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::{FnAbiExt, HasTyCtxt};
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, HasDataLayout, LayoutOf, Primitive};
22 use rustc_target::spec::PanicStrategy;
24 use std::cmp::Ordering;
27 fn get_simple_intrinsic(cx: &CodegenCx<'ll, '_>, name: Symbol) -> Option<(&'ll Type, &'ll Value)> {
28 let llvm_name = match name {
29 sym::sqrtf32 => "llvm.sqrt.f32",
30 sym::sqrtf64 => "llvm.sqrt.f64",
31 sym::powif32 => "llvm.powi.f32",
32 sym::powif64 => "llvm.powi.f64",
33 sym::sinf32 => "llvm.sin.f32",
34 sym::sinf64 => "llvm.sin.f64",
35 sym::cosf32 => "llvm.cos.f32",
36 sym::cosf64 => "llvm.cos.f64",
37 sym::powf32 => "llvm.pow.f32",
38 sym::powf64 => "llvm.pow.f64",
39 sym::expf32 => "llvm.exp.f32",
40 sym::expf64 => "llvm.exp.f64",
41 sym::exp2f32 => "llvm.exp2.f32",
42 sym::exp2f64 => "llvm.exp2.f64",
43 sym::logf32 => "llvm.log.f32",
44 sym::logf64 => "llvm.log.f64",
45 sym::log10f32 => "llvm.log10.f32",
46 sym::log10f64 => "llvm.log10.f64",
47 sym::log2f32 => "llvm.log2.f32",
48 sym::log2f64 => "llvm.log2.f64",
49 sym::fmaf32 => "llvm.fma.f32",
50 sym::fmaf64 => "llvm.fma.f64",
51 sym::fabsf32 => "llvm.fabs.f32",
52 sym::fabsf64 => "llvm.fabs.f64",
53 sym::minnumf32 => "llvm.minnum.f32",
54 sym::minnumf64 => "llvm.minnum.f64",
55 sym::maxnumf32 => "llvm.maxnum.f32",
56 sym::maxnumf64 => "llvm.maxnum.f64",
57 sym::copysignf32 => "llvm.copysign.f32",
58 sym::copysignf64 => "llvm.copysign.f64",
59 sym::floorf32 => "llvm.floor.f32",
60 sym::floorf64 => "llvm.floor.f64",
61 sym::ceilf32 => "llvm.ceil.f32",
62 sym::ceilf64 => "llvm.ceil.f64",
63 sym::truncf32 => "llvm.trunc.f32",
64 sym::truncf64 => "llvm.trunc.f64",
65 sym::rintf32 => "llvm.rint.f32",
66 sym::rintf64 => "llvm.rint.f64",
67 sym::nearbyintf32 => "llvm.nearbyint.f32",
68 sym::nearbyintf64 => "llvm.nearbyint.f64",
69 sym::roundf32 => "llvm.round.f32",
70 sym::roundf64 => "llvm.round.f64",
73 Some(cx.get_intrinsic(&llvm_name))
76 impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
77 fn codegen_intrinsic_call(
79 instance: ty::Instance<'tcx>,
80 fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
81 args: &[OperandRef<'tcx, &'ll Value>],
86 let callee_ty = instance.ty(tcx, ty::ParamEnv::reveal_all());
88 let (def_id, substs) = match *callee_ty.kind() {
89 ty::FnDef(def_id, substs) => (def_id, substs),
90 _ => bug!("expected fn item type, found {}", callee_ty),
93 let sig = callee_ty.fn_sig(tcx);
94 let sig = tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), sig);
95 let arg_tys = sig.inputs();
96 let ret_ty = sig.output();
97 let name = tcx.item_name(def_id);
98 let name_str = &*name.as_str();
100 let llret_ty = self.layout_of(ret_ty).llvm_type(self);
101 let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
103 let simple = get_simple_intrinsic(self, name);
104 let llval = match name {
105 _ if simple.is_some() => {
106 let (simple_ty, simple_fn) = simple.unwrap();
110 &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
115 self.call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(true)])
117 sym::unlikely => self
118 .call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(false)]),
129 sym::breakpoint => self.call_intrinsic("llvm.debugtrap", &[]),
131 self.call_intrinsic("llvm.va_copy", &[args[0].immediate(), args[1].immediate()])
134 match fn_abi.ret.layout.abi {
135 abi::Abi::Scalar(ref scalar) => {
137 Primitive::Int(..) => {
138 if self.cx().size_of(ret_ty).bytes() < 4 {
139 // `va_arg` should not be called on a integer type
140 // less than 4 bytes in length. If it is, promote
141 // the integer to a `i32` and truncate the result
142 // back to the smaller type.
143 let promoted_result = emit_va_arg(self, args[0], tcx.types.i32);
144 self.trunc(promoted_result, llret_ty)
146 emit_va_arg(self, args[0], ret_ty)
149 Primitive::F64 | Primitive::Pointer => {
150 emit_va_arg(self, args[0], ret_ty)
152 // `va_arg` should never be used with the return type f32.
