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",
74 sym::ptr_mask => "llvm.ptrmask",
77 Some(cx.get_intrinsic(llvm_name))
80 impl<'ll, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'_, 'll, 'tcx> {
81 fn codegen_intrinsic_call(
83 instance: ty::Instance<'tcx>,
84 fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
85 args: &[OperandRef<'tcx, &'ll Value>],
90 let callee_ty = instance.ty(tcx, ty::ParamEnv::reveal_all());
92 let ty::FnDef(def_id, substs) = *callee_ty.kind() else {
93 bug!("expected fn item type, found {}", callee_ty);
96 let sig = callee_ty.fn_sig(tcx);
97 let sig = tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), sig);
98 let arg_tys = sig.inputs();
99 let ret_ty = sig.output();
100 let name = tcx.item_name(def_id);
102 let llret_ty = self.layout_of(ret_ty).llvm_type(self);
103 let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
105 let simple = get_simple_intrinsic(self, name);
106 let llval = match name {
107 _ if simple.is_some() => {
108 let (simple_ty, simple_fn) = simple.unwrap();
113 &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
118 self.call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(true)])
120 sym::unlikely => self
121 .call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(false)]),
132 sym::breakpoint => self.call_intrinsic("llvm.debugtrap", &[]),
134 self.call_intrinsic("llvm.va_copy", &[args[0].immediate(), args[1].immediate()])
137 match fn_abi.ret.layout.abi {
138 abi::Abi::Scalar(scalar) => {
139 match scalar.primitive() {
140 Primitive::Int(..) => {
141 if self.cx().size_of(ret_ty).bytes() < 4 {
142 // `va_arg` should not be called on an integer type
143 // less than 4 bytes in length. If it is, promote
144 // the integer to an `i32` and truncate the result
145 // back to the smaller type.
146 let promoted_result = emit_va_arg(self, args[0], tcx.types.i32);
147 self.trunc(promoted_result, llret_ty)
149 emit_va_arg(self, args[0], ret_ty)
152 Primitive::F64 | Primitive::Pointer => {
153 emit_va_arg(self, args[0], ret_ty)
155 // `va_arg` should never be used with the return type f32.
156 Primitive::F32 => bug!("the va_arg intrinsic does not work with `f32`"),
159 _ => bug!("the va_arg intrinsic does not work with non-scalar types"),
163 sym::volatile_load | sym::unaligned_volatile_load => {
164 let tp_ty = substs.type_at(0);
165 let ptr = args[0].immediate();
166 let load = if let PassMode::Cast(ty, _) = &fn_abi.ret.mode {
167 let llty = ty.llvm_type(self);
168 let ptr = self.pointercast(ptr, self.type_ptr_to(llty));
169 self.volatile_load(llty, ptr)
171 self.volatile_load(self.layout_of(tp_ty).llvm_type(self), ptr)
173 let align = if name == sym::unaligned_volatile_load {
176 self.align_of(tp_ty).bytes() as u32
179 llvm::LLVMSetAlignment(load, align);
181 self.to_immediate(load, self.layout_of(tp_ty))
183 sym::volatile_store => {
184 let dst = args[0].deref(self.cx());
185 args[1].val.volatile_store(self, dst);
188 sym::unaligned_volatile_store => {
189 let dst = args[0].deref(self.cx());
190 args[1].val.unaligned_volatile_store(self, dst);
193 sym::prefetch_read_data
194 | sym::prefetch_write_data
195 | sym::prefetch_read_instruction
196 | sym::prefetch_write_instruction => {
197 let (rw, cache_type) = match name {
198 sym::prefetch_read_data => (0, 1),
199 sym::prefetch_write_data => (1, 1),
200 sym::prefetch_read_instruction => (0, 0),
201 sym::prefetch_write_instruction => (1, 0),
210 self.const_i32(cache_type),
223 | sym::saturating_add
224 | sym::saturating_sub => {
226 match int_type_width_signed(ty, self) {
227 Some((width, signed)) => match name {
228 sym::ctlz | sym::cttz => {
229 let y = self.const_bool(false);
231 &format!("llvm.{}.i{}", name, width),
232 &[args[0].immediate(), y],
235 sym::ctlz_nonzero => {
236 let y = self.const_bool(true);
237 let llvm_name = &format!("llvm.ctlz.i{}", width);
238 self.call_intrinsic(llvm_name, &[args[0].immediate(), y])
240 sym::cttz_nonzero => {
241 let y = self.const_bool(true);
242 let llvm_name = &format!("llvm.cttz.i{}", width);
243 self.call_intrinsic(llvm_name, &[args[0].immediate(), y])
245 sym::ctpop => self.call_intrinsic(
246 &format!("llvm.ctpop.i{}", width),
247 &[args[0].immediate()],
251 args[0].immediate() // byte swap a u8/i8 is just a no-op
254 &format!("llvm.bswap.i{}", width),
255 &[args[0].immediate()],
259 sym::bitreverse => self.call_intrinsic(
260 &format!("llvm.bitreverse.i{}", width),
261 &[args[0].immediate()],
263 sym::rotate_left | sym::rotate_right => {
264 let is_left = name == sym::rotate_left;
265 let val = args[0].immediate();
266 let raw_shift = args[1].immediate();
267 // rotate = funnel shift with first two args the same
269 &format!("llvm.fsh{}.i{}", if is_left { 'l' } else { 'r' }, width);
270 self.call_intrinsic(llvm_name, &[val, val, raw_shift])
272 sym::saturating_add | sym::saturating_sub => {
273 let is_add = name == sym::saturating_add;
274 let lhs = args[0].immediate();
275 let rhs = args[1].immediate();
276 let llvm_name = &format!(
278 if signed { 's' } else { 'u' },
279 if is_add { "add" } else { "sub" },
282 self.call_intrinsic(llvm_name, &[lhs, rhs])
287 span_invalid_monomorphization_error(
291 "invalid monomorphization of `{}` intrinsic: \
292 expected basic integer type, found `{}`",
303 let tp_ty = substs.type_at(0);
304 let layout = self.layout_of(tp_ty).layout;
305 let use_integer_compare = match layout.abi() {
306 Scalar(_) | ScalarPair(_, _) => true,
307 Uninhabited | Vector { .. } => false,
308 Aggregate { .. } => {
309 // For rusty ABIs, small aggregates are actually passed
310 // as `RegKind::Integer` (see `FnAbi::adjust_for_abi`),
311 // so we re-use that same threshold here.
