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);
343 let result_val_span = [result.llval];
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.
350 // For zero-sized types, the location pointed to by the result may be
351 // uninitialized. Do not "use" the result in this case; instead just clobber
353 let (constraint, inputs): (&str, &[_]) = if result.layout.is_zst() {
356 ("r,~{memory}", &result_val_span)
358 crate::asm::inline_asm_call(
366 llvm::AsmDialect::Att,
371 .unwrap_or_else(|| bug!("failed to generate inline asm call for `black_box`"));
373 // We have copied the value to `result` already.
377 _ if name.as_str().starts_with("simd_") => {
378 match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
384 _ => bug!("unknown intrinsic '{}'", name),
387 if !fn_abi.ret.is_ignore() {
388 if let PassMode::Cast(ty, _) = &fn_abi.ret.mode {
389 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
390 let ptr = self.pointercast(result.llval, ptr_llty);
391 self.store(llval, ptr, result.align);
393 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
395 .store(self, result);
400 fn abort(&mut self) {
401 self.call_intrinsic("llvm.trap", &[]);
404 fn assume(&mut self, val: Self::Value) {
405 self.call_intrinsic("llvm.assume", &[val]);
408 fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
409 self.call_intrinsic("llvm.expect.i1", &[cond, self.const_bool(expected)])
412 fn type_test(&mut self, pointer: Self::Value, typeid: Self::Value) -> Self::Value {
413 // Test the called operand using llvm.type.test intrinsic. The LowerTypeTests link-time
414 // optimization pass replaces calls to this intrinsic with code to test type membership.
415 let i8p_ty = self.type_i8p();
416 let bitcast = self.bitcast(pointer, i8p_ty);
417 self.call_intrinsic("llvm.type.test", &[bitcast, typeid])
420 fn type_checked_load(
422 llvtable: &'ll Value,
423 vtable_byte_offset: u64,
426 let vtable_byte_offset = self.const_i32(vtable_byte_offset as i32);
427 let type_checked_load =
428 self.call_intrinsic("llvm.type.checked.load", &[llvtable, vtable_byte_offset, typeid]);
429 self.extract_value(type_checked_load, 0)
432 fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
433 self.call_intrinsic("llvm.va_start", &[va_list])
436 fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
437 self.call_intrinsic("llvm.va_end", &[va_list])
441 fn try_intrinsic<'ll>(
442 bx: &mut Builder<'_, 'll, '_>,
443 try_func: &'ll Value,
445 catch_func: &'ll Value,
448 if bx.sess().panic_strategy() == PanicStrategy::Abort {
449 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
450 bx.call(try_func_ty, None, try_func, &[data], None);
451 // Return 0 unconditionally from the intrinsic call;
452 // we can never unwind.
453 let ret_align = bx.tcx().data_layout.i32_align.abi;
454 bx.store(bx.const_i32(0), dest, ret_align);
455 } else if wants_msvc_seh(bx.sess()) {
456 codegen_msvc_try(bx, try_func, data, catch_func, dest);
457 } else if bx.sess().target.os == "emscripten" {
458 codegen_emcc_try(bx, try_func, data, catch_func, dest);
460 codegen_gnu_try(bx, try_func, data, catch_func, dest);
464 // MSVC's definition of the `rust_try` function.
466 // This implementation uses the new exception handling instructions in LLVM
467 // which have support in LLVM for SEH on MSVC targets. Although these
468 // instructions are meant to work for all targets, as of the time of this
469 // writing, however, LLVM does not recommend the usage of these new instructions
470 // as the old ones are still more optimized.
471 fn codegen_msvc_try<'ll>(
472 bx: &mut Builder<'_, 'll, '_>,
473 try_func: &'ll Value,
475 catch_func: &'ll Value,
478 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
479 bx.set_personality_fn(bx.eh_personality());
481 let normal = bx.append_sibling_block("normal");
482 let catchswitch = bx.append_sibling_block("catchswitch");
483 let catchpad_rust = bx.append_sibling_block("catchpad_rust");
484 let catchpad_foreign = bx.append_sibling_block("catchpad_foreign");
485 let caught = bx.append_sibling_block("caught");
487 let try_func = llvm::get_param(bx.llfn(), 0);
488 let data = llvm::get_param(bx.llfn(), 1);
489 let catch_func = llvm::get_param(bx.llfn(), 2);
491 // We're generating an IR snippet that looks like:
493 // declare i32 @rust_try(%try_func, %data, %catch_func) {
494 // %slot = alloca i8*
495 // invoke %try_func(%data) to label %normal unwind label %catchswitch
501 // %cs = catchswitch within none [%catchpad_rust, %catchpad_foreign] unwind to caller
504 // %tok = catchpad within %cs [%type_descriptor, 8, %slot]
506 // call %catch_func(%data, %ptr)
507 // catchret from %tok to label %caught
510 // %tok = catchpad within %cs [null, 64, null]
511 // call %catch_func(%data, null)
512 // catchret from %tok to label %caught
518 // This structure follows the basic usage of throw/try/catch in LLVM.
519 // For example, compile this C++ snippet to see what LLVM generates:
521 // struct rust_panic {
522 // rust_panic(const rust_panic&);
529 // void (*try_func)(void*),
531 // void (*catch_func)(void*, void*) noexcept
536 // } catch(rust_panic& a) {
537 // catch_func(data, &a);
540 // catch_func(data, NULL);
545 // More information can be found in libstd's seh.rs implementation.
546 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
547 let slot = bx.alloca(bx.type_i8p(), ptr_align);
548 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
549 bx.invoke(try_func_ty, None, try_func, &[data], normal, catchswitch, None);
551 bx.switch_to_block(normal);
552 bx.ret(bx.const_i32(0));
554 bx.switch_to_block(catchswitch);
555 let cs = bx.catch_switch(None, None, &[catchpad_rust, catchpad_foreign]);
557 // We can't use the TypeDescriptor defined in libpanic_unwind because it
558 // might be in another DLL and the SEH encoding only supports specifying
559 // a TypeDescriptor from the current module.