153 Primitive::F32 => bug!("the va_arg intrinsic does not work with `f32`"),
156 _ => bug!("the va_arg intrinsic does not work with non-scalar types"),
160 sym::volatile_load | sym::unaligned_volatile_load => {
161 let tp_ty = substs.type_at(0);
162 let ptr = args[0].immediate();
163 let load = if let PassMode::Cast(ty) = fn_abi.ret.mode {
164 let llty = ty.llvm_type(self);
165 let ptr = self.pointercast(ptr, self.type_ptr_to(llty));
166 self.volatile_load(llty, ptr)
168 self.volatile_load(self.layout_of(tp_ty).llvm_type(self), ptr)
170 let align = if name == sym::unaligned_volatile_load {
173 self.align_of(tp_ty).bytes() as u32
176 llvm::LLVMSetAlignment(load, align);
178 self.to_immediate(load, self.layout_of(tp_ty))
180 sym::volatile_store => {
181 let dst = args[0].deref(self.cx());
182 args[1].val.volatile_store(self, dst);
185 sym::unaligned_volatile_store => {
186 let dst = args[0].deref(self.cx());
187 args[1].val.unaligned_volatile_store(self, dst);
190 sym::prefetch_read_data
191 | sym::prefetch_write_data
192 | sym::prefetch_read_instruction
193 | sym::prefetch_write_instruction => {
194 let (rw, cache_type) = match name {
195 sym::prefetch_read_data => (0, 1),
196 sym::prefetch_write_data => (1, 1),
197 sym::prefetch_read_instruction => (0, 0),
198 sym::prefetch_write_instruction => (1, 0),
207 self.const_i32(cache_type),
220 | sym::saturating_add
221 | sym::saturating_sub => {
223 match int_type_width_signed(ty, self) {
224 Some((width, signed)) => match name {
225 sym::ctlz | sym::cttz => {
226 let y = self.const_bool(false);
228 &format!("llvm.{}.i{}", name, width),
229 &[args[0].immediate(), y],
232 sym::ctlz_nonzero | sym::cttz_nonzero => {
233 let y = self.const_bool(true);
234 let llvm_name = &format!("llvm.{}.i{}", &name_str[..4], width);
235 self.call_intrinsic(llvm_name, &[args[0].immediate(), y])
237 sym::ctpop => self.call_intrinsic(
238 &format!("llvm.ctpop.i{}", width),
239 &[args[0].immediate()],
243 args[0].immediate() // byte swap a u8/i8 is just a no-op
246 &format!("llvm.bswap.i{}", width),
247 &[args[0].immediate()],
251 sym::bitreverse => self.call_intrinsic(
252 &format!("llvm.bitreverse.i{}", width),
253 &[args[0].immediate()],
255 sym::rotate_left | sym::rotate_right => {
256 let is_left = name == sym::rotate_left;
257 let val = args[0].immediate();
258 let raw_shift = args[1].immediate();
259 // rotate = funnel shift with first two args the same
261 &format!("llvm.fsh{}.i{}", if is_left { 'l' } else { 'r' }, width);
262 self.call_intrinsic(llvm_name, &[val, val, raw_shift])
264 sym::saturating_add | sym::saturating_sub => {
265 let is_add = name == sym::saturating_add;
266 let lhs = args[0].immediate();
267 let rhs = args[1].immediate();
268 let llvm_name = &format!(
270 if signed { 's' } else { 'u' },
271 if is_add { "add" } else { "sub" },
274 self.call_intrinsic(llvm_name, &[lhs, rhs])
279 span_invalid_monomorphization_error(
283 "invalid monomorphization of `{}` intrinsic: \
284 expected basic integer type, found `{}`",
295 let tp_ty = substs.type_at(0);
296 let layout = self.layout_of(tp_ty).layout;
297 let use_integer_compare = match layout.abi {
298 Scalar(_) | ScalarPair(_, _) => true,
299 Uninhabited | Vector { .. } => false,
300 Aggregate { .. } => {
301 // For rusty ABIs, small aggregates are actually passed
302 // as `RegKind::Integer` (see `FnAbi::adjust_for_abi`),
303 // so we re-use that same threshold here.
304 layout.size <= self.data_layout().pointer_size * 2
308 let a = args[0].immediate();
309 let b = args[1].immediate();
310 if layout.size.bytes() == 0 {
311 self.const_bool(true)
312 } else if use_integer_compare {
313 let integer_ty = self.type_ix(layout.size.bits());
314 let ptr_ty = self.type_ptr_to(integer_ty);
315 let a_ptr = self.bitcast(a, ptr_ty);
316 let a_val = self.load(integer_ty, a_ptr, layout.align.abi);
317 let b_ptr = self.bitcast(b, ptr_ty);
318 let b_val = self.load(integer_ty, b_ptr, layout.align.abi);
319 self.icmp(IntPredicate::IntEQ, a_val, b_val)
321 let i8p_ty = self.type_i8p();
322 let a_ptr = self.bitcast(a, i8p_ty);
323 let b_ptr = self.bitcast(b, i8p_ty);
324 let n = self.const_usize(layout.size.bytes());
325 let cmp = self.call_intrinsic("memcmp", &[a_ptr, b_ptr, n]);
326 self.icmp(IntPredicate::IntEQ, cmp, self.const_i32(0))
330 _ if name_str.starts_with("simd_") => {
331 match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
337 _ => bug!("unknown intrinsic '{}'", name),
340 if !fn_abi.ret.is_ignore() {
341 if let PassMode::Cast(ty) = fn_abi.ret.mode {
342 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
343 let ptr = self.pointercast(result.llval, ptr_llty);
344 self.store(llval, ptr, result.align);
346 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
348 .store(self, result);
353 fn abort(&mut self) {
354 self.call_intrinsic("llvm.trap", &[]);
357 fn assume(&mut self, val: Self::Value) {
358 self.call_intrinsic("llvm.assume", &[val]);
361 fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
362 self.call_intrinsic("llvm.expect.i1", &[cond, self.const_bool(expected)])
365 fn sideeffect(&mut self) {
366 // This kind of check would make a ton of sense in the caller, but currently the only
367 // caller of this function is in `rustc_codegen_ssa`, which is agnostic to whether LLVM
368 // codegen backend being used, and so is unable to check the LLVM version.
369 if unsafe { llvm::LLVMRustVersionMajor() } < 12 {
370 self.call_intrinsic("llvm.sideeffect", &[]);
374 fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
375 self.call_intrinsic("llvm.va_start", &[va_list])
378 fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
379 self.call_intrinsic("llvm.va_end", &[va_list])
384 bx: &mut Builder<'a, 'll, 'tcx>,
385 try_func: &'ll Value,
387 catch_func: &'ll Value,
390 if bx.sess().panic_strategy() == PanicStrategy::Abort {
391 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
392 bx.call(try_func_ty, try_func, &[data], None);
393 // Return 0 unconditionally from the intrinsic call;
394 // we can never unwind.
395 let ret_align = bx.tcx().data_layout.i32_align.abi;
396 bx.store(bx.const_i32(0), dest, ret_align);
397 } else if wants_msvc_seh(bx.sess()) {
398 codegen_msvc_try(bx, try_func, data, catch_func, dest);
399 } else if bx.sess().target.is_like_emscripten {
400 codegen_emcc_try(bx, try_func, data, catch_func, dest);
402 codegen_gnu_try(bx, try_func, data, catch_func, dest);
406 // MSVC's definition of the `rust_try` function.