312 layout.size() <= self.data_layout().pointer_size * 2
316 let a = args[0].immediate();
317 let b = args[1].immediate();
318 if layout.size().bytes() == 0 {
319 self.const_bool(true)
320 } else if use_integer_compare {
321 let integer_ty = self.type_ix(layout.size().bits());
322 let ptr_ty = self.type_ptr_to(integer_ty);
323 let a_ptr = self.bitcast(a, ptr_ty);
324 let a_val = self.load(integer_ty, a_ptr, layout.align().abi);
325 let b_ptr = self.bitcast(b, ptr_ty);
326 let b_val = self.load(integer_ty, b_ptr, layout.align().abi);
327 self.icmp(IntPredicate::IntEQ, a_val, b_val)
329 let i8p_ty = self.type_i8p();
330 let a_ptr = self.bitcast(a, i8p_ty);
331 let b_ptr = self.bitcast(b, i8p_ty);
332 let n = self.const_usize(layout.size().bytes());
333 let cmp = self.call_intrinsic("memcmp", &[a_ptr, b_ptr, n]);
334 match self.cx.sess().target.arch.as_ref() {
335 "avr" | "msp430" => self.icmp(IntPredicate::IntEQ, cmp, self.const_i16(0)),
336 _ => self.icmp(IntPredicate::IntEQ, cmp, self.const_i32(0)),
342 args[0].val.store(self, result);
344 // We need to "use" the argument in some way LLVM can't introspect, and on
345 // targets that support it we can typically leverage inline assembly to do
346 // this. LLVM's interpretation of inline assembly is that it's, well, a black
347 // box. This isn't the greatest implementation since it probably deoptimizes
348 // more than we want, but it's so far good enough.
349 crate::asm::inline_asm_call(
357 llvm::AsmDialect::Att,
362 .unwrap_or_else(|| bug!("failed to generate inline asm call for `black_box`"));
364 // We have copied the value to `result` already.
368 _ if name.as_str().starts_with("simd_") => {
369 match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
375 _ => bug!("unknown intrinsic '{}'", name),
378 if !fn_abi.ret.is_ignore() {
379 if let PassMode::Cast(ty, _) = &fn_abi.ret.mode {
380 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
381 let ptr = self.pointercast(result.llval, ptr_llty);
382 self.store(llval, ptr, result.align);
384 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
386 .store(self, result);
391 fn abort(&mut self) {
392 self.call_intrinsic("llvm.trap", &[]);
395 fn assume(&mut self, val: Self::Value) {
396 self.call_intrinsic("llvm.assume", &[val]);
399 fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
400 self.call_intrinsic("llvm.expect.i1", &[cond, self.const_bool(expected)])
403 fn type_test(&mut self, pointer: Self::Value, typeid: Self::Value) -> Self::Value {
404 // Test the called operand using llvm.type.test intrinsic. The LowerTypeTests link-time
405 // optimization pass replaces calls to this intrinsic with code to test type membership.
406 let i8p_ty = self.type_i8p();
407 let bitcast = self.bitcast(pointer, i8p_ty);
408 self.call_intrinsic("llvm.type.test", &[bitcast, typeid])
411 fn type_checked_load(
413 llvtable: &'ll Value,
414 vtable_byte_offset: u64,
417 let vtable_byte_offset = self.const_i32(vtable_byte_offset as i32);
418 self.call_intrinsic("llvm.type.checked.load", &[llvtable, vtable_byte_offset, typeid])
421 fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
422 self.call_intrinsic("llvm.va_start", &[va_list])
425 fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
426 self.call_intrinsic("llvm.va_end", &[va_list])
430 fn try_intrinsic<'ll>(
431 bx: &mut Builder<'_, 'll, '_>,
432 try_func: &'ll Value,
434 catch_func: &'ll Value,
437 if bx.sess().panic_strategy() == PanicStrategy::Abort {
438 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
439 bx.call(try_func_ty, None, try_func, &[data], None);
440 // Return 0 unconditionally from the intrinsic call;
441 // we can never unwind.
442 let ret_align = bx.tcx().data_layout.i32_align.abi;
443 bx.store(bx.const_i32(0), dest, ret_align);
444 } else if wants_msvc_seh(bx.sess()) {
445 codegen_msvc_try(bx, try_func, data, catch_func, dest);
446 } else if bx.sess().target.os == "emscripten" {
447 codegen_emcc_try(bx, try_func, data, catch_func, dest);
449 codegen_gnu_try(bx, try_func, data, catch_func, dest);
453 // MSVC's definition of the `rust_try` function.
455 // This implementation uses the new exception handling instructions in LLVM
456 // which have support in LLVM for SEH on MSVC targets. Although these
457 // instructions are meant to work for all targets, as of the time of this
458 // writing, however, LLVM does not recommend the usage of these new instructions
459 // as the old ones are still more optimized.
460 fn codegen_msvc_try<'ll>(
461 bx: &mut Builder<'_, 'll, '_>,
462 try_func: &'ll Value,
464 catch_func: &'ll Value,
467 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
468 bx.set_personality_fn(bx.eh_personality());
470 let normal = bx.append_sibling_block("normal");
471 let catchswitch = bx.append_sibling_block("catchswitch");
472 let catchpad_rust = bx.append_sibling_block("catchpad_rust");
473 let catchpad_foreign = bx.append_sibling_block("catchpad_foreign");
474 let caught = bx.append_sibling_block("caught");
476 let try_func = llvm::get_param(bx.llfn(), 0);
477 let data = llvm::get_param(bx.llfn(), 1);
478 let catch_func = llvm::get_param(bx.llfn(), 2);
480 // We're generating an IR snippet that looks like:
482 // declare i32 @rust_try(%try_func, %data, %catch_func) {
483 // %slot = alloca i8*
484 // invoke %try_func(%data) to label %normal unwind label %catchswitch
490 // %cs = catchswitch within none [%catchpad_rust, %catchpad_foreign] unwind to caller
493 // %tok = catchpad within %cs [%type_descriptor, 8, %slot]
495 // call %catch_func(%data, %ptr)
496 // catchret from %tok to label %caught
499 // %tok = catchpad within %cs [null, 64, null]
500 // call %catch_func(%data, null)
501 // catchret from %tok to label %caught
507 // This structure follows the basic usage of throw/try/catch in LLVM.
508 // For example, compile this C++ snippet to see what LLVM generates:
510 // struct rust_panic {
511 // rust_panic(const rust_panic&);
518 // void (*try_func)(void*),
520 // void (*catch_func)(void*, void*) noexcept
525 // } catch(rust_panic& a) {
526 // catch_func(data, &a);
529 // catch_func(data, NULL);
534 // More information can be found in libstd's seh.rs implementation.