561 // However this isn't an issue since the MSVC runtime uses string
562 // comparison on the type name to match TypeDescriptors rather than
565 // So instead we generate a new TypeDescriptor in each module that uses
566 // `try` and let the linker merge duplicate definitions in the same
569 // When modifying, make sure that the type_name string exactly matches
570 // the one used in library/panic_unwind/src/seh.rs.
571 let type_info_vtable = bx.declare_global("??_7type_info@@6B@", bx.type_i8p());
572 let type_name = bx.const_bytes(b"rust_panic\0");
574 bx.const_struct(&[type_info_vtable, bx.const_null(bx.type_i8p()), type_name], false);
575 let tydesc = bx.declare_global("__rust_panic_type_info", bx.val_ty(type_info));
577 llvm::LLVMRustSetLinkage(tydesc, llvm::Linkage::LinkOnceODRLinkage);
578 llvm::SetUniqueComdat(bx.llmod, tydesc);
579 llvm::LLVMSetInitializer(tydesc, type_info);
582 // The flag value of 8 indicates that we are catching the exception by
583 // reference instead of by value. We can't use catch by value because
584 // that requires copying the exception object, which we don't support
585 // since our exception object effectively contains a Box.
587 // Source: MicrosoftCXXABI::getAddrOfCXXCatchHandlerType in clang
588 bx.switch_to_block(catchpad_rust);
589 let flags = bx.const_i32(8);
590 let funclet = bx.catch_pad(cs, &[tydesc, flags, slot]);
591 let ptr = bx.load(bx.type_i8p(), slot, ptr_align);
592 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
593 bx.call(catch_ty, None, catch_func, &[data, ptr], Some(&funclet));
594 bx.catch_ret(&funclet, caught);
596 // The flag value of 64 indicates a "catch-all".
597 bx.switch_to_block(catchpad_foreign);
598 let flags = bx.const_i32(64);
599 let null = bx.const_null(bx.type_i8p());
600 let funclet = bx.catch_pad(cs, &[null, flags, null]);
601 bx.call(catch_ty, None, catch_func, &[data, null], Some(&funclet));
602 bx.catch_ret(&funclet, caught);
604 bx.switch_to_block(caught);
605 bx.ret(bx.const_i32(1));
608 // Note that no invoke is used here because by definition this function
609 // can't panic (that's what it's catching).
610 let ret = bx.call(llty, None, llfn, &[try_func, data, catch_func], None);
611 let i32_align = bx.tcx().data_layout.i32_align.abi;
612 bx.store(ret, dest, i32_align);
615 // Definition of the standard `try` function for Rust using the GNU-like model
616 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
619 // This codegen is a little surprising because we always call a shim
620 // function instead of inlining the call to `invoke` manually here. This is done
621 // because in LLVM we're only allowed to have one personality per function
622 // definition. The call to the `try` intrinsic is being inlined into the
623 // function calling it, and that function may already have other personality
624 // functions in play. By calling a shim we're guaranteed that our shim will have
625 // the right personality function.
626 fn codegen_gnu_try<'ll>(
627 bx: &mut Builder<'_, 'll, '_>,
628 try_func: &'ll Value,
630 catch_func: &'ll Value,
633 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
634 // Codegens the shims described above:
637 // invoke %try_func(%data) normal %normal unwind %catch
643 // (%ptr, _) = landingpad
644 // call %catch_func(%data, %ptr)
646 let then = bx.append_sibling_block("then");
647 let catch = bx.append_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, None, try_func, &[data], then, catch, None);
655 bx.switch_to_block(then);
656 bx.ret(bx.const_i32(0));
658 // Type indicator for the exception being thrown.
660 // The first value in this tuple is a pointer to the exception object
661 // being thrown. The second value is a "selector" indicating which of
662 // the landing pad clauses the exception's type had been matched to.
663 // rust_try ignores the selector.
664 bx.switch_to_block(catch);
665 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
666 let vals = bx.landing_pad(lpad_ty, bx.eh_personality(), 1);
667 let tydesc = bx.const_null(bx.type_i8p());
668 bx.add_clause(vals, tydesc);
669 let ptr = bx.extract_value(vals, 0);
670 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
671 bx.call(catch_ty, None, catch_func, &[data, ptr], None);
672 bx.ret(bx.const_i32(1));
675 // Note that no invoke is used here because by definition this function
676 // can't panic (that's what it's catching).
677 let ret = bx.call(llty, None, llfn, &[try_func, data, catch_func], None);
678 let i32_align = bx.tcx().data_layout.i32_align.abi;
679 bx.store(ret, dest, i32_align);
682 // Variant of codegen_gnu_try used for emscripten where Rust panics are
683 // implemented using C++ exceptions. Here we use exceptions of a specific type
684 // (`struct rust_panic`) to represent Rust panics.
685 fn codegen_emcc_try<'ll>(
686 bx: &mut Builder<'_, 'll, '_>,
687 try_func: &'ll Value,
689 catch_func: &'ll Value,
692 let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
693 // Codegens the shims described above:
696 // invoke %try_func(%data) normal %normal unwind %catch
702 // (%ptr, %selector) = landingpad
703 // %rust_typeid = @llvm.eh.typeid.for(@_ZTI10rust_panic)
704 // %is_rust_panic = %selector == %rust_typeid
705 // %catch_data = alloca { i8*, i8 }
706 // %catch_data[0] = %ptr
707 // %catch_data[1] = %is_rust_panic
708 // call %catch_func(%data, %catch_data)
710 let then = bx.append_sibling_block("then");
711 let catch = bx.append_sibling_block("catch");
713 let try_func = llvm::get_param(bx.llfn(), 0);
714 let data = llvm::get_param(bx.llfn(), 1);
715 let catch_func = llvm::get_param(bx.llfn(), 2);
716 let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
717 bx.invoke(try_func_ty, None, try_func, &[data], then, catch, None);
719 bx.switch_to_block(then);
720 bx.ret(bx.const_i32(0));
722 // Type indicator for the exception being thrown.
