with length {} (same as input type `{}`), \
1315 found `{}` with length {}",
1324 "expected {} argument with length {} (same as input type `{}`), \
1325 found `{}` with length {}",
1333 // The return type must match the first argument type
1334 require!(ret_ty == in_ty, "expected return type `{}`, found `{}`", in_ty, ret_ty);
1336 // This counts how many pointers
1337 fn ptr_count(t: Ty<'_>) -> usize {
1339 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1345 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1347 ty::RawPtr(p) => non_ptr(p.ty),
1352 // The second argument must be a simd vector with an element type that's a pointer
1353 // to the element type of the first argument
1354 let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1355 let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1356 let (pointer_count, underlying_ty) = match element_ty1.kind() {
1357 ty::RawPtr(p) if p.ty == in_elem => (ptr_count(element_ty1), non_ptr(element_ty1)),
1361 "expected element type `{}` of second argument `{}` \
1362 to be a pointer to the element type `{}` of the first \
1363 argument `{}`, found `{}` != `*_ {}`",
1374 assert!(pointer_count > 0);
1375 assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1376 assert_eq!(underlying_ty, non_ptr(element_ty0));
1378 // The element type of the third argument must be a signed integer type of any width:
1379 let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1380 match element_ty2.kind() {
1385 "expected element type `{}` of third argument `{}` \
1386 to be a signed integer type",
1393 // Alignment of T, must be a constant integer value:
1394 let alignment_ty = bx.type_i32();
1395 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1397 // Truncate the mask vector to a vector of i1s:
1398 let (mask, mask_ty) = {
1399 let i1 = bx.type_i1();
1400 let i1xn = bx.type_vector(i1, in_len);
1401 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1404 // Type of the vector of pointers:
1405 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1406 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1408 // Type of the vector of elements:
1409 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1410 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1412 let llvm_intrinsic =
1413 format!("llvm.masked.gather.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1414 let fn_ty = bx.type_func(
1415 &[llvm_pointer_vec_ty, alignment_ty, mask_ty, llvm_elem_vec_ty],
1418 let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1420 bx.call(fn_ty, f, &[args[1].immediate(), alignment, mask, args[0].immediate()], None);
1424 if name == sym::simd_scatter {
1425 // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1426 // mask: <N x i{M}>) -> ()
1427 // * N: number of elements in the input vectors
1428 // * T: type of the element to load
1429 // * M: any integer width is supported, will be truncated to i1
1431 // All types must be simd vector types
1432 require_simd!(in_ty, "first");
1433 require_simd!(arg_tys[1], "second");
1434 require_simd!(arg_tys[2], "third");
1436 // Of the same length:
1437 let (element_len1, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1438 let (element_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1440 in_len == element_len1,
1441 "expected {} argument with length {} (same as input type `{}`), \
1442 found `{}` with length {}",
1450 in_len == element_len2,
1451 "expected {} argument with length {} (same as input type `{}`), \
1452 found `{}` with length {}",
1460 // This counts how many pointers
1461 fn ptr_count(t: Ty<'_>) -> usize {
1463 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1469 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1471 ty::RawPtr(p) => non_ptr(p.ty),
1476 // The second argument must be a simd vector with an element type that's a pointer
1477 // to the element type of the first argument
1478 let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1479 let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1480 let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1481 let (pointer_count, underlying_ty) = match element_ty1.kind() {
1482 ty::RawPtr(p) if p.ty == in_elem && p.mutbl == hir::Mutability::Mut => {
1483 (ptr_count(element_ty1), non_ptr(element_ty1))
1488 "expected element type `{}` of second argument `{}` \
1489 to be a pointer to the element type `{}` of the first \
1490 argument `{}`, found `{}` != `*mut {}`",
1501 assert!(pointer_count > 0);
1502 assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1503 assert_eq!(underlying_ty, non_ptr(element_ty0));
1505 // The element type of the third argument must be a signed integer type of any width:
1506 match element_ty2.kind() {
1511 "expected element type `{}` of third argument `{}` \
1512 be a signed integer type",