535 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
536 let slot = bx.alloca(bx.type_i8p(), ptr_align);
537 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
538 bx.invoke(try_func_ty, None, try_func, &[data], normal, catchswitch, None);
540 bx.switch_to_block(normal);
541 bx.ret(bx.const_i32(0));
543 bx.switch_to_block(catchswitch);
544 let cs = bx.catch_switch(None, None, &[catchpad_rust, catchpad_foreign]);
546 // We can't use the TypeDescriptor defined in libpanic_unwind because it
547 // might be in another DLL and the SEH encoding only supports specifying
548 // a TypeDescriptor from the current module.
550 // However this isn't an issue since the MSVC runtime uses string
551 // comparison on the type name to match TypeDescriptors rather than
554 // So instead we generate a new TypeDescriptor in each module that uses
555 // `try` and let the linker merge duplicate definitions in the same
558 // When modifying, make sure that the type_name string exactly matches
559 // the one used in src/libpanic_unwind/seh.rs.
560 let type_info_vtable = bx.declare_global("??_7type_info@@6B@", bx.type_i8p());
561 let type_name = bx.const_bytes(b"rust_panic\0");
563 bx.const_struct(&[type_info_vtable, bx.const_null(bx.type_i8p()), type_name], false);
564 let tydesc = bx.declare_global("__rust_panic_type_info", bx.val_ty(type_info));
566 llvm::LLVMRustSetLinkage(tydesc, llvm::Linkage::LinkOnceODRLinkage);
567 llvm::SetUniqueComdat(bx.llmod, tydesc);
568 llvm::LLVMSetInitializer(tydesc, type_info);
571 // The flag value of 8 indicates that we are catching the exception by
572 // reference instead of by value. We can't use catch by value because
573 // that requires copying the exception object, which we don't support
574 // since our exception object effectively contains a Box.
576 // Source: MicrosoftCXXABI::getAddrOfCXXCatchHandlerType in clang
577 bx.switch_to_block(catchpad_rust);
578 let flags = bx.const_i32(8);
579 let funclet = bx.catch_pad(cs, &[tydesc, flags, slot]);
580 let ptr = bx.load(bx.type_i8p(), slot, ptr_align);
581 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
582 bx.call(catch_ty, None, catch_func, &[data, ptr], Some(&funclet));
583 bx.catch_ret(&funclet, caught);
585 // The flag value of 64 indicates a "catch-all".
586 bx.switch_to_block(catchpad_foreign);
587 let flags = bx.const_i32(64);
588 let null = bx.const_null(bx.type_i8p());
589 let funclet = bx.catch_pad(cs, &[null, flags, null]);
590 bx.call(catch_ty, None, catch_func, &[data, null], Some(&funclet));
591 bx.catch_ret(&funclet, caught);
593 bx.switch_to_block(caught);
594 bx.ret(bx.const_i32(1));
597 // Note that no invoke is used here because by definition this function
598 // can't panic (that's what it's catching).
599 let ret = bx.call(llty, None, llfn, &[try_func, data, catch_func], None);
600 let i32_align = bx.tcx().data_layout.i32_align.abi;
601 bx.store(ret, dest, i32_align);
604 // Definition of the standard `try` function for Rust using the GNU-like model
605 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
608 // This codegen is a little surprising because we always call a shim
609 // function instead of inlining the call to `invoke` manually here. This is done
610 // because in LLVM we're only allowed to have one personality per function
611 // definition. The call to the `try` intrinsic is being inlined into the
612 // function calling it, and that function may already have other personality
613 // functions in play. By calling a shim we're guaranteed that our shim will have
614 // the right personality function.
615 fn codegen_gnu_try<'ll>(
616 bx: &mut Builder<'_, 'll, '_>,
617 try_func: &'ll Value,
619 catch_func: &'ll Value,
622 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
623 // Codegens the shims described above:
626 // invoke %try_func(%data) normal %normal unwind %catch
632 // (%ptr, _) = landingpad
633 // call %catch_func(%data, %ptr)
635 let then = bx.append_sibling_block("then");
636 let catch = bx.append_sibling_block("catch");
638 let try_func = llvm::get_param(bx.llfn(), 0);
639 let data = llvm::get_param(bx.llfn(), 1);
640 let catch_func = llvm::get_param(bx.llfn(), 2);
641 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
642 bx.invoke(try_func_ty, None, try_func, &[data], then, catch, None);
644 bx.switch_to_block(then);
645 bx.ret(bx.const_i32(0));
647 // Type indicator for the exception being thrown.
649 // The first value in this tuple is a pointer to the exception object
650 // being thrown. The second value is a "selector" indicating which of
651 // the landing pad clauses the exception's type had been matched to.
652 // rust_try ignores the selector.
653 bx.switch_to_block(catch);
654 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
655 let vals = bx.landing_pad(lpad_ty, bx.eh_personality(), 1);
656 let tydesc = bx.const_null(bx.type_i8p());
657 bx.add_clause(vals, tydesc);
658 let ptr = bx.extract_value(vals, 0);
659 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
660 bx.call(catch_ty, None, catch_func, &[data, ptr], None);
661 bx.ret(bx.const_i32(1));
664 // Note that no invoke is used here because by definition this function
665 // can't panic (that's what it's catching).
666 let ret = bx.call(llty, None, llfn, &[try_func, data, catch_func], None);
667 let i32_align = bx.tcx().data_layout.i32_align.abi;
668 bx.store(ret, dest, i32_align);
671 // Variant of codegen_gnu_try used for emscripten where Rust panics are
672 // implemented using C++ exceptions. Here we use exceptions of a specific type
673 // (`struct rust_panic`) to represent Rust panics.
674 fn codegen_emcc_try<'ll>(
675 bx: &mut Builder<'_, 'll, '_>,
676 try_func: &'ll Value,
678 catch_func: &'ll Value,
681 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
682 // Codegens the shims described above:
685 // invoke %try_func(%data) normal %normal unwind %catch
691 // (%ptr, %selector) = landingpad
692 // %rust_typeid = @llvm.eh.typeid.for(@_ZTI10rust_panic)
693 // %is_rust_panic = %selector == %rust_typeid
694 // %catch_data = alloca { i8*, i8 }
695 // %catch_data[0] = %ptr
696 // %catch_data[1] = %is_rust_panic
697 // call %catch_func(%data, %catch_data)
699 let then = bx.append_sibling_block("then");
700 let catch = bx.append_sibling_block("catch");
702 let try_func = llvm::get_param(bx.llfn(), 0);
703 let data = llvm::get_param(bx.llfn(), 1);
704 let catch_func = llvm::get_param(bx.llfn(), 2);
705 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
706 bx.invoke(try_func_ty, None, try_func, &[data], then, catch, None);
708 bx.switch_to_block(then);
709 bx.ret(bx.const_i32(0));
711 // Type indicator for the exception being thrown.