408 // This implementation uses the new exception handling instructions in LLVM
409 // which have support in LLVM for SEH on MSVC targets. Although these
410 // instructions are meant to work for all targets, as of the time of this
411 // writing, however, LLVM does not recommend the usage of these new instructions
412 // as the old ones are still more optimized.
414 bx: &mut Builder<'a, 'll, 'tcx>,
415 try_func: &'ll Value,
417 catch_func: &'ll Value,
420 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
421 bx.set_personality_fn(bx.eh_personality());
423 let mut normal = bx.build_sibling_block("normal");
424 let mut catchswitch = bx.build_sibling_block("catchswitch");
425 let mut catchpad_rust = bx.build_sibling_block("catchpad_rust");
426 let mut catchpad_foreign = bx.build_sibling_block("catchpad_foreign");
427 let mut caught = bx.build_sibling_block("caught");
429 let try_func = llvm::get_param(bx.llfn(), 0);
430 let data = llvm::get_param(bx.llfn(), 1);
431 let catch_func = llvm::get_param(bx.llfn(), 2);
433 // We're generating an IR snippet that looks like:
435 // declare i32 @rust_try(%try_func, %data, %catch_func) {
436 // %slot = alloca i8*
437 // invoke %try_func(%data) to label %normal unwind label %catchswitch
443 // %cs = catchswitch within none [%catchpad_rust, %catchpad_foreign] unwind to caller
446 // %tok = catchpad within %cs [%type_descriptor, 8, %slot]
448 // call %catch_func(%data, %ptr)
449 // catchret from %tok to label %caught
452 // %tok = catchpad within %cs [null, 64, null]
453 // call %catch_func(%data, null)
454 // catchret from %tok to label %caught
460 // This structure follows the basic usage of throw/try/catch in LLVM.
461 // For example, compile this C++ snippet to see what LLVM generates:
463 // struct rust_panic {
464 // rust_panic(const rust_panic&);
471 // void (*try_func)(void*),
473 // void (*catch_func)(void*, void*) noexcept
478 // } catch(rust_panic& a) {
479 // catch_func(data, &a);
482 // catch_func(data, NULL);
487 // More information can be found in libstd's seh.rs implementation.
488 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
489 let slot = bx.alloca(bx.type_i8p(), ptr_align);
490 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
491 bx.invoke(try_func_ty, try_func, &[data], normal.llbb(), catchswitch.llbb(), None);
493 normal.ret(bx.const_i32(0));
495 let cs = catchswitch.catch_switch(None, None, 2);
496 catchswitch.add_handler(cs, catchpad_rust.llbb());
497 catchswitch.add_handler(cs, catchpad_foreign.llbb());
499 // We can't use the TypeDescriptor defined in libpanic_unwind because it
500 // might be in another DLL and the SEH encoding only supports specifying
501 // a TypeDescriptor from the current module.
503 // However this isn't an issue since the MSVC runtime uses string
504 // comparison on the type name to match TypeDescriptors rather than
507 // So instead we generate a new TypeDescriptor in each module that uses
508 // `try` and let the linker merge duplicate definitions in the same
511 // When modifying, make sure that the type_name string exactly matches
512 // the one used in src/libpanic_unwind/seh.rs.
513 let type_info_vtable = bx.declare_global("??_7type_info@@6B@", bx.type_i8p());
514 let type_name = bx.const_bytes(b"rust_panic\0");
516 bx.const_struct(&[type_info_vtable, bx.const_null(bx.type_i8p()), type_name], false);
517 let tydesc = bx.declare_global("__rust_panic_type_info", bx.val_ty(type_info));
519 llvm::LLVMRustSetLinkage(tydesc, llvm::Linkage::LinkOnceODRLinkage);
520 llvm::SetUniqueComdat(bx.llmod, tydesc);
521 llvm::LLVMSetInitializer(tydesc, type_info);
524 // The flag value of 8 indicates that we are catching the exception by
525 // reference instead of by value. We can't use catch by value because
526 // that requires copying the exception object, which we don't support
527 // since our exception object effectively contains a Box.
529 // Source: MicrosoftCXXABI::getAddrOfCXXCatchHandlerType in clang
530 let flags = bx.const_i32(8);
531 let funclet = catchpad_rust.catch_pad(cs, &[tydesc, flags, slot]);
532 let ptr = catchpad_rust.load(bx.type_i8p(), slot, ptr_align);
533 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
534 catchpad_rust.call(catch_ty, catch_func, &[data, ptr], Some(&funclet));
535 catchpad_rust.catch_ret(&funclet, caught.llbb());
537 // The flag value of 64 indicates a "catch-all".
538 let flags = bx.const_i32(64);
539 let null = bx.const_null(bx.type_i8p());
540 let funclet = catchpad_foreign.catch_pad(cs, &[null, flags, null]);
541 catchpad_foreign.call(catch_ty, catch_func, &[data, null], Some(&funclet));
542 catchpad_foreign.catch_ret(&funclet, caught.llbb());
544 caught.ret(bx.const_i32(1));
547 // Note that no invoke is used here because by definition this function
548 // can't panic (that's what it's catching).
549 let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
550 let i32_align = bx.tcx().data_layout.i32_align.abi;
551 bx.store(ret, dest, i32_align);
554 // Definition of the standard `try` function for Rust using the GNU-like model
555 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
558 // This codegen is a little surprising because we always call a shim
559 // function instead of inlining the call to `invoke` manually here. This is done
560 // because in LLVM we're only allowed to have one personality per function
561 // definition. The call to the `try` intrinsic is being inlined into the
562 // function calling it, and that function may already have other personality
563 // functions in play. By calling a shim we're guaranteed that our shim will have
564 // the right personality function.