724 // The first value in this tuple is a pointer to the exception object
725 // being thrown. The second value is a "selector" indicating which of
726 // the landing pad clauses the exception's type had been matched to.
727 bx.switch_to_block(catch);
728 let tydesc = bx.eh_catch_typeinfo();
729 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
730 let vals = bx.landing_pad(lpad_ty, bx.eh_personality(), 2);
731 bx.add_clause(vals, tydesc);
732 bx.add_clause(vals, bx.const_null(bx.type_i8p()));
733 let ptr = bx.extract_value(vals, 0);
734 let selector = bx.extract_value(vals, 1);
736 // Check if the typeid we got is the one for a Rust panic.
737 let rust_typeid = bx.call_intrinsic("llvm.eh.typeid.for", &[tydesc]);
738 let is_rust_panic = bx.icmp(IntPredicate::IntEQ, selector, rust_typeid);
739 let is_rust_panic = bx.zext(is_rust_panic, bx.type_bool());
741 // We need to pass two values to catch_func (ptr and is_rust_panic), so
742 // create an alloca and pass a pointer to that.
743 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
744 let i8_align = bx.tcx().data_layout.i8_align.abi;
745 let catch_data_type = bx.type_struct(&[bx.type_i8p(), bx.type_bool()], false);
746 let catch_data = bx.alloca(catch_data_type, ptr_align);
748 bx.inbounds_gep(catch_data_type, catch_data, &[bx.const_usize(0), bx.const_usize(0)]);
749 bx.store(ptr, catch_data_0, ptr_align);
751 bx.inbounds_gep(catch_data_type, catch_data, &[bx.const_usize(0), bx.const_usize(1)]);
752 bx.store(is_rust_panic, catch_data_1, i8_align);
753 let catch_data = bx.bitcast(catch_data, bx.type_i8p());
755 let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
756 bx.call(catch_ty, None, catch_func, &[data, catch_data], None);
757 bx.ret(bx.const_i32(1));
760 // Note that no invoke is used here because by definition this function
761 // can't panic (that's what it's catching).
762 let ret = bx.call(llty, None, llfn, &[try_func, data, catch_func], None);
763 let i32_align = bx.tcx().data_layout.i32_align.abi;
764 bx.store(ret, dest, i32_align);
767 // Helper function to give a Block to a closure to codegen a shim function.
768 // This is currently primarily used for the `try` intrinsic functions above.
769 fn gen_fn<'ll, 'tcx>(
770 cx: &CodegenCx<'ll, 'tcx>,
772 rust_fn_sig: ty::PolyFnSig<'tcx>,
773 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
774 ) -> (&'ll Type, &'ll Value) {
775 let fn_abi = cx.fn_abi_of_fn_ptr(rust_fn_sig, ty::List::empty());
776 let llty = fn_abi.llvm_type(cx);
777 let llfn = cx.declare_fn(name, fn_abi);
778 cx.set_frame_pointer_type(llfn);
779 cx.apply_target_cpu_attr(llfn);
780 // FIXME(eddyb) find a nicer way to do this.
781 unsafe { llvm::LLVMRustSetLinkage(llfn, llvm::Linkage::InternalLinkage) };
782 let llbb = Builder::append_block(cx, llfn, "entry-block");
783 let bx = Builder::build(cx, llbb);
788 // Helper function used to get a handle to the `__rust_try` function used to
791 // This function is only generated once and is then cached.
792 fn get_rust_try_fn<'ll, 'tcx>(
793 cx: &CodegenCx<'ll, 'tcx>,
794 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
795 ) -> (&'ll Type, &'ll Value) {
796 if let Some(llfn) = cx.rust_try_fn.get() {
800 // Define the type up front for the signature of the rust_try function.