713 // The first value in this tuple is a pointer to the exception object
714 // being thrown. The second value is a "selector" indicating which of
715 // the landing pad clauses the exception's type had been matched to.
716 bx.switch_to_block(catch);
717 let tydesc = bx.eh_catch_typeinfo();
718 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
719 let vals = bx.landing_pad(lpad_ty, bx.eh_personality(), 2);
720 bx.add_clause(vals, tydesc);
721 bx.add_clause(vals, bx.const_null(bx.type_i8p()));
722 let ptr = bx.extract_value(vals, 0);
723 let selector = bx.extract_value(vals, 1);
725 // Check if the typeid we got is the one for a Rust panic.
726 let rust_typeid = bx.call_intrinsic("llvm.eh.typeid.for", &[tydesc]);
727 let is_rust_panic = bx.icmp(IntPredicate::IntEQ, selector, rust_typeid);
728 let is_rust_panic = bx.zext(is_rust_panic, bx.type_bool());
730 // We need to pass two values to catch_func (ptr and is_rust_panic), so
731 // create an alloca and pass a pointer to that.
732 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
733 let i8_align = bx.tcx().data_layout.i8_align.abi;
734 let catch_data_type = bx.type_struct(&[bx.type_i8p(), bx.type_bool()], false);
735 let catch_data = bx.alloca(catch_data_type, ptr_align);
737 bx.inbounds_gep(catch_data_type, catch_data, &[bx.const_usize(0), bx.const_usize(0)]);
738 bx.store(ptr, catch_data_0, ptr_align);
740 bx.inbounds_gep(catch_data_type, catch_data, &[bx.const_usize(0), bx.const_usize(1)]);
741 bx.store(is_rust_panic, catch_data_1, i8_align);
742 let catch_data = bx.bitcast(catch_data, bx.type_i8p());
744 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
745 bx.call(catch_ty, None, catch_func, &[data, catch_data], None);
746 bx.ret(bx.const_i32(1));
749 // Note that no invoke is used here because by definition this function
750 // can't panic (that's what it's catching).
751 let ret = bx.call(llty, None, llfn, &[try_func, data, catch_func], None);
752 let i32_align = bx.tcx().data_layout.i32_align.abi;
753 bx.store(ret, dest, i32_align);
756 // Helper function to give a Block to a closure to codegen a shim function.
757 // This is currently primarily used for the `try` intrinsic functions above.
758 fn gen_fn<'ll, 'tcx>(
759 cx: &CodegenCx<'ll, 'tcx>,
761 rust_fn_sig: ty::PolyFnSig<'tcx>,
762 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
763 ) -> (&'ll Type, &'ll Value) {
764 let fn_abi = cx.fn_abi_of_fn_ptr(rust_fn_sig, ty::List::empty());
765 let llty = fn_abi.llvm_type(cx);
766 let llfn = cx.declare_fn(name, fn_abi);
767 cx.set_frame_pointer_type(llfn);
768 cx.apply_target_cpu_attr(llfn);
769 // FIXME(eddyb) find a nicer way to do this.
770 unsafe { llvm::LLVMRustSetLinkage(llfn, llvm::Linkage::InternalLinkage) };
771 let llbb = Builder::append_block(cx, llfn, "entry-block");
772 let bx = Builder::build(cx, llbb);
777 // Helper function used to get a handle to the `__rust_try` function used to
780 // This function is only generated once and is then cached.
781 fn get_rust_try_fn<'ll, 'tcx>(
782 cx: &CodegenCx<'ll, 'tcx>,
783 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
784 ) -> (&'ll Type, &'ll Value) {
785 if let Some(llfn) = cx.rust_try_fn.get() {
789 // Define the type up front for the signature of the rust_try function.
791 let i8p = tcx.mk_mut_ptr(tcx.types.i8);
792 // `unsafe fn(*mut i8) -> ()`
793 let try_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
797 hir::Unsafety::Unsafe,
800 // `unsafe fn(*mut i8, *mut i8) -> ()`
801 let catch_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
802 [i8p, i8p].iter().cloned(),
805 hir::Unsafety::Unsafe,
808 // `unsafe fn(unsafe fn(*mut i8) -> (), *mut i8, unsafe fn(*mut i8, *mut i8) -> ()) -> i32`
809 let rust_fn_sig = ty::Binder::dummy(cx.tcx.mk_fn_sig(
810 [try_fn_ty, i8p, catch_fn_ty].into_iter(),
813 hir::Unsafety::Unsafe,
816 let rust_try = gen_fn(cx, "__rust_try", rust_fn_sig, codegen);
817 cx.rust_try_fn.set(Some(rust_try));
821 fn generic_simd_intrinsic<'ll, 'tcx>(
822 bx: &mut Builder<'_, 'll, 'tcx>,
825 args: &[OperandRef<'tcx, &'ll Value>],
829 ) -> Result<&'ll Value, ()> {
830 // macros for error handling:
831 #[allow(unused_macro_rules)]
832 macro_rules! emit_error {
836 ($msg: tt, $($fmt: tt)*) => {
837 span_invalid_monomorphization_error(
839 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
844 macro_rules! return_error {
847 emit_error!($($fmt)*);
853 macro_rules! require {
854 ($cond: expr, $($fmt: tt)*) => {
856 return_error!($($fmt)*);
861 macro_rules! require_simd {
862 ($ty: expr, $position: expr) => {
863 require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
869 tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), callee_ty.fn_sig(tcx));
870 let arg_tys = sig.inputs();
872 if name == sym::simd_select_bitmask {
873 require_simd!(arg_tys[1], "argument");
874 let (len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
876 let expected_int_bits = (len.max(8) - 1).next_power_of_two();
877 let expected_bytes = len / 8 + ((len % 8 > 0) as u64);
879 let mask_ty = arg_tys[0];
880 let mask = match mask_ty.kind() {
881 ty::Int(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
882 ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
884 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
885 && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
886 == Some(expected_bytes) =>
888 let place = PlaceRef::alloca(bx, args[0].layout);
889 args[0].val.store(bx, place);
890 let int_ty = bx.type_ix(expected_bytes * 8);
891 let ptr = bx.pointercast(place.llval, bx.cx.type_ptr_to(int_ty));
892 bx.load(int_ty, ptr, Align::ONE)
895 "invalid bitmask `{}`, expected `u{}` or `[u8; {}]`",
902 let i1 = bx.type_i1();
903 let im = bx.type_ix(len);
904 let i1xn = bx.type_vector(i1, len);
905 let m_im = bx.trunc(mask, im);
906 let m_i1s = bx.bitcast(m_im, i1xn);
907 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
910 // every intrinsic below takes a SIMD vector as its first argument
911 require_simd!(arg_tys[0], "input");
912 let in_ty = arg_tys[0];
914 let comparison = match name {
915 sym::simd_eq => Some(hir::BinOpKind::Eq),
916 sym::simd_ne => Some(hir::BinOpKind::Ne),
917 sym::simd_lt => Some(hir::BinOpKind::Lt),
918 sym::simd_le => Some(hir::BinOpKind::Le),
919 sym::simd_gt => Some(hir::BinOpKind::Gt),
920 sym::simd_ge => Some(hir::BinOpKind::Ge),
924 let (in_len, in_elem) = arg_tys[0].simd_size_and_type(bx.tcx());
925 if let Some(cmp_op) = comparison {
926 require_simd!(ret_ty, "return");
928 let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
931 "expected return type with length {} (same as input type `{}`), \
932 found `{}` with length {}",
939 bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
940 "expected return type with integer elements, found `{}` with non-integer `{}`",
945 return Ok(compare_simd_types(
955 if let Some(stripped) = name.as_str().strip_prefix("simd_shuffle") {
956 // If this intrinsic is the older "simd_shuffleN" form, simply parse the integer.