566 bx: &mut Builder<'a, 'll, 'tcx>,
567 try_func: &'ll Value,
569 catch_func: &'ll Value,
572 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
573 // Codegens the shims described above:
576 // invoke %try_func(%data) normal %normal unwind %catch
582 // (%ptr, _) = landingpad
583 // call %catch_func(%data, %ptr)
585 let mut then = bx.build_sibling_block("then");
586 let mut catch = bx.build_sibling_block("catch");
588 let try_func = llvm::get_param(bx.llfn(), 0);
589 let data = llvm::get_param(bx.llfn(), 1);
590 let catch_func = llvm::get_param(bx.llfn(), 2);
591 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
592 bx.invoke(try_func_ty, try_func, &[data], then.llbb(), catch.llbb(), None);
593 then.ret(bx.const_i32(0));
595 // Type indicator for the exception being thrown.
597 // The first value in this tuple is a pointer to the exception object
598 // being thrown. The second value is a "selector" indicating which of
599 // the landing pad clauses the exception's type had been matched to.
600 // rust_try ignores the selector.
601 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
602 let vals = catch.landing_pad(lpad_ty, bx.eh_personality(), 1);
603 let tydesc = bx.const_null(bx.type_i8p());
604 catch.add_clause(vals, tydesc);
605 let ptr = catch.extract_value(vals, 0);
606 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
607 catch.call(catch_ty, catch_func, &[data, ptr], None);
608 catch.ret(bx.const_i32(1));
611 // Note that no invoke is used here because by definition this function
612 // can't panic (that's what it's catching).
613 let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
614 let i32_align = bx.tcx().data_layout.i32_align.abi;
615 bx.store(ret, dest, i32_align);
618 // Variant of codegen_gnu_try used for emscripten where Rust panics are
619 // implemented using C++ exceptions. Here we use exceptions of a specific type
620 // (`struct rust_panic`) to represent Rust panics.
622 bx: &mut Builder<'a, 'll, 'tcx>,
623 try_func: &'ll Value,
625 catch_func: &'ll Value,
628 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
629 // Codegens the shims described above:
632 // invoke %try_func(%data) normal %normal unwind %catch
638 // (%ptr, %selector) = landingpad
639 // %rust_typeid = @llvm.eh.typeid.for(@_ZTI10rust_panic)
640 // %is_rust_panic = %selector == %rust_typeid
641 // %catch_data = alloca { i8*, i8 }
642 // %catch_data[0] = %ptr
643 // %catch_data[1] = %is_rust_panic
644 // call %catch_func(%data, %catch_data)
646 let mut then = bx.build_sibling_block("then");
647 let mut catch = bx.build_sibling_block("catch");
649 let try_func = llvm::get_param(bx.llfn(), 0);
650 let data = llvm::get_param(bx.llfn(), 1);
651 let catch_func = llvm::get_param(bx.llfn(), 2);
652 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
653 bx.invoke(try_func_ty, try_func, &[data], then.llbb(), catch.llbb(), None);
654 then.ret(bx.const_i32(0));
656 // Type indicator for the exception being thrown.
658 // The first value in this tuple is a pointer to the exception object
659 // being thrown. The second value is a "selector" indicating which of
660 // the landing pad clauses the exception's type had been matched to.
661 let tydesc = bx.eh_catch_typeinfo();
662 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
663 let vals = catch.landing_pad(lpad_ty, bx.eh_personality(), 2);
664 catch.add_clause(vals, tydesc);
665 catch.add_clause(vals, bx.const_null(bx.type_i8p()));
666 let ptr = catch.extract_value(vals, 0);
667 let selector = catch.extract_value(vals, 1);
669 // Check if the typeid we got is the one for a Rust panic.
670 let rust_typeid = catch.call_intrinsic("llvm.eh.typeid.for", &[tydesc]);
671 let is_rust_panic = catch.icmp(IntPredicate::IntEQ, selector, rust_typeid);
672 let is_rust_panic = catch.zext(is_rust_panic, bx.type_bool());
674 // We need to pass two values to catch_func (ptr and is_rust_panic), so
675 // create an alloca and pass a pointer to that.
676 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
677 let i8_align = bx.tcx().data_layout.i8_align.abi;
678 let catch_data_type = bx.type_struct(&[bx.type_i8p(), bx.type_bool()], false);
679 let catch_data = catch.alloca(catch_data_type, ptr_align);
680 let catch_data_0 = catch.inbounds_gep(
683 &[bx.const_usize(0), bx.const_usize(0)],
685 catch.store(ptr, catch_data_0, ptr_align);
686 let catch_data_1 = catch.inbounds_gep(
689 &[bx.const_usize(0), bx.const_usize(1)],
691 catch.store(is_rust_panic, catch_data_1, i8_align);
692 let catch_data = catch.bitcast(catch_data, bx.type_i8p());
694 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
695 catch.call(catch_ty, catch_func, &[data, catch_data], None);
696 catch.ret(bx.const_i32(1));
699 // Note that no invoke is used here because by definition this function
700 // can't panic (that's what it's catching).
701 let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
702 let i32_align = bx.tcx().data_layout.i32_align.abi;
703 bx.store(ret, dest, i32_align);
706 // Helper function to give a Block to a closure to codegen a shim function.
707 // This is currently primarily used for the `try` intrinsic functions above.
708 fn gen_fn<'ll, 'tcx>(
709 cx: &CodegenCx<'ll, 'tcx>,
711 rust_fn_sig: ty::PolyFnSig<'tcx>,
712 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
713 ) -> (&'ll Type, &'ll Value) {
714 let fn_abi = FnAbi::of_fn_ptr(cx, rust_fn_sig, &[]);
715 let llty = fn_abi.llvm_type(cx, false);
716 let llfn = cx.declare_fn(name, &fn_abi);
717 cx.set_frame_pointer_type(llfn);
718 cx.apply_target_cpu_attr(llfn);
719 // FIXME(eddyb) find a nicer way to do this.
720 unsafe { llvm::LLVMRustSetLinkage(llfn, llvm::Linkage::InternalLinkage) };
721 let llbb = Builder::append_block(cx, llfn, "entry-block");
722 let bx = Builder::build(cx, llbb);
727 // Helper function used to get a handle to the `__rust_try` function used to
730 // This function is only generated once and is then cached.
731 fn get_rust_try_fn<'ll, 'tcx>(
732 cx: &CodegenCx<'ll, 'tcx>,
733 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
734 ) -> (&'ll Type, &'ll Value) {
735 if let Some(llfn) = cx.rust_try_fn.get() {
739 // Define the type up front for the signature of the rust_try function.