802 let i8p = tcx.mk_mut_ptr(tcx.types.i8);
803 // `unsafe fn(*mut i8) -> ()`
804 let try_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
808 hir::Unsafety::Unsafe,
811 // `unsafe fn(*mut i8, *mut i8) -> ()`
812 let catch_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
813 [i8p, i8p].iter().cloned(),
816 hir::Unsafety::Unsafe,
819 // `unsafe fn(unsafe fn(*mut i8) -> (), *mut i8, unsafe fn(*mut i8, *mut i8) -> ()) -> i32`
820 let rust_fn_sig = ty::Binder::dummy(cx.tcx.mk_fn_sig(
821 [try_fn_ty, i8p, catch_fn_ty].into_iter(),
824 hir::Unsafety::Unsafe,
827 let rust_try = gen_fn(cx, "__rust_try", rust_fn_sig, codegen);
828 cx.rust_try_fn.set(Some(rust_try));
832 fn generic_simd_intrinsic<'ll, 'tcx>(
833 bx: &mut Builder<'_, 'll, 'tcx>,
836 args: &[OperandRef<'tcx, &'ll Value>],
840 ) -> Result<&'ll Value, ()> {
841 // macros for error handling:
842 #[allow(unused_macro_rules)]
843 macro_rules! emit_error {
847 ($msg: tt, $($fmt: tt)*) => {
848 span_invalid_monomorphization_error(
850 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
855 macro_rules! return_error {
858 emit_error!($($fmt)*);
864 macro_rules! require {
865 ($cond: expr, $($fmt: tt)*) => {
867 return_error!($($fmt)*);
872 macro_rules! require_simd {
873 ($ty: expr, $position: expr) => {
874 require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
880 tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), callee_ty.fn_sig(tcx));
881 let arg_tys = sig.inputs();
883 if name == sym::simd_select_bitmask {
884 require_simd!(arg_tys[1], "argument");
885 let (len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
887 let expected_int_bits = (len.max(8) - 1).next_power_of_two();
888 let expected_bytes = len / 8 + ((len % 8 > 0) as u64);
890 let mask_ty = arg_tys[0];
891 let mask = match mask_ty.kind() {
892 ty::Int(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
893 ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
895 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
896 && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
897 == Some(expected_bytes) =>
899 let place = PlaceRef::alloca(bx, args[0].layout);
900 args[0].val.store(bx, place);
901 let int_ty = bx.type_ix(expected_bytes * 8);
902 let ptr = bx.pointercast(place.llval, bx.cx.type_ptr_to(int_ty));
903 bx.load(int_ty, ptr, Align::ONE)
906 "invalid bitmask `{}`, expected `u{}` or `[u8; {}]`",
913 let i1 = bx.type_i1();
914 let im = bx.type_ix(len);
915 let i1xn = bx.type_vector(i1, len);
916 let m_im = bx.trunc(mask, im);
917 let m_i1s = bx.bitcast(m_im, i1xn);
918 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
921 // every intrinsic below takes a SIMD vector as its first argument
922 require_simd!(arg_tys[0], "input");
923 let in_ty = arg_tys[0];
925 let comparison = match name {
926 sym::simd_eq => Some(hir::BinOpKind::Eq),
927 sym::simd_ne => Some(hir::BinOpKind::Ne),
928 sym::simd_lt => Some(hir::BinOpKind::Lt),
929 sym::simd_le => Some(hir::BinOpKind::Le),
930 sym::simd_gt => Some(hir::BinOpKind::Gt),
931 sym::simd_ge => Some(hir::BinOpKind::Ge),
935 let (in_len, in_elem) = arg_tys[0].simd_size_and_type(bx.tcx());
936 if let Some(cmp_op) = comparison {
937 require_simd!(ret_ty, "return");
939 let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
942 "expected return type with length {} (same as input type `{}`), \
943 found `{}` with length {}",
950 bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
951 "expected return type with integer elements, found `{}` with non-integer `{}`",
956 return Ok(compare_simd_types(
966 if let Some(stripped) = name.as_str().strip_prefix("simd_shuffle") {
967 // If this intrinsic is the older "simd_shuffleN" form, simply parse the integer.
968 // If there is no suffix, use the index array length.
969 let n: u64 = if stripped.is_empty() {
970 // Make sure this is actually an array, since typeck only checks the length-suffixed
971 // version of this intrinsic.
972 match args[2].layout.ty.kind() {
973 ty::Array(ty, len) if matches!(ty.kind(), ty::Uint(ty::UintTy::U32)) => {
974 len.try_eval_usize(bx.cx.tcx, ty::ParamEnv::reveal_all()).unwrap_or_else(|| {
975 span_bug!(span, "could not evaluate shuffle index array length")
979 "simd_shuffle index must be an array of `u32`, got `{}`",
984 stripped.parse().unwrap_or_else(|_| {
985 span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
989 require_simd!(ret_ty, "return");
990 let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
993 "expected return type of length {}, found `{}` with length {}",
1000 "expected return element type `{}` (element of input `{}`), \
1001 found `{}` with element type `{}`",
1008 let total_len = u128::from(in_len) * 2;
1010 let vector = args[2].immediate();
1012 let indices: Option<Vec<_>> = (0..n)
1015 let val = bx.const_get_elt(vector, i as u64);
1016 match bx.const_to_opt_u128(val, true) {
1018 emit_error!("shuffle index #{} is not a constant", arg_idx);
1021 Some(idx) if idx >= total_len => {
1023 "shuffle index #{} is out of bounds (limit {})",
1029 Some(idx) => Some(bx.const_i32(idx as i32)),
1033 let Some(indices) = indices else {
1034 return Ok(bx.const_null(llret_ty));
1037 return Ok(bx.shuffle_vector(
1038 args[0].immediate(),
1039 args[1].immediate(),
1040 bx.const_vector(&indices),
1044 if name == sym::simd_insert {
1046 in_elem == arg_tys[2],
1047 "expected inserted type `{}` (element of input `{}`), found `{}`",
1052 return Ok(bx.insert_element(
1053 args[0].immediate(),
1054 args[2].immediate(),
1055 args[1].immediate(),
1058 if name == sym::simd_extract {
1061 "expected return type `{}` (element of input `{}`), found `{}`",
1066 return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()));
1069 if name == sym::simd_select {
1070 let m_elem_ty = in_elem;
1072 require_simd!(arg_tys[1], "argument");
1073 let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1076 "mismatched lengths: mask length `{}` != other vector length `{}`",
1080 match m_elem_ty.kind() {
1082 _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
1084 // truncate the mask to a vector of i1s
1085 let i1 = bx.type_i1();
1086 let i1xn = bx.type_vector(i1, m_len as u64);
1087 let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1088 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1091 if name == sym::simd_bitmask {
1092 // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1093 // vector mask and returns the most significant bit (MSB) of each lane in the form
1095 // * an unsigned integer
1096 // * an array of `u8`
1097 // If the vector has less than 8 lanes, a u8 is returned with zeroed trailing bits.
1099 // The bit order of the result depends on the byte endianness, LSB-first for little
1100 // endian and MSB-first for big endian.