957 // If there is no suffix, use the index array length.
958 let n: u64 = if stripped.is_empty() {
959 // Make sure this is actually an array, since typeck only checks the length-suffixed
960 // version of this intrinsic.
961 match args[2].layout.ty.kind() {
962 ty::Array(ty, len) if matches!(ty.kind(), ty::Uint(ty::UintTy::U32)) => {
963 len.try_eval_usize(bx.cx.tcx, ty::ParamEnv::reveal_all()).unwrap_or_else(|| {
964 span_bug!(span, "could not evaluate shuffle index array length")
968 "simd_shuffle index must be an array of `u32`, got `{}`",
973 stripped.parse().unwrap_or_else(|_| {
974 span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
978 require_simd!(ret_ty, "return");
979 let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
982 "expected return type of length {}, found `{}` with length {}",
989 "expected return element type `{}` (element of input `{}`), \
990 found `{}` with element type `{}`",
997 let total_len = u128::from(in_len) * 2;
999 let vector = args[2].immediate();
1001 let indices: Option<Vec<_>> = (0..n)
1004 let val = bx.const_get_elt(vector, i as u64);
1005 match bx.const_to_opt_u128(val, true) {
1007 emit_error!("shuffle index #{} is not a constant", arg_idx);
1010 Some(idx) if idx >= total_len => {
1012 "shuffle index #{} is out of bounds (limit {})",
1018 Some(idx) => Some(bx.const_i32(idx as i32)),
1022 let Some(indices) = indices else {
1023 return Ok(bx.const_null(llret_ty));
1026 return Ok(bx.shuffle_vector(
1027 args[0].immediate(),
1028 args[1].immediate(),
1029 bx.const_vector(&indices),
1033 if name == sym::simd_insert {
1035 in_elem == arg_tys[2],
1036 "expected inserted type `{}` (element of input `{}`), found `{}`",
1041 return Ok(bx.insert_element(
1042 args[0].immediate(),
1043 args[2].immediate(),
1044 args[1].immediate(),
1047 if name == sym::simd_extract {
1050 "expected return type `{}` (element of input `{}`), found `{}`",
1055 return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()));
1058 if name == sym::simd_select {
1059 let m_elem_ty = in_elem;
1061 require_simd!(arg_tys[1], "argument");
1062 let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1065 "mismatched lengths: mask length `{}` != other vector length `{}`",
1069 match m_elem_ty.kind() {
1071 _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
1073 // truncate the mask to a vector of i1s
1074 let i1 = bx.type_i1();
1075 let i1xn = bx.type_vector(i1, m_len as u64);
1076 let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1077 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1080 if name == sym::simd_bitmask {
1081 // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1082 // vector mask and returns the most significant bit (MSB) of each lane in the form
1084 // * an unsigned integer
1085 // * an array of `u8`
1086 // If the vector has less than 8 lanes, a u8 is returned with zeroed trailing bits.
1088 // The bit order of the result depends on the byte endianness, LSB-first for little
1089 // endian and MSB-first for big endian.
1090 let expected_int_bits = in_len.max(8);
1091 let expected_bytes = expected_int_bits / 8 + ((expected_int_bits % 8 > 0) as u64);
1093 // Integer vector <i{in_bitwidth} x in_len>:
1094 let (i_xn, in_elem_bitwidth) = match in_elem.kind() {
1096 args[0].immediate(),
1097 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1100 args[0].immediate(),
1101 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1104 "vector argument `{}`'s element type `{}`, expected integer element type",
1110 // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1113 bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
1116 let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1117 // Truncate vector to an <i1 x N>
1118 let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len));
1119 // Bitcast <i1 x N> to iN:
1120 let i_ = bx.bitcast(i1xn, bx.type_ix(in_len));
1122 match ret_ty.kind() {
1123 ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => {
1124 // Zero-extend iN to the bitmask type:
1125 return Ok(bx.zext(i_, bx.type_ix(expected_int_bits)));
1127 ty::Array(elem, len)
1128 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
1129 && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
1130 == Some(expected_bytes) =>
1132 // Zero-extend iN to the array length:
1133 let ze = bx.zext(i_, bx.type_ix(expected_bytes * 8));
1135 // Convert the integer to a byte array
1136 let ptr = bx.alloca(bx.type_ix(expected_bytes * 8), Align::ONE);
1137 bx.store(ze, ptr, Align::ONE);
1138 let array_ty = bx.type_array(bx.type_i8(), expected_bytes);
1139 let ptr = bx.pointercast(ptr, bx.cx.type_ptr_to(array_ty));
1140 return Ok(bx.load(array_ty, ptr, Align::ONE));
1143 "cannot return `{}`, expected `u{}` or `[u8; {}]`",
1151 fn simd_simple_float_intrinsic<'ll, 'tcx>(
1156 bx: &mut Builder<'_, 'll, 'tcx>,
1158 args: &[OperandRef<'tcx, &'ll Value>],
1159 ) -> Result<&'ll Value, ()> {
1160 #[allow(unused_macro_rules)]
1161 macro_rules! emit_error {
1165 ($msg: tt, $($fmt: tt)*) => {
1166 span_invalid_monomorphization_error(
1168 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1172 macro_rules! return_error {
1175 emit_error!($($fmt)*);
1181 let (elem_ty_str, elem_ty) = if let ty::Float(f) = in_elem.kind() {
1182 let elem_ty = bx.cx.type_float_from_ty(*f);
1183 match f.bit_width() {
1184 32 => ("f32", elem_ty),
1185 64 => ("f64", elem_ty),
1188 "unsupported element type `{}` of floating-point vector `{}`",
1195 return_error!("`{}` is not a floating-point type", in_ty);
1198 let vec_ty = bx.type_vector(elem_ty, in_len);
1200 let (intr_name, fn_ty) = match name {
1201 sym::simd_ceil => ("ceil", bx.type_func(&[vec_ty], vec_ty)),
1202 sym::simd_fabs => ("fabs", bx.type_func(&[vec_ty], vec_ty)),
1203 sym::simd_fcos => ("cos", bx.type_func(&[vec_ty], vec_ty)),
1204 sym::simd_fexp2 => ("exp2", bx.type_func(&[vec_ty], vec_ty)),
1205 sym::simd_fexp => ("exp", bx.type_func(&[vec_ty], vec_ty)),
1206 sym::simd_flog10 => ("log10", bx.type_func(&[vec_ty], vec_ty)),
1207 sym::simd_flog2 => ("log2", bx.type_func(&[vec_ty], vec_ty)),
1208 sym::simd_flog => ("log", bx.type_func(&[vec_ty], vec_ty)),
1209 sym::simd_floor => ("floor", bx.type_func(&[vec_ty], vec_ty)),
1210 sym::simd_fma => ("fma", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)),
1211 sym::simd_fpowi => ("powi", bx.type_func(&[vec_ty, bx.type_i32()], vec_ty)),
1212 sym::simd_fpow => ("pow", bx.type_func(&[vec_ty, vec_ty], vec_ty)),
1213 sym::simd_fsin => ("sin", bx.type_func(&[vec_ty], vec_ty)),
1214 sym::simd_fsqrt => ("sqrt", bx.type_func(&[vec_ty], vec_ty)),
1215 sym::simd_round => ("round", bx.type_func(&[vec_ty], vec_ty)),
1216 sym::simd_trunc => ("trunc", bx.type_func(&[vec_ty], vec_ty)),
1217 _ => return_error!("unrecognized intrinsic `{}`", name),