741 let i8p = tcx.mk_mut_ptr(tcx.types.i8);
742 // `unsafe fn(*mut i8) -> ()`
743 let try_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
747 hir::Unsafety::Unsafe,
750 // `unsafe fn(*mut i8, *mut i8) -> ()`
751 let catch_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
752 [i8p, i8p].iter().cloned(),
755 hir::Unsafety::Unsafe,
758 // `unsafe fn(unsafe fn(*mut i8) -> (), *mut i8, unsafe fn(*mut i8, *mut i8) -> ()) -> i32`
759 let rust_fn_sig = ty::Binder::dummy(cx.tcx.mk_fn_sig(
760 vec![try_fn_ty, i8p, catch_fn_ty].into_iter(),
763 hir::Unsafety::Unsafe,
766 let rust_try = gen_fn(cx, "__rust_try", rust_fn_sig, codegen);
767 cx.rust_try_fn.set(Some(rust_try));
771 fn generic_simd_intrinsic(
772 bx: &mut Builder<'a, 'll, 'tcx>,
775 args: &[OperandRef<'tcx, &'ll Value>],
779 ) -> Result<&'ll Value, ()> {
780 // macros for error handling:
781 macro_rules! emit_error {
785 ($msg: tt, $($fmt: tt)*) => {
786 span_invalid_monomorphization_error(
788 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
793 macro_rules! return_error {
796 emit_error!($($fmt)*);
802 macro_rules! require {
803 ($cond: expr, $($fmt: tt)*) => {
805 return_error!($($fmt)*);
810 macro_rules! require_simd {
811 ($ty: expr, $position: expr) => {
812 require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
818 tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), callee_ty.fn_sig(tcx));
819 let arg_tys = sig.inputs();
820 let name_str = &*name.as_str();
822 if name == sym::simd_select_bitmask {
823 let in_ty = arg_tys[0];
824 let m_len = match in_ty.kind() {
825 // Note that this `.unwrap()` crashes for isize/usize, that's sort
826 // of intentional as there's not currently a use case for that.
827 ty::Int(i) => i.bit_width().unwrap(),
828 ty::Uint(i) => i.bit_width().unwrap(),
829 _ => return_error!("`{}` is not an integral type", in_ty),
831 require_simd!(arg_tys[1], "argument");
832 let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
834 // Allow masks for vectors with fewer than 8 elements to be
835 // represented with a u8 or i8.
836 m_len == v_len || (m_len == 8 && v_len < 8),
837 "mismatched lengths: mask length `{}` != other vector length `{}`",
841 let i1 = bx.type_i1();
842 let im = bx.type_ix(v_len);
843 let i1xn = bx.type_vector(i1, v_len);
844 let m_im = bx.trunc(args[0].immediate(), im);
845 let m_i1s = bx.bitcast(m_im, i1xn);
846 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
849 // every intrinsic below takes a SIMD vector as its first argument
850 require_simd!(arg_tys[0], "input");
851 let in_ty = arg_tys[0];
853 let comparison = match name {
854 sym::simd_eq => Some(hir::BinOpKind::Eq),
855 sym::simd_ne => Some(hir::BinOpKind::Ne),
856 sym::simd_lt => Some(hir::BinOpKind::Lt),
857 sym::simd_le => Some(hir::BinOpKind::Le),
858 sym::simd_gt => Some(hir::BinOpKind::Gt),
859 sym::simd_ge => Some(hir::BinOpKind::Ge),
863 let (in_len, in_elem) = arg_tys[0].simd_size_and_type(bx.tcx());
864 if let Some(cmp_op) = comparison {
865 require_simd!(ret_ty, "return");
867 let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
870 "expected return type with length {} (same as input type `{}`), \
871 found `{}` with length {}",
878 bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
879 "expected return type with integer elements, found `{}` with non-integer `{}`",
884 return Ok(compare_simd_types(
894 if let Some(stripped) = name_str.strip_prefix("simd_shuffle") {
895 let n: u64 = stripped.parse().unwrap_or_else(|_| {
896 span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
899 require_simd!(ret_ty, "return");
901 let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
904 "expected return type of length {}, found `{}` with length {}",
911 "expected return element type `{}` (element of input `{}`), \
912 found `{}` with element type `{}`",
919 let total_len = u128::from(in_len) * 2;
921 let vector = args[2].immediate();
923 let indices: Option<Vec<_>> = (0..n)
926 let val = bx.const_get_elt(vector, i as u64);
927 match bx.const_to_opt_u128(val, true) {
929 emit_error!("shuffle index #{} is not a constant", arg_idx);
932 Some(idx) if idx >= total_len => {
934 "shuffle index #{} is out of bounds (limit {})",
940 Some(idx) => Some(bx.const_i32(idx as i32)),
944 let indices = match indices {
946 None => return Ok(bx.const_null(llret_ty)),
949 return Ok(bx.shuffle_vector(
952 bx.const_vector(&indices),
956 if name == sym::simd_insert {
958 in_elem == arg_tys[2],
959 "expected inserted type `{}` (element of input `{}`), found `{}`",
964 return Ok(bx.insert_element(
970 if name == sym::simd_extract {
973 "expected return type `{}` (element of input `{}`), found `{}`",
978 return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()));
981 if name == sym::simd_select {
982 let m_elem_ty = in_elem;
984 require_simd!(arg_tys[1], "argument");
985 let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
988 "mismatched lengths: mask length `{}` != other vector length `{}`",
992 match m_elem_ty.kind() {
994 _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
996 // truncate the mask to a vector of i1s
997 let i1 = bx.type_i1();
998 let i1xn = bx.type_vector(i1, m_len as u64);
999 let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1000 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1003 if name == sym::simd_bitmask {
1004 // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1005 // vector mask and returns an unsigned integer containing the most
1006 // significant bit (MSB) of each lane.