1519 // Alignment of T, must be a constant integer value:
1520 let alignment_ty = bx.type_i32();
1521 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1523 // Truncate the mask vector to a vector of i1s:
1524 let (mask, mask_ty) = {
1525 let i1 = bx.type_i1();
1526 let i1xn = bx.type_vector(i1, in_len);
1527 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1530 let ret_t = bx.type_void();
1532 // Type of the vector of pointers:
1533 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1534 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1536 // Type of the vector of elements:
1537 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1538 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1540 let llvm_intrinsic =
1541 format!("llvm.masked.scatter.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1543 bx.type_func(&[llvm_elem_vec_ty, llvm_pointer_vec_ty, alignment_ty, mask_ty], ret_t);
1544 let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1546 bx.call(fn_ty, f, &[args[0].immediate(), args[1].immediate(), alignment, mask], None);
1550 macro_rules! arith_red {
1551 ($name:ident : $integer_reduce:ident, $float_reduce:ident, $ordered:expr, $op:ident,
1552 $identity:expr) => {
1553 if name == sym::$name {
1556 "expected return type `{}` (element of input `{}`), found `{}`",
1561 return match in_elem.kind() {
1562 ty::Int(_) | ty::Uint(_) => {
1563 let r = bx.$integer_reduce(args[0].immediate());
1565 // if overflow occurs, the result is the
1566 // mathematical result modulo 2^n:
1567 Ok(bx.$op(args[1].immediate(), r))
1569 Ok(bx.$integer_reduce(args[0].immediate()))
1573 let acc = if $ordered {
1574 // ordered arithmetic reductions take an accumulator
1577 // unordered arithmetic reductions use the identity accumulator
1578 match f.bit_width() {
1579 32 => bx.const_real(bx.type_f32(), $identity),
1580 64 => bx.const_real(bx.type_f64(), $identity),
1583 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1592 Ok(bx.$float_reduce(acc, args[0].immediate()))
1595 "unsupported {} from `{}` with element `{}` to `{}`",
1606 arith_red!(simd_reduce_add_ordered: vector_reduce_add, vector_reduce_fadd, true, add, 0.0);
1607 arith_red!(simd_reduce_mul_ordered: vector_reduce_mul, vector_reduce_fmul, true, mul, 1.0);
1609 simd_reduce_add_unordered: vector_reduce_add,
1610 vector_reduce_fadd_fast,
1616 simd_reduce_mul_unordered: vector_reduce_mul,
1617 vector_reduce_fmul_fast,
1623 macro_rules! minmax_red {
1624 ($name:ident: $int_red:ident, $float_red:ident) => {
1625 if name == sym::$name {
1628 "expected return type `{}` (element of input `{}`), found `{}`",
1633 return match in_elem.kind() {
1634 ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
1635 ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
1636 ty::Float(_f) => Ok(bx.$float_red(args[0].immediate())),
1638 "unsupported {} from `{}` with element `{}` to `{}`",
1649 minmax_red!(simd_reduce_min: vector_reduce_min, vector_reduce_fmin);
1650 minmax_red!(simd_reduce_max: vector_reduce_max, vector_reduce_fmax);
1652 minmax_red!(simd_reduce_min_nanless: vector_reduce_min, vector_reduce_fmin_fast);
1653 minmax_red!(simd_reduce_max_nanless: vector_reduce_max, vector_reduce_fmax_fast);
1655 macro_rules! bitwise_red {
1656 ($name:ident : $red:ident, $boolean:expr) => {
1657 if name == sym::$name {
1658 let input = if !$boolean {
1661 "expected return type `{}` (element of input `{}`), found `{}`",
1668 match in_elem.kind() {
1669 ty::Int(_) | ty::Uint(_) => {}
1671 "unsupported {} from `{}` with element `{}` to `{}`",
1679 // boolean reductions operate on vectors of i1s:
1680 let i1 = bx.type_i1();
1681 let i1xn = bx.type_vector(i1, in_len as u64);
1682 bx.trunc(args[0].immediate(), i1xn)
1684 return match in_elem.kind() {
1685 ty::Int(_) | ty::Uint(_) => {
1686 let r = bx.$red(input);
1687 Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
1690 "unsupported {} from `{}` with element `{}` to `{}`",
1701 bitwise_red!(simd_reduce_and: vector_reduce_and, false);
1702 bitwise_red!(simd_reduce_or: vector_reduce_or, false);
1703 bitwise_red!(simd_reduce_xor: vector_reduce_xor, false);
1704 bitwise_red!(simd_reduce_all: vector_reduce_and, true);
1705 bitwise_red!(simd_reduce_any: vector_reduce_or, true);
1707 if name == sym::simd_cast || name == sym::simd_as {
1708 require_simd!(ret_ty, "return");
1709 let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1712 "expected return type with length {} (same as input type `{}`), \
1713 found `{}` with length {}",
1719 // casting cares about nominal type, not just structural type
1720 if in_elem == out_elem {
1721 return Ok(args[0].immediate());
1726 Int(/* is signed? */ bool),
1730 let (in_style, in_width) = match in_elem.kind() {
1731 // vectors of pointer-sized integers should've been
1732 // disallowed before here, so this unwrap is safe.