1219 let llvm_name = &format!("llvm.{0}.v{1}{2}", intr_name, in_len, elem_ty_str);
1220 let f = bx.declare_cfn(llvm_name, llvm::UnnamedAddr::No, fn_ty);
1225 &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
1250 return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args);
1254 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1255 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1260 bx: &Builder<'_, '_, '_>,
1262 let p0s: String = "p0".repeat(no_pointers);
1263 match *elem_ty.kind() {
1264 ty::Int(v) => format!(
1268 // Normalize to prevent crash if v: IntTy::Isize
1269 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1271 ty::Uint(v) => format!(
1275 // Normalize to prevent crash if v: UIntTy::Usize
1276 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1278 ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1279 _ => unreachable!(),
1283 fn llvm_vector_ty<'ll>(
1284 cx: &CodegenCx<'ll, '_>,
1287 mut no_pointers: usize,
1289 // FIXME: use cx.layout_of(ty).llvm_type() ?
1290 let mut elem_ty = match *elem_ty.kind() {
1291 ty::Int(v) => cx.type_int_from_ty(v),
1292 ty::Uint(v) => cx.type_uint_from_ty(v),
1293 ty::Float(v) => cx.type_float_from_ty(v),
1294 _ => unreachable!(),
1296 while no_pointers > 0 {
1297 elem_ty = cx.type_ptr_to(elem_ty);
1300 cx.type_vector(elem_ty, vec_len)
1303 if name == sym::simd_gather {
1304 // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1305 // mask: <N x i{M}>) -> <N x T>
1306 // * N: number of elements in the input vectors
1307 // * T: type of the element to load
1308 // * M: any integer width is supported, will be truncated to i1
1310 // All types must be simd vector types
1311 require_simd!(in_ty, "first");
1312 require_simd!(arg_tys[1], "second");
1313 require_simd!(arg_tys[2], "third");
1314 require_simd!(ret_ty, "return");
1316 // Of the same length:
1317 let (out_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1318 let (out_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1321 "expected {} argument with length {} (same as input type `{}`), \
1322 found `{}` with length {}",
1331 "expected {} argument with length {} (same as input type `{}`), \
1332 found `{}` with length {}",
1340 // The return type must match the first argument type
1341 require!(ret_ty == in_ty, "expected return type `{}`, found `{}`", in_ty, ret_ty);
1343 // This counts how many pointers
1344 fn ptr_count(t: Ty<'_>) -> usize {
1346 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1352 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1354 ty::RawPtr(p) => non_ptr(p.ty),
1359 // The second argument must be a simd vector with an element type that's a pointer
1360 // to the element type of the first argument
1361 let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1362 let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1363 let (pointer_count, underlying_ty) = match element_ty1.kind() {
1364 ty::RawPtr(p) if p.ty == in_elem => (ptr_count(element_ty1), non_ptr(element_ty1)),
1368 "expected element type `{}` of second argument `{}` \
1369 to be a pointer to the element type `{}` of the first \
1370 argument `{}`, found `{}` != `*_ {}`",
1381 assert!(pointer_count > 0);
1382 assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1383 assert_eq!(underlying_ty, non_ptr(element_ty0));
1385 // The element type of the third argument must be a signed integer type of any width:
1386 let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1387 match element_ty2.kind() {
1392 "expected element type `{}` of third argument `{}` \
1393 to 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 // Type of the vector of pointers:
1412 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1413 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1415 // Type of the vector of elements:
1416 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1417 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1419 let llvm_intrinsic =
1420 format!("llvm.masked.gather.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1421 let fn_ty = bx.type_func(
1422 &[llvm_pointer_vec_ty, alignment_ty, mask_ty, llvm_elem_vec_ty],
1425 let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1430 &[args[1].immediate(), alignment, mask, args[0].immediate()],
1436 if name == sym::simd_scatter {
1437 // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1438 // mask: <N x i{M}>) -> ()
1439 // * N: number of elements in the input vectors
1440 // * T: type of the element to load
1441 // * M: any integer width is supported, will be truncated to i1
1443 // All types must be simd vector types
1444 require_simd!(in_ty, "first");
1445 require_simd!(arg_tys[1], "second");
1446 require_simd!(arg_tys[2], "third");
1448 // Of the same length:
1449 let (element_len1, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1450 let (element_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1452 in_len == element_len1,
1453 "expected {} argument with length {} (same as input type `{}`), \
1454 found `{}` with length {}",
1462 in_len == element_len2,
1463 "expected {} argument with length {} (same as input type `{}`), \
1464 found `{}` with length {}",
1472 // This counts how many pointers
1473 fn ptr_count(t: Ty<'_>) -> usize {
1475 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1481 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1483 ty::RawPtr(p) => non_ptr(p.ty),
1488 // The second argument must be a simd vector with an element type that's a pointer
1489 // to the element type of the first argument
1490 let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1491 let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1492 let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1493 let (pointer_count, underlying_ty) = match element_ty1.kind() {
1494 ty::RawPtr(p) if p.ty == in_elem && p.mutbl == hir::Mutability::Mut => {
1495 (ptr_count(element_ty1), non_ptr(element_ty1))
1500 "expected element type `{}` of second argument `{}` \
1501 to be a pointer to the element type `{}` of the first \
1502 argument `{}`, found `{}` != `*mut {}`",
1513 assert!(pointer_count > 0);
1514 assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1515 assert_eq!(underlying_ty, non_ptr(element_ty0));
1517 // The element type of the third argument must be a signed integer type of any width:
1518 match element_ty2.kind() {
1523 "expected element type `{}` of third argument `{}` \
1524 be a signed integer type",
1531 // Alignment of T, must be a constant integer value:
1532 let alignment_ty = bx.type_i32();
1533 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1535 // Truncate the mask vector to a vector of i1s:
1536 let (mask, mask_ty) = {
1537 let i1 = bx.type_i1();
1538 let i1xn = bx.type_vector(i1, in_len);