1101 let expected_int_bits = in_len.max(8);
1102 let expected_bytes = expected_int_bits / 8 + ((expected_int_bits % 8 > 0) as u64);
1104 // Integer vector <i{in_bitwidth} x in_len>:
1105 let (i_xn, in_elem_bitwidth) = match in_elem.kind() {
1107 args[0].immediate(),
1108 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1111 args[0].immediate(),
1112 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1115 "vector argument `{}`'s element type `{}`, expected integer element type",
1121 // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1124 bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
1127 let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1128 // Truncate vector to an <i1 x N>
1129 let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len));
1130 // Bitcast <i1 x N> to iN:
1131 let i_ = bx.bitcast(i1xn, bx.type_ix(in_len));
1133 match ret_ty.kind() {
1134 ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => {
1135 // Zero-extend iN to the bitmask type:
1136 return Ok(bx.zext(i_, bx.type_ix(expected_int_bits)));
1138 ty::Array(elem, len)
1139 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
1140 && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
1141 == Some(expected_bytes) =>
1143 // Zero-extend iN to the array length:
1144 let ze = bx.zext(i_, bx.type_ix(expected_bytes * 8));
1146 // Convert the integer to a byte array
1147 let ptr = bx.alloca(bx.type_ix(expected_bytes * 8), Align::ONE);
1148 bx.store(ze, ptr, Align::ONE);
1149 let array_ty = bx.type_array(bx.type_i8(), expected_bytes);
1150 let ptr = bx.pointercast(ptr, bx.cx.type_ptr_to(array_ty));
1151 return Ok(bx.load(array_ty, ptr, Align::ONE));
1154 "cannot return `{}`, expected `u{}` or `[u8; {}]`",
1162 fn simd_simple_float_intrinsic<'ll, 'tcx>(
1167 bx: &mut Builder<'_, 'll, 'tcx>,
1169 args: &[OperandRef<'tcx, &'ll Value>],
1170 ) -> Result<&'ll Value, ()> {
1171 #[allow(unused_macro_rules)]
1172 macro_rules! emit_error {
1176 ($msg: tt, $($fmt: tt)*) => {
1177 span_invalid_monomorphization_error(
1179 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1183 macro_rules! return_error {
1186 emit_error!($($fmt)*);
1192 let (elem_ty_str, elem_ty) = if let ty::Float(f) = in_elem.kind() {
1193 let elem_ty = bx.cx.type_float_from_ty(*f);
1194 match f.bit_width() {
1195 32 => ("f32", elem_ty),
1196 64 => ("f64", elem_ty),
1199 "unsupported element type `{}` of floating-point vector `{}`",
1206 return_error!("`{}` is not a floating-point type", in_ty);
1209 let vec_ty = bx.type_vector(elem_ty, in_len);
1211 let (intr_name, fn_ty) = match name {
1212 sym::simd_ceil => ("ceil", bx.type_func(&[vec_ty], vec_ty)),
1213 sym::simd_fabs => ("fabs", bx.type_func(&[vec_ty], vec_ty)),
1214 sym::simd_fcos => ("cos", bx.type_func(&[vec_ty], vec_ty)),
1215 sym::simd_fexp2 => ("exp2", bx.type_func(&[vec_ty], vec_ty)),
1216 sym::simd_fexp => ("exp", bx.type_func(&[vec_ty], vec_ty)),
1217 sym::simd_flog10 => ("log10", bx.type_func(&[vec_ty], vec_ty)),
1218 sym::simd_flog2 => ("log2", bx.type_func(&[vec_ty], vec_ty)),
1219 sym::simd_flog => ("log", bx.type_func(&[vec_ty], vec_ty)),
1220 sym::simd_floor => ("floor", bx.type_func(&[vec_ty], vec_ty)),
1221 sym::simd_fma => ("fma", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)),
1222 sym::simd_fpowi => ("powi", bx.type_func(&[vec_ty, bx.type_i32()], vec_ty)),
1223 sym::simd_fpow => ("pow", bx.type_func(&[vec_ty, vec_ty], vec_ty)),
1224 sym::simd_fsin => ("sin", bx.type_func(&[vec_ty], vec_ty)),
1225 sym::simd_fsqrt => ("sqrt", bx.type_func(&[vec_ty], vec_ty)),
1226 sym::simd_round => ("round", bx.type_func(&[vec_ty], vec_ty)),
1227 sym::simd_trunc => ("trunc", bx.type_func(&[vec_ty], vec_ty)),
1228 _ => return_error!("unrecognized intrinsic `{}`", name),
1230 let llvm_name = &format!("llvm.{0}.v{1}{2}", intr_name, in_len, elem_ty_str);
1231 let f = bx.declare_cfn(llvm_name, llvm::UnnamedAddr::No, fn_ty);
1236 &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
1261 return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args);
1265 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1266 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1271 bx: &Builder<'_, '_, '_>,
1273 let p0s: String = "p0".repeat(no_pointers);
1274 match *elem_ty.kind() {
1275 ty::Int(v) => format!(
1279 // Normalize to prevent crash if v: IntTy::Isize
1280 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1282 ty::Uint(v) => format!(
1286 // Normalize to prevent crash if v: UIntTy::Usize
1287 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1289 ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1290 _ => unreachable!(),
1294 fn llvm_vector_ty<'ll>(
1295 cx: &CodegenCx<'ll, '_>,
1298 mut no_pointers: usize,
1300 // FIXME: use cx.layout_of(ty).llvm_type() ?