1008 // If the vector has less than 8 lanes, an u8 is returned with zeroed
1010 let expected_int_bits = in_len.max(8);
1011 match ret_ty.kind() {
1012 ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => (),
1013 _ => return_error!("bitmask `{}`, expected `u{}`", ret_ty, expected_int_bits),
1016 // Integer vector <i{in_bitwidth} x in_len>:
1017 let (i_xn, in_elem_bitwidth) = match in_elem.kind() {
1019 args[0].immediate(),
1020 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1023 args[0].immediate(),
1024 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1027 "vector argument `{}`'s element type `{}`, expected integer element type",
1033 // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1036 bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
1039 let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1040 // Truncate vector to an <i1 x N>
1041 let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len));
1042 // Bitcast <i1 x N> to iN:
1043 let i_ = bx.bitcast(i1xn, bx.type_ix(in_len));
1044 // Zero-extend iN to the bitmask type:
1045 return Ok(bx.zext(i_, bx.type_ix(expected_int_bits)));
1048 fn simd_simple_float_intrinsic(
1050 in_elem: &::rustc_middle::ty::TyS<'_>,
1051 in_ty: &::rustc_middle::ty::TyS<'_>,
1053 bx: &mut Builder<'a, 'll, 'tcx>,
1055 args: &[OperandRef<'tcx, &'ll Value>],
1056 ) -> Result<&'ll Value, ()> {
1057 macro_rules! emit_error {
1061 ($msg: tt, $($fmt: tt)*) => {
1062 span_invalid_monomorphization_error(
1064 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1068 macro_rules! return_error {
1071 emit_error!($($fmt)*);
1077 let (elem_ty_str, elem_ty) = if let ty::Float(f) = in_elem.kind() {
1078 let elem_ty = bx.cx.type_float_from_ty(*f);
1079 match f.bit_width() {
1080 32 => ("f32", elem_ty),
1081 64 => ("f64", elem_ty),
1084 "unsupported element type `{}` of floating-point vector `{}`",
1091 return_error!("`{}` is not a floating-point type", in_ty);
1094 let vec_ty = bx.type_vector(elem_ty, in_len);
1096 let (intr_name, fn_ty) = match name {
1097 sym::simd_ceil => ("ceil", bx.type_func(&[vec_ty], vec_ty)),
1098 sym::simd_fabs => ("fabs", bx.type_func(&[vec_ty], vec_ty)),
1099 sym::simd_fcos => ("cos", bx.type_func(&[vec_ty], vec_ty)),
1100 sym::simd_fexp2 => ("exp2", bx.type_func(&[vec_ty], vec_ty)),
1101 sym::simd_fexp => ("exp", bx.type_func(&[vec_ty], vec_ty)),
1102 sym::simd_flog10 => ("log10", bx.type_func(&[vec_ty], vec_ty)),
1103 sym::simd_flog2 => ("log2", bx.type_func(&[vec_ty], vec_ty)),
1104 sym::simd_flog => ("log", bx.type_func(&[vec_ty], vec_ty)),
1105 sym::simd_floor => ("floor", bx.type_func(&[vec_ty], vec_ty)),
1106 sym::simd_fma => ("fma", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)),
1107 sym::simd_fpowi => ("powi", bx.type_func(&[vec_ty, bx.type_i32()], vec_ty)),
1108 sym::simd_fpow => ("pow", bx.type_func(&[vec_ty, vec_ty], vec_ty)),
1109 sym::simd_fsin => ("sin", bx.type_func(&[vec_ty], vec_ty)),
1110 sym::simd_fsqrt => ("sqrt", bx.type_func(&[vec_ty], vec_ty)),
1111 sym::simd_round => ("round", bx.type_func(&[vec_ty], vec_ty)),
1112 sym::simd_trunc => ("trunc", bx.type_func(&[vec_ty], vec_ty)),
1113 _ => return_error!("unrecognized intrinsic `{}`", name),
1115 let llvm_name = &format!("llvm.{0}.v{1}{2}", intr_name, in_len, elem_ty_str);
1116 let f = bx.declare_cfn(&llvm_name, llvm::UnnamedAddr::No, fn_ty);
1118 bx.call(fn_ty, f, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None);
1141 return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args);
1145 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1146 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1147 fn llvm_vector_str(elem_ty: Ty<'_>, vec_len: u64, no_pointers: usize) -> String {
1148 let p0s: String = "p0".repeat(no_pointers);
1149 match *elem_ty.kind() {
1150 ty::Int(v) => format!("v{}{}i{}", vec_len, p0s, v.bit_width().unwrap()),
1151 ty::Uint(v) => format!("v{}{}i{}", vec_len, p0s, v.bit_width().unwrap()),
1152 ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1153 _ => unreachable!(),
1158 cx: &CodegenCx<'ll, '_>,
1161 mut no_pointers: usize,
1163 // FIXME: use cx.layout_of(ty).llvm_type() ?