1539 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1542 let ret_t = bx.type_void();
1544 // Type of the vector of pointers:
1545 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1546 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1548 // Type of the vector of elements:
1549 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1550 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1552 let llvm_intrinsic =
1553 format!("llvm.masked.scatter.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1555 bx.type_func(&[llvm_elem_vec_ty, llvm_pointer_vec_ty, alignment_ty, mask_ty], ret_t);
1556 let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1561 &[args[0].immediate(), args[1].immediate(), alignment, mask],
1567 macro_rules! arith_red {
1568 ($name:ident : $integer_reduce:ident, $float_reduce:ident, $ordered:expr, $op:ident,
1569 $identity:expr) => {
1570 if name == sym::$name {
1573 "expected return type `{}` (element of input `{}`), found `{}`",
1578 return match in_elem.kind() {
1579 ty::Int(_) | ty::Uint(_) => {
1580 let r = bx.$integer_reduce(args[0].immediate());
1582 // if overflow occurs, the result is the
1583 // mathematical result modulo 2^n:
1584 Ok(bx.$op(args[1].immediate(), r))
1586 Ok(bx.$integer_reduce(args[0].immediate()))
1590 let acc = if $ordered {
1591 // ordered arithmetic reductions take an accumulator
1594 // unordered arithmetic reductions use the identity accumulator
1595 match f.bit_width() {
1596 32 => bx.const_real(bx.type_f32(), $identity),
1597 64 => bx.const_real(bx.type_f64(), $identity),
1600 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1609 Ok(bx.$float_reduce(acc, args[0].immediate()))
1612 "unsupported {} from `{}` with element `{}` to `{}`",
1623 arith_red!(simd_reduce_add_ordered: vector_reduce_add, vector_reduce_fadd, true, add, 0.0);
1624 arith_red!(simd_reduce_mul_ordered: vector_reduce_mul, vector_reduce_fmul, true, mul, 1.0);
1626 simd_reduce_add_unordered: vector_reduce_add,
1627 vector_reduce_fadd_fast,
1633 simd_reduce_mul_unordered: vector_reduce_mul,
1634 vector_reduce_fmul_fast,
1640 macro_rules! minmax_red {
1641 ($name:ident: $int_red:ident, $float_red:ident) => {
1642 if name == sym::$name {
1645 "expected return type `{}` (element of input `{}`), found `{}`",
1650 return match in_elem.kind() {
1651 ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
1652 ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
1653 ty::Float(_f) => Ok(bx.$float_red(args[0].immediate())),
1655 "unsupported {} from `{}` with element `{}` to `{}`",
1666 minmax_red!(simd_reduce_min: vector_reduce_min, vector_reduce_fmin);
1667 minmax_red!(simd_reduce_max: vector_reduce_max, vector_reduce_fmax);
1669 minmax_red!(simd_reduce_min_nanless: vector_reduce_min, vector_reduce_fmin_fast);
1670 minmax_red!(simd_reduce_max_nanless: vector_reduce_max, vector_reduce_fmax_fast);
1672 macro_rules! bitwise_red {
1673 ($name:ident : $red:ident, $boolean:expr) => {
1674 if name == sym::$name {
1675 let input = if !$boolean {
1678 "expected return type `{}` (element of input `{}`), found `{}`",
1685 match in_elem.kind() {
1686 ty::Int(_) | ty::Uint(_) => {}
1688 "unsupported {} from `{}` with element `{}` to `{}`",
1696 // boolean reductions operate on vectors of i1s:
1697 let i1 = bx.type_i1();
1698 let i1xn = bx.type_vector(i1, in_len as u64);
1699 bx.trunc(args[0].immediate(), i1xn)
1701 return match in_elem.kind() {
1702 ty::Int(_) | ty::Uint(_) => {
1703 let r = bx.$red(input);
1704 Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
1707 "unsupported {} from `{}` with element `{}` to `{}`",
1718 bitwise_red!(simd_reduce_and: vector_reduce_and, false);
1719 bitwise_red!(simd_reduce_or: vector_reduce_or, false);
1720 bitwise_red!(simd_reduce_xor: vector_reduce_xor, false);
1721 bitwise_red!(simd_reduce_all: vector_reduce_and, true);
1722 bitwise_red!(simd_reduce_any: vector_reduce_or, true);
1724 if name == sym::simd_cast_ptr {
1725 require_simd!(ret_ty, "return");
1726 let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1729 "expected return type with length {} (same as input type `{}`), \
1730 found `{}` with length {}",
1737 match in_elem.kind() {
1739 let (metadata, check_sized) = p.ty.ptr_metadata_ty(bx.tcx, |ty| {
1740 bx.tcx.normalize_erasing_regions(ty::ParamEnv::reveal_all(), ty)
1742 assert!(!check_sized); // we are in codegen, so we shouldn't see these types
1743 require!(metadata.is_unit(), "cannot cast fat pointer `{}`", in_elem)
1745 _ => return_error!("expected pointer, got `{}`", in_elem),
1747 match out_elem.kind() {
1749 let (metadata, check_sized) = p.ty.ptr_metadata_ty(bx.tcx, |ty| {
1750 bx.tcx.normalize_erasing_regions(ty::ParamEnv::reveal_all(), ty)
1752 assert!(!check_sized); // we are in codegen, so we shouldn't see these types
1753 require!(metadata.is_unit(), "cannot cast to fat pointer `{}`", out_elem)
1755 _ => return_error!("expected pointer, got `{}`", out_elem),
1758 if in_elem == out_elem {
1759 return Ok(args[0].immediate());
1761 return Ok(bx.pointercast(args[0].immediate(), llret_ty));
1765 if name == sym::simd_expose_addr {
1766 require_simd!(ret_ty, "return");
1767 let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1770 "expected return type with length {} (same as input type `{}`), \
1771 found `{}` with length {}",
1778 match in_elem.kind() {
1780 _ => return_error!("expected pointer, got `{}`", in_elem),
1782 match out_elem.kind() {
1783 ty::Uint(ty::UintTy::Usize) => {}
1784 _ => return_error!("expected `usize`, got `{}`", out_elem),
1787 return Ok(bx.ptrtoint(args[0].immediate(), llret_ty));
1790 if name == sym::simd_from_exposed_addr {
1791 require_simd!(ret_ty, "return");
1792 let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1795 "expected return type with length {} (same as input type `{}`), \
1796 found `{}` with length {}",
1803 match in_elem.kind() {
1804 ty::Uint(ty::UintTy::Usize) => {}
1805 _ => return_error!("expected `usize`, got `{}`", in_elem),
1807 match out_elem.kind() {
1809 _ => return_error!("expected pointer, got `{}`", out_elem),
1812 return Ok(bx.inttoptr(args[0].immediate(), llret_ty));
1815 if name == sym::simd_cast || name == sym::simd_as {
1816 require_simd!(ret_ty, "return");
1817 let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1820 "expected return type with length {} (same as input type `{}`), \
1821 found `{}` with length {}",
1827 // casting cares about nominal type, not just structural type
1828 if in_elem == out_elem {
1829 return Ok(args[0].immediate());
1834 Int(/* is signed? */ bool),
1838 let (in_style, in_width) = match in_elem.kind() {
1839 // vectors of pointer-sized integers should've been
1840 // disallowed before here, so this unwrap is safe.