1301 let mut elem_ty = match *elem_ty.kind() {
1302 ty::Int(v) => cx.type_int_from_ty(v),
1303 ty::Uint(v) => cx.type_uint_from_ty(v),
1304 ty::Float(v) => cx.type_float_from_ty(v),
1305 _ => unreachable!(),
1307 while no_pointers > 0 {
1308 elem_ty = cx.type_ptr_to(elem_ty);
1311 cx.type_vector(elem_ty, vec_len)
1314 if name == sym::simd_gather {
1315 // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1316 // mask: <N x i{M}>) -> <N x T>
1317 // * N: number of elements in the input vectors
1318 // * T: type of the element to load
1319 // * M: any integer width is supported, will be truncated to i1
1321 // All types must be simd vector types
1322 require_simd!(in_ty, "first");
1323 require_simd!(arg_tys[1], "second");
1324 require_simd!(arg_tys[2], "third");
1325 require_simd!(ret_ty, "return");
1327 // Of the same length:
1328 let (out_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1329 let (out_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1332 "expected {} argument with length {} (same as input type `{}`), \
1333 found `{}` with length {}",
1342 "expected {} argument with length {} (same as input type `{}`), \
1343 found `{}` with length {}",
1351 // The return type must match the first argument type
1352 require!(ret_ty == in_ty, "expected return type `{}`, found `{}`", in_ty, ret_ty);
1354 // This counts how many pointers
1355 fn ptr_count(t: Ty<'_>) -> usize {
1357 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1363 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1365 ty::RawPtr(p) => non_ptr(p.ty),
1370 // The second argument must be a simd vector with an element type that's a pointer
1371 // to the element type of the first argument
1372 let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1373 let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1374 let (pointer_count, underlying_ty) = match element_ty1.kind() {
1375 ty::RawPtr(p) if p.ty == in_elem => (ptr_count(element_ty1), non_ptr(element_ty1)),
1379 "expected element type `{}` of second argument `{}` \
1380 to be a pointer to the element type `{}` of the first \
1381 argument `{}`, found `{}` != `*_ {}`",
1392 assert!(pointer_count > 0);
1393 assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1394 assert_eq!(underlying_ty, non_ptr(element_ty0));
1396 // The element type of the third argument must be a signed integer type of any width:
1397 let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1398 match element_ty2.kind() {
1403 "expected element type `{}` of third argument `{}` \
1404 to be a signed integer type",
1411 // Alignment of T, must be a constant integer value:
1412 let alignment_ty = bx.type_i32();
1413 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1415 // Truncate the mask vector to a vector of i1s:
1416 let (mask, mask_ty) = {
1417 let i1 = bx.type_i1();
1418 let i1xn = bx.type_vector(i1, in_len);
1419 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1422 // Type of the vector of pointers:
1423 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1424 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1426 // Type of the vector of elements:
1427 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1428 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1430 let llvm_intrinsic =
1431 format!("llvm.masked.gather.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1432 let fn_ty = bx.type_func(
1433 &[llvm_pointer_vec_ty, alignment_ty, mask_ty, llvm_elem_vec_ty],
1436 let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1441 &[args[1].immediate(), alignment, mask, args[0].immediate()],
1447 if name == sym::simd_scatter {
1448 // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1449 // mask: <N x i{M}>) -> ()
1450 // * N: number of elements in the input vectors
1451 // * T: type of the element to load
1452 // * M: any integer width is supported, will be truncated to i1
1454 // All types must be simd vector types
1455 require_simd!(in_ty, "first");
1456 require_simd!(arg_tys[1], "second");
1457 require_simd!(arg_tys[2], "third");
1459 // Of the same length:
1460 let (element_len1, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1461 let (element_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1463 in_len == element_len1,
1464 "expected {} argument with length {} (same as input type `{}`), \
1465 found `{}` with length {}",
1473 in_len == element_len2,
1474 "expected {} argument with length {} (same as input type `{}`), \
1475 found `{}` with length {}",
1483 // This counts how many pointers
1484 fn ptr_count(t: Ty<'_>) -> usize {
1486 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1492 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1494 ty::RawPtr(p) => non_ptr(p.ty),
1499 // The second argument must be a simd vector with an element type that's a pointer
1500 // to the element type of the first argument
1501 let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1502 let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1503 let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1504 let (pointer_count, underlying_ty) = match element_ty1.kind() {
1505 ty::RawPtr(p) if p.ty == in_elem && p.mutbl.is_mut() => {
1506 (ptr_count(element_ty1), non_ptr(element_ty1))
1511 "expected element type `{}` of second argument `{}` \
1512 to be a pointer to the element type `{}` of the first \
1513 argument `{}`, found `{}` != `*mut {}`",
1524 assert!(pointer_count > 0);
1525 assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1526 assert_eq!(underlying_ty, non_ptr(element_ty0));
1528 // The element type of the third argument must be a signed integer type of any width:
1529 match element_ty2.kind() {
1534 "expected element type `{}` of third argument `{}` \
1535 be a signed integer type",
1542 // Alignment of T, must be a constant integer value:
1543 let alignment_ty = bx.type_i32();
1544 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1546 // Truncate the mask vector to a vector of i1s:
1547 let (mask, mask_ty) = {
1548 let i1 = bx.type_i1();
1549 let i1xn = bx.type_vector(i1, in_len);
1550 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1553 let ret_t = bx.type_void();
1555 // Type of the vector of pointers:
1556 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1557 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1559 // Type of the vector of elements:
1560 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1561 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1563 let llvm_intrinsic =
1564 format!("llvm.masked.scatter.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1566 bx.type_func(&[llvm_elem_vec_ty, llvm_pointer_vec_ty, alignment_ty, mask_ty], ret_t);
1567 let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1572 &[args[0].immediate(), args[1].immediate(), alignment, mask],
1578 macro_rules! arith_red {
1579 ($name:ident : $integer_reduce:ident, $float_reduce:ident, $ordered:expr, $op:ident,
1580 $identity:expr) => {
1581 if name == sym::$name {
1584 "expected return type `{}` (element of input `{}`), found `{}`",
1589 return match in_elem.kind() {
1590 ty::Int(_) | ty::Uint(_) => {
1591 let r = bx.$integer_reduce(args[0].immediate());
1593 // if overflow occurs, the result is the
1594 // mathematical result modulo 2^n:
1595 Ok(bx.$op(args[1].immediate(), r))
1597 Ok(bx.$integer_reduce(args[0].immediate()))
1601 let acc = if $ordered {
1602 // ordered arithmetic reductions take an accumulator
1605 // unordered arithmetic reductions use the identity accumulator
1606 match f.bit_width() {
1607 32 => bx.const_real(bx.type_f32(), $identity),
1608 64 => bx.const_real(bx.type_f64(), $identity),
1611 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1620 Ok(bx.$float_reduce(acc, args[0].immediate()))
1623 "unsupported {} from `{}` with element `{}` to `{}`",
1634 arith_red!(simd_reduce_add_ordered: vector_reduce_add, vector_reduce_fadd, true, add, 0.0);
1635 arith_red!(simd_reduce_mul_ordered: vector_reduce_mul, vector_reduce_fmul, true, mul, 1.0);
1637 simd_reduce_add_unordered: vector_reduce_add,
1638 vector_reduce_fadd_fast,
1644 simd_reduce_mul_unordered: vector_reduce_mul,
1645 vector_reduce_fmul_fast,
1651 macro_rules! minmax_red {
1652 ($name:ident: $int_red:ident, $float_red:ident) => {
1653 if name == sym::$name {
1656 "expected return type `{}` (element of input `{}`), found `{}`",
1661 return match in_elem.kind() {
1662 ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
1663 ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
1664 ty::Float(_f) => Ok(bx.$float_red(args[0].immediate())),
1666 "unsupported {} from `{}` with element `{}` to `{}`",
1677 minmax_red!(simd_reduce_min: vector_reduce_min, vector_reduce_fmin);
1678 minmax_red!(simd_reduce_max: vector_reduce_max, vector_reduce_fmax);
1680 minmax_red!(simd_reduce_min_nanless: vector_reduce_min, vector_reduce_fmin_fast);
1681 minmax_red!(simd_reduce_max_nanless: vector_reduce_max, vector_reduce_fmax_fast);
1683 macro_rules! bitwise_red {
1684 ($name:ident : $red:ident, $boolean:expr) => {
1685 if name == sym::$name {
1686 let input = if !$boolean {
1689 "expected return type `{}` (element of input `{}`), found `{}`",
1696 match in_elem.kind() {
1697 ty::Int(_) | ty::Uint(_) => {}
1699 "unsupported {} from `{}` with element `{}` to `{}`",
1707 // boolean reductions operate on vectors of i1s:
1708 let i1 = bx.type_i1();
1709 let i1xn = bx.type_vector(i1, in_len as u64);
1710 bx.trunc(args[0].immediate(), i1xn)
1712 return match in_elem.kind() {
1713 ty::Int(_) | ty::Uint(_) => {
1714 let r = bx.$red(input);
1715 Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
1718 "unsupported {} from `{}` with element `{}` to `{}`",
1729 bitwise_red!(simd_reduce_and: vector_reduce_and, false);
1730 bitwise_red!(simd_reduce_or: vector_reduce_or, false);
1731 bitwise_red!(simd_reduce_xor: vector_reduce_xor, false);
1732 bitwise_red!(simd_reduce_all: vector_reduce_and, true);
1733 bitwise_red!(simd_reduce_any: vector_reduce_or, true);
1735 if name == sym::simd_cast_ptr {
1736 require_simd!(ret_ty, "return");
1737 let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1740 "expected return type with length {} (same as input type `{}`), \
1741 found `{}` with length {}",
1748 match in_elem.kind() {
1750 let (metadata, check_sized) = p.ty.ptr_metadata_ty(bx.tcx, |ty| {
1751 bx.tcx.normalize_erasing_regions(ty::ParamEnv::reveal_all(), ty)
1753 assert!(!check_sized); // we are in codegen, so we shouldn't see these types
1754 require!(metadata.is_unit(), "cannot cast fat pointer `{}`", in_elem)
1756 _ => return_error!("expected pointer, got `{}`", in_elem),
1758 match out_elem.kind() {
1760 let (metadata, check_sized) = p.ty.ptr_metadata_ty(bx.tcx, |ty| {
1761 bx.tcx.normalize_erasing_regions(ty::ParamEnv::reveal_all(), ty)
1763 assert!(!check_sized); // we are in codegen, so we shouldn't see these types
1764 require!(metadata.is_unit(), "cannot cast to fat pointer `{}`", out_elem)
1766 _ => return_error!("expected pointer, got `{}`", out_elem),
1769 if in_elem == out_elem {
1770 return Ok(args[0].immediate());
1772 return Ok(bx.pointercast(args[0].immediate(), llret_ty));
1776 if name == sym::simd_expose_addr {
1777 require_simd!(ret_ty, "return");
1778 let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1781 "expected return type with length {} (same as input type `{}`), \
1782 found `{}` with length {}",
1789 match in_elem.kind() {
1791 _ => return_error!("expected pointer, got `{}`", in_elem),
1793 match out_elem.kind() {
1794 ty::Uint(ty::UintTy::Usize) => {}
1795 _ => return_error!("expected `usize`, got `{}`", out_elem),
1798 return Ok(bx.ptrtoint(args[0].immediate(), llret_ty));
1801 if name == sym::simd_from_exposed_addr {
1802 require_simd!(ret_ty, "return");
1803 let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1806 "expected return type with length {} (same as input type `{}`), \
1807 found `{}` with length {}",
1814 match in_elem.kind() {
1815 ty::Uint(ty::UintTy::Usize) => {}
1816 _ => return_error!("expected `usize`, got `{}`", in_elem),
1818 match out_elem.kind() {
1820 _ => return_error!("expected pointer, got `{}`", out_elem),
1823 return Ok(bx.inttoptr(args[0].immediate(), llret_ty));
1826 if name == sym::simd_cast || name == sym::simd_as {
1827 require_simd!(ret_ty, "return");
1828 let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1831 "expected return type with length {} (same as input type `{}`), \
1832 found `{}` with length {}",
1838 // casting cares about nominal type, not just structural type
1839 if in_elem == out_elem {
1840 return Ok(args[0].immediate());
1845 Int(/* is signed? */ bool),
1849 let (in_style, in_width) = match in_elem.kind() {
1850 // vectors of pointer-sized integers should've been
1851 // disallowed before here, so this unwrap is safe.