1735 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1739 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1741 ty::Float(f) => (Style::Float, f.bit_width()),
1742 _ => (Style::Unsupported, 0),
1744 let (out_style, out_width) = match out_elem.kind() {
1747 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1751 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1753 ty::Float(f) => (Style::Float, f.bit_width()),
1754 _ => (Style::Unsupported, 0),
1757 match (in_style, out_style) {
1758 (Style::Int(in_is_signed), Style::Int(_)) => {
1759 return Ok(match in_width.cmp(&out_width) {
1760 Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1761 Ordering::Equal => args[0].immediate(),
1764 bx.sext(args[0].immediate(), llret_ty)
1766 bx.zext(args[0].immediate(), llret_ty)
1771 (Style::Int(in_is_signed), Style::Float) => {
1772 return Ok(if in_is_signed {
1773 bx.sitofp(args[0].immediate(), llret_ty)
1775 bx.uitofp(args[0].immediate(), llret_ty)
1778 (Style::Float, Style::Int(out_is_signed)) => {
1779 return Ok(match (out_is_signed, name == sym::simd_as) {
1780 (false, false) => bx.fptoui(args[0].immediate(), llret_ty),
1781 (true, false) => bx.fptosi(args[0].immediate(), llret_ty),
1782 (_, true) => bx.cast_float_to_int(out_is_signed, args[0].immediate(), llret_ty),
1785 (Style::Float, Style::Float) => {
1786 return Ok(match in_width.cmp(&out_width) {
1787 Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1788 Ordering::Equal => args[0].immediate(),
1789 Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
1792 _ => { /* Unsupported. Fallthrough. */ }
1796 "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
1803 macro_rules! arith_binary {
1804 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1805 $(if name == sym::$name {
1806 match in_elem.kind() {
1807 $($(ty::$p(_))|* => {
1808 return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
1813 "unsupported operation on `{}` with element `{}`",
1820 simd_add: Uint, Int => add, Float => fadd;
1821 simd_sub: Uint, Int => sub, Float => fsub;
1822 simd_mul: Uint, Int => mul, Float => fmul;
1823 simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
1824 simd_rem: Uint => urem, Int => srem, Float => frem;
1825 simd_shl: Uint, Int => shl;
1826 simd_shr: Uint => lshr, Int => ashr;
1827 simd_and: Uint, Int => and;
1828 simd_or: Uint, Int => or;
1829 simd_xor: Uint, Int => xor;
1830 simd_fmax: Float => maxnum;
1831 simd_fmin: Float => minnum;
1834 macro_rules! arith_unary {
1835 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1836 $(if name == sym::$name {
1837 match in_elem.kind() {
1838 $($(ty::$p(_))|* => {
1839 return Ok(bx.$call(args[0].immediate()))
1844 "unsupported operation on `{}` with element `{}`",
1851 simd_neg: Int => neg, Float => fneg;
1854 if name == sym::simd_arith_offset {
1855 // This also checks that the first operand is a ptr type.
1856 let pointee = in_elem.builtin_deref(true).unwrap_or_else(|| {
1857 span_bug!(span, "must be called with a vector of pointer types as first argument")
1859 let layout = bx.layout_of(pointee.ty);
1860 let ptrs = args[0].immediate();
1861 // The second argument must be a ptr-sized integer.
1862 // (We don't care about the signedness, this is wrapping anyway.)
1863 let (_offsets_len, offsets_elem) = arg_tys[1].simd_size_and_type(bx.tcx());
1864 if !matches!(offsets_elem.kind(), ty::Int(ty::IntTy::Isize) | ty::Uint(ty::UintTy::Usize)) {
1867 "must be called with a vector of pointer-sized integers as second argument"
1870 let offsets = args[1].immediate();
1872 return Ok(bx.gep(bx.backend_type(layout), ptrs, &[offsets]));
1875 if name == sym::simd_saturating_add || name == sym::simd_saturating_sub {
1876 let lhs = args[0].immediate();
1877 let rhs = args[1].immediate();
1878 let is_add = name == sym::simd_saturating_add;
1879 let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
1880 let (signed, elem_width, elem_ty) = match *in_elem.kind() {
1881 ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
1882 ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
1885 "expected element type `{}` of vector type `{}` \
1886 to be a signed or unsigned integer type",
1887 arg_tys[0].simd_size_and_type(bx.tcx()).1,
1892 let llvm_intrinsic = &format!(
1893 "llvm.{}{}.sat.v{}i{}",
1894 if signed { 's' } else { 'u' },
1895 if is_add { "add" } else { "sub" },
1899 let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
1901 let fn_ty = bx.type_func(&[vec_ty, vec_ty], vec_ty);
1902 let f = bx.declare_cfn(llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1903 let v = bx.call(fn_ty, f, &[lhs, rhs], None);
1907 span_bug!(span, "unknown SIMD intrinsic");
1910 // Returns the width of an int Ty, and if it's signed or not
1911 // Returns None if the type is not an integer
1912 // FIXME: there’s multiple of this functions, investigate using some of the already existing
1914 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
1917 Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), true))
1920 Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), false))