1164 let mut elem_ty = match *elem_ty.kind() {
1165 ty::Int(v) => cx.type_int_from_ty(v),
1166 ty::Uint(v) => cx.type_uint_from_ty(v),
1167 ty::Float(v) => cx.type_float_from_ty(v),
1168 _ => unreachable!(),
1170 while no_pointers > 0 {
1171 elem_ty = cx.type_ptr_to(elem_ty);
1174 cx.type_vector(elem_ty, vec_len)
1177 if name == sym::simd_gather {
1178 // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1179 // mask: <N x i{M}>) -> <N x T>
1180 // * N: number of elements in the input vectors
1181 // * T: type of the element to load
1182 // * M: any integer width is supported, will be truncated to i1
1184 // All types must be simd vector types
1185 require_simd!(in_ty, "first");
1186 require_simd!(arg_tys[1], "second");
1187 require_simd!(arg_tys[2], "third");
1188 require_simd!(ret_ty, "return");
1190 // Of the same length:
1191 let (out_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1192 let (out_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1195 "expected {} argument with length {} (same as input type `{}`), \
1196 found `{}` with length {}",
1205 "expected {} argument with length {} (same as input type `{}`), \
1206 found `{}` with length {}",
1214 // The return type must match the first argument type
1215 require!(ret_ty == in_ty, "expected return type `{}`, found `{}`", in_ty, ret_ty);
1217 // This counts how many pointers
1218 fn ptr_count(t: Ty<'_>) -> usize {
1220 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1226 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1228 ty::RawPtr(p) => non_ptr(p.ty),
1233 // The second argument must be a simd vector with an element type that's a pointer
1234 // to the element type of the first argument
1235 let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1236 let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1237 let (pointer_count, underlying_ty) = match element_ty1.kind() {
1238 ty::RawPtr(p) if p.ty == in_elem => (ptr_count(element_ty1), non_ptr(element_ty1)),
1242 "expected element type `{}` of second argument `{}` \
1243 to be a pointer to the element type `{}` of the first \
1244 argument `{}`, found `{}` != `*_ {}`",
1255 assert!(pointer_count > 0);
1256 assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1257 assert_eq!(underlying_ty, non_ptr(element_ty0));
1259 // The element type of the third argument must be a signed integer type of any width:
1260 let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1261 match element_ty2.kind() {
1266 "expected element type `{}` of third argument `{}` \
1267 to be a signed integer type",
1274 // Alignment of T, must be a constant integer value:
1275 let alignment_ty = bx.type_i32();
1276 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1278 // Truncate the mask vector to a vector of i1s:
1279 let (mask, mask_ty) = {
1280 let i1 = bx.type_i1();
1281 let i1xn = bx.type_vector(i1, in_len);
1282 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1285 // Type of the vector of pointers:
1286 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1287 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count);
1289 // Type of the vector of elements:
1290 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1291 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1);
1293 let llvm_intrinsic =
1294 format!("llvm.masked.gather.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1295 let fn_ty = bx.type_func(
1296 &[llvm_pointer_vec_ty, alignment_ty, mask_ty, llvm_elem_vec_ty],
1299 let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1301 bx.call(fn_ty, f, &[args[1].immediate(), alignment, mask, args[0].immediate()], None);
1305 if name == sym::simd_scatter {
1306 // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1307 // mask: <N x i{M}>) -> ()
1308 // * N: number of elements in the input vectors
1309 // * T: type of the element to load
1310 // * M: any integer width is supported, will be truncated to i1
1312 // All types must be simd vector types
1313 require_simd!(in_ty, "first");
1314 require_simd!(arg_tys[1], "second");
1315 require_simd!(arg_tys[2], "third");
1317 // Of the same length:
1318 let (element_len1, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1319 let (element_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1321 in_len == element_len1,
1322 "expected {} argument with length {} (same as input type `{}`), \
1323 found `{}` with length {}",
1331 in_len == element_len2,
1332 "expected {} argument with length {} (same as input type `{}`), \
1333 found `{}` with length {}",
1341 // This counts how many pointers
1342 fn ptr_count(t: Ty<'_>) -> usize {
1344 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1350 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1352 ty::RawPtr(p) => non_ptr(p.ty),
1357 // The second argument must be a simd vector with an element type that's a pointer
1358 // to the element type of the first argument
1359 let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1360 let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1361 let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1362 let (pointer_count, underlying_ty) = match element_ty1.kind() {
1363 ty::RawPtr(p) if p.ty == in_elem && p.mutbl == hir::Mutability::Mut => {
1364 (ptr_count(element_ty1), non_ptr(element_ty1))
1369 "expected element type `{}` of second argument `{}` \
1370 to be a pointer to the element type `{}` of the first \
1371 argument `{}`, found `{}` != `*mut {}`",
1382 assert!(pointer_count > 0);
1383 assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1384 assert_eq!(underlying_ty, non_ptr(element_ty0));
1386 // The element type of the third argument must be a signed integer type of any width:
1387 match element_ty2.kind() {
1392 "expected element type `{}` of third argument `{}` \
1393 be a signed integer type",
1400 // Alignment of T, must be a constant integer value:
1401 let alignment_ty = bx.type_i32();
1402 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1404 // Truncate the mask vector to a vector of i1s:
1405 let (mask, mask_ty) = {
1406 let i1 = bx.type_i1();
1407 let i1xn = bx.type_vector(i1, in_len);
1408 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1411 let ret_t = bx.type_void();
1413 // Type of the vector of pointers:
1414 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1415 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count);
1417 // Type of the vector of elements:
1418 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1419 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1);
1421 let llvm_intrinsic =
1422 format!("llvm.masked.scatter.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1424 bx.type_func(&[llvm_elem_vec_ty, llvm_pointer_vec_ty, alignment_ty, mask_ty], ret_t);
1425 let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1427 bx.call(fn_ty, f, &[args[0].immediate(), args[1].immediate(), alignment, mask], None);
1431 macro_rules! arith_red {
1432 ($name:ident : $integer_reduce:ident, $float_reduce:ident, $ordered:expr, $op:ident,
1433 $identity:expr) => {
1434 if name == sym::$name {
1437 "expected return type `{}` (element of input `{}`), found `{}`",
1442 return match in_elem.kind() {
1443 ty::Int(_) | ty::Uint(_) => {
1444 let r = bx.$integer_reduce(args[0].immediate());
1446 // if overflow occurs, the result is the
1447 // mathematical result modulo 2^n:
1448 Ok(bx.$op(args[1].immediate(), r))
1450 Ok(bx.$integer_reduce(args[0].immediate()))
1454 let acc = if $ordered {
1455 // ordered arithmetic reductions take an accumulator
1458 // unordered arithmetic reductions use the identity accumulator
1459 match f.bit_width() {
1460 32 => bx.const_real(bx.type_f32(), $identity),
1461 64 => bx.const_real(bx.type_f64(), $identity),
1464 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1473 Ok(bx.$float_reduce(acc, args[0].immediate()))
1476 "unsupported {} from `{}` with element `{}` to `{}`",
1487 arith_red!(simd_reduce_add_ordered: vector_reduce_add, vector_reduce_fadd, true, add, 0.0);
1488 arith_red!(simd_reduce_mul_ordered: vector_reduce_mul, vector_reduce_fmul, true, mul, 1.0);
1490 simd_reduce_add_unordered: vector_reduce_add,
1491 vector_reduce_fadd_fast,
1497 simd_reduce_mul_unordered: vector_reduce_mul,
1498 vector_reduce_fmul_fast,
1504 macro_rules! minmax_red {
1505 ($name:ident: $int_red:ident, $float_red:ident) => {
1506 if name == sym::$name {
1509 "expected return type `{}` (element of input `{}`), found `{}`",
1514 return match in_elem.kind() {
1515 ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
1516 ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
1517 ty::Float(_f) => Ok(bx.$float_red(args[0].immediate())),
1519 "unsupported {} from `{}` with element `{}` to `{}`",
1530 minmax_red!(simd_reduce_min: vector_reduce_min, vector_reduce_fmin);
1531 minmax_red!(simd_reduce_max: vector_reduce_max, vector_reduce_fmax);
1533 minmax_red!(simd_reduce_min_nanless: vector_reduce_min, vector_reduce_fmin_fast);
1534 minmax_red!(simd_reduce_max_nanless: vector_reduce_max, vector_reduce_fmax_fast);
1536 macro_rules! bitwise_red {
1537 ($name:ident : $red:ident, $boolean:expr) => {
1538 if name == sym::$name {
1539 let input = if !$boolean {
1542 "expected return type `{}` (element of input `{}`), found `{}`",
1549 match in_elem.kind() {
1550 ty::Int(_) | ty::Uint(_) => {}
1552 "unsupported {} from `{}` with element `{}` to `{}`",
1560 // boolean reductions operate on vectors of i1s:
1561 let i1 = bx.type_i1();
1562 let i1xn = bx.type_vector(i1, in_len as u64);
1563 bx.trunc(args[0].immediate(), i1xn)
1565 return match in_elem.kind() {
1566 ty::Int(_) | ty::Uint(_) => {
1567 let r = bx.$red(input);
1568 Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
1571 "unsupported {} from `{}` with element `{}` to `{}`",
1582 bitwise_red!(simd_reduce_and: vector_reduce_and, false);
1583 bitwise_red!(simd_reduce_or: vector_reduce_or, false);
1584 bitwise_red!(simd_reduce_xor: vector_reduce_xor, false);
1585 bitwise_red!(simd_reduce_all: vector_reduce_and, true);
1586 bitwise_red!(simd_reduce_any: vector_reduce_or, true);
1588 if name == sym::simd_cast {
1589 require_simd!(ret_ty, "return");
1590 let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1593 "expected return type with length {} (same as input type `{}`), \
1594 found `{}` with length {}",
1600 // casting cares about nominal type, not just structural type
1601 if in_elem == out_elem {
1602 return Ok(args[0].immediate());
1607 Int(/* is signed? */ bool),
1611 let (in_style, in_width) = match in_elem.kind() {
1612 // vectors of pointer-sized integers should've been
1613 // disallowed before here, so this unwrap is safe.
1614 ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1615 ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1616 ty::Float(f) => (Style::Float, f.bit_width()),
1617 _ => (Style::Unsupported, 0),
1619 let (out_style, out_width) = match out_elem.kind() {
1620 ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1621 ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1622 ty::Float(f) => (Style::Float, f.bit_width()),
1623 _ => (Style::Unsupported, 0),
1626 match (in_style, out_style) {
1627 (Style::Int(in_is_signed), Style::Int(_)) => {
1628 return Ok(match in_width.cmp(&out_width) {
1629 Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1630 Ordering::Equal => args[0].immediate(),
1633 bx.sext(args[0].immediate(), llret_ty)
1635 bx.zext(args[0].immediate(), llret_ty)
1640 (Style::Int(in_is_signed), Style::Float) => {
1641 return Ok(if in_is_signed {
1642 bx.sitofp(args[0].immediate(), llret_ty)
1644 bx.uitofp(args[0].immediate(), llret_ty)
1647 (Style::Float, Style::Int(out_is_signed)) => {
1648 return Ok(if out_is_signed {
1649 bx.fptosi(args[0].immediate(), llret_ty)
1651 bx.fptoui(args[0].immediate(), llret_ty)
1654 (Style::Float, Style::Float) => {
1655 return Ok(match in_width.cmp(&out_width) {
1656 Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1657 Ordering::Equal => args[0].immediate(),
1658 Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
1661 _ => { /* Unsupported. Fallthrough. */ }
1665 "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
1672 macro_rules! arith_binary {
1673 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1674 $(if name == sym::$name {
1675 match in_elem.kind() {
1676 $($(ty::$p(_))|* => {
1677 return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
1682 "unsupported operation on `{}` with element `{}`",
1689 simd_add: Uint, Int => add, Float => fadd;
1690 simd_sub: Uint, Int => sub, Float => fsub;
1691 simd_mul: Uint, Int => mul, Float => fmul;
1692 simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
1693 simd_rem: Uint => urem, Int => srem, Float => frem;
1694 simd_shl: Uint, Int => shl;
1695 simd_shr: Uint => lshr, Int => ashr;
1696 simd_and: Uint, Int => and;
1697 simd_or: Uint, Int => or;
1698 simd_xor: Uint, Int => xor;
1699 simd_fmax: Float => maxnum;
1700 simd_fmin: Float => minnum;
1703 macro_rules! arith_unary {
1704 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1705 $(if name == sym::$name {
1706 match in_elem.kind() {
1707 $($(ty::$p(_))|* => {
1708 return Ok(bx.$call(args[0].immediate()))
1713 "unsupported operation on `{}` with element `{}`",
1720 simd_neg: Int => neg, Float => fneg;
1723 if name == sym::simd_saturating_add || name == sym::simd_saturating_sub {
1724 let lhs = args[0].immediate();
1725 let rhs = args[1].immediate();
1726 let is_add = name == sym::simd_saturating_add;
1727 let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
1728 let (signed, elem_width, elem_ty) = match *in_elem.kind() {
1729 ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
1730 ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
1733 "expected element type `{}` of vector type `{}` \
1734 to be a signed or unsigned integer type",
1735 arg_tys[0].simd_size_and_type(bx.tcx()).1,
1740 let llvm_intrinsic = &format!(
1741 "llvm.{}{}.sat.v{}i{}",
1742 if signed { 's' } else { 'u' },
1743 if is_add { "add" } else { "sub" },
1747 let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
1749 let fn_ty = bx.type_func(&[vec_ty, vec_ty], vec_ty);
1750 let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1751 let v = bx.call(fn_ty, f, &[lhs, rhs], None);
1755 span_bug!(span, "unknown SIMD intrinsic");
1758 // Returns the width of an int Ty, and if it's signed or not
1759 // Returns None if the type is not an integer
1760 // FIXME: there’s multiple of this functions, investigate using some of the already existing
1762 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
1765 Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), true))
1768 Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), false))