1843 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1847 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1849 ty::Float(f) => (Style::Float, f.bit_width()),
1850 _ => (Style::Unsupported, 0),
1852 let (out_style, out_width) = match out_elem.kind() {
1855 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1859 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1861 ty::Float(f) => (Style::Float, f.bit_width()),
1862 _ => (Style::Unsupported, 0),
1865 match (in_style, out_style) {
1866 (Style::Int(in_is_signed), Style::Int(_)) => {
1867 return Ok(match in_width.cmp(&out_width) {
1868 Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1869 Ordering::Equal => args[0].immediate(),
1872 bx.sext(args[0].immediate(), llret_ty)
1874 bx.zext(args[0].immediate(), llret_ty)
1879 (Style::Int(in_is_signed), Style::Float) => {
1880 return Ok(if in_is_signed {
1881 bx.sitofp(args[0].immediate(), llret_ty)
1883 bx.uitofp(args[0].immediate(), llret_ty)
1886 (Style::Float, Style::Int(out_is_signed)) => {
1887 return Ok(match (out_is_signed, name == sym::simd_as) {
1888 (false, false) => bx.fptoui(args[0].immediate(), llret_ty),
1889 (true, false) => bx.fptosi(args[0].immediate(), llret_ty),
1890 (_, true) => bx.cast_float_to_int(out_is_signed, args[0].immediate(), llret_ty),
1893 (Style::Float, Style::Float) => {
1894 return Ok(match in_width.cmp(&out_width) {
1895 Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1896 Ordering::Equal => args[0].immediate(),
1897 Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
1900 _ => { /* Unsupported. Fallthrough. */ }
1904 "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
1911 macro_rules! arith_binary {
1912 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1913 $(if name == sym::$name {
1914 match in_elem.kind() {
1915 $($(ty::$p(_))|* => {
1916 return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
1921 "unsupported operation on `{}` with element `{}`",
1928 simd_add: Uint, Int => add, Float => fadd;
1929 simd_sub: Uint, Int => sub, Float => fsub;
1930 simd_mul: Uint, Int => mul, Float => fmul;
1931 simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
1932 simd_rem: Uint => urem, Int => srem, Float => frem;
1933 simd_shl: Uint, Int => shl;
1934 simd_shr: Uint => lshr, Int => ashr;
1935 simd_and: Uint, Int => and;
1936 simd_or: Uint, Int => or;
1937 simd_xor: Uint, Int => xor;
1938 simd_fmax: Float => maxnum;
1939 simd_fmin: Float => minnum;
1942 macro_rules! arith_unary {
1943 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1944 $(if name == sym::$name {
1945 match in_elem.kind() {
1946 $($(ty::$p(_))|* => {
1947 return Ok(bx.$call(args[0].immediate()))
1952 "unsupported operation on `{}` with element `{}`",
1959 simd_neg: Int => neg, Float => fneg;
1962 if name == sym::simd_arith_offset {
1963 // This also checks that the first operand is a ptr type.
1964 let pointee = in_elem.builtin_deref(true).unwrap_or_else(|| {
1965 span_bug!(span, "must be called with a vector of pointer types as first argument")
1967 let layout = bx.layout_of(pointee.ty);
1968 let ptrs = args[0].immediate();
1969 // The second argument must be a ptr-sized integer.
1970 // (We don't care about the signedness, this is wrapping anyway.)
1971 let (_offsets_len, offsets_elem) = arg_tys[1].simd_size_and_type(bx.tcx());
1972 if !matches!(offsets_elem.kind(), ty::Int(ty::IntTy::Isize) | ty::Uint(ty::UintTy::Usize)) {
1975 "must be called with a vector of pointer-sized integers as second argument"
1978 let offsets = args[1].immediate();
1980 return Ok(bx.gep(bx.backend_type(layout), ptrs, &[offsets]));
1983 if name == sym::simd_saturating_add || name == sym::simd_saturating_sub {
1984 let lhs = args[0].immediate();
1985 let rhs = args[1].immediate();
1986 let is_add = name == sym::simd_saturating_add;
1987 let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
1988 let (signed, elem_width, elem_ty) = match *in_elem.kind() {
1989 ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
1990 ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
1993 "expected element type `{}` of vector type `{}` \
1994 to be a signed or unsigned integer type",
1995 arg_tys[0].simd_size_and_type(bx.tcx()).1,
2000 let llvm_intrinsic = &format!(
2001 "llvm.{}{}.sat.v{}i{}",
2002 if signed { 's' } else { 'u' },
2003 if is_add { "add" } else { "sub" },
2007 let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
2009 let fn_ty = bx.type_func(&[vec_ty, vec_ty], vec_ty);
2010 let f = bx.declare_cfn(llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
2011 let v = bx.call(fn_ty, None, f, &[lhs, rhs], None);
2015 span_bug!(span, "unknown SIMD intrinsic");
2018 // Returns the width of an int Ty, and if it's signed or not
2019 // Returns None if the type is not an integer
2020 // FIXME: there’s multiple of this functions, investigate using some of the already existing
2022 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
2025 Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), true))
2028 Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), false))