1854 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1858 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1860 ty::Float(f) => (Style::Float, f.bit_width()),
1861 _ => (Style::Unsupported, 0),
1863 let (out_style, out_width) = match out_elem.kind() {
1866 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1870 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1872 ty::Float(f) => (Style::Float, f.bit_width()),
1873 _ => (Style::Unsupported, 0),
1876 match (in_style, out_style) {
1877 (Style::Int(in_is_signed), Style::Int(_)) => {
1878 return Ok(match in_width.cmp(&out_width) {
1879 Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1880 Ordering::Equal => args[0].immediate(),
1883 bx.sext(args[0].immediate(), llret_ty)
1885 bx.zext(args[0].immediate(), llret_ty)
1890 (Style::Int(in_is_signed), Style::Float) => {
1891 return Ok(if in_is_signed {
1892 bx.sitofp(args[0].immediate(), llret_ty)
1894 bx.uitofp(args[0].immediate(), llret_ty)
1897 (Style::Float, Style::Int(out_is_signed)) => {
1898 return Ok(match (out_is_signed, name == sym::simd_as) {
1899 (false, false) => bx.fptoui(args[0].immediate(), llret_ty),
1900 (true, false) => bx.fptosi(args[0].immediate(), llret_ty),
1901 (_, true) => bx.cast_float_to_int(out_is_signed, args[0].immediate(), llret_ty),
1904 (Style::Float, Style::Float) => {
1905 return Ok(match in_width.cmp(&out_width) {
1906 Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1907 Ordering::Equal => args[0].immediate(),
1908 Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
1911 _ => { /* Unsupported. Fallthrough. */ }
1915 "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
1922 macro_rules! arith_binary {
1923 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1924 $(if name == sym::$name {
1925 match in_elem.kind() {
1926 $($(ty::$p(_))|* => {
1927 return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
1932 "unsupported operation on `{}` with element `{}`",
1939 simd_add: Uint, Int => add, Float => fadd;
1940 simd_sub: Uint, Int => sub, Float => fsub;
1941 simd_mul: Uint, Int => mul, Float => fmul;
1942 simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
1943 simd_rem: Uint => urem, Int => srem, Float => frem;
1944 simd_shl: Uint, Int => shl;
1945 simd_shr: Uint => lshr, Int => ashr;
1946 simd_and: Uint, Int => and;
1947 simd_or: Uint, Int => or;
1948 simd_xor: Uint, Int => xor;
1949 simd_fmax: Float => maxnum;
1950 simd_fmin: Float => minnum;
1953 macro_rules! arith_unary {
1954 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1955 $(if name == sym::$name {
1956 match in_elem.kind() {
1957 $($(ty::$p(_))|* => {
1958 return Ok(bx.$call(args[0].immediate()))
1963 "unsupported operation on `{}` with element `{}`",
1970 simd_neg: Int => neg, Float => fneg;
1973 if name == sym::simd_arith_offset {
1974 // This also checks that the first operand is a ptr type.
1975 let pointee = in_elem.builtin_deref(true).unwrap_or_else(|| {
1976 span_bug!(span, "must be called with a vector of pointer types as first argument")
1978 let layout = bx.layout_of(pointee.ty);
1979 let ptrs = args[0].immediate();
1980 // The second argument must be a ptr-sized integer.
1981 // (We don't care about the signedness, this is wrapping anyway.)
1982 let (_offsets_len, offsets_elem) = arg_tys[1].simd_size_and_type(bx.tcx());
1983 if !matches!(offsets_elem.kind(), ty::Int(ty::IntTy::Isize) | ty::Uint(ty::UintTy::Usize)) {
1986 "must be called with a vector of pointer-sized integers as second argument"
1989 let offsets = args[1].immediate();
1991 return Ok(bx.gep(bx.backend_type(layout), ptrs, &[offsets]));
1994 if name == sym::simd_saturating_add || name == sym::simd_saturating_sub {
1995 let lhs = args[0].immediate();
1996 let rhs = args[1].immediate();
1997 let is_add = name == sym::simd_saturating_add;
1998 let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
1999 let (signed, elem_width, elem_ty) = match *in_elem.kind() {
2000 ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
2001 ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
2004 "expected element type `{}` of vector type `{}` \
2005 to be a signed or unsigned integer type",
2006 arg_tys[0].simd_size_and_type(bx.tcx()).1,
2011 let llvm_intrinsic = &format!(
2012 "llvm.{}{}.sat.v{}i{}",
2013 if signed { 's' } else { 'u' },
2014 if is_add { "add" } else { "sub" },
2018 let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
2020 let fn_ty = bx.type_func(&[vec_ty, vec_ty], vec_ty);
2021 let f = bx.declare_cfn(llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
2022 let v = bx.call(fn_ty, None, f, &[lhs, rhs], None);
2026 span_bug!(span, "unknown SIMD intrinsic");
2029 // Returns the width of an int Ty, and if it's signed or not
2030 // Returns None if the type is not an integer
2031 // FIXME: there’s multiple of this functions, investigate using some of the already existing
2033 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
2036 Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), true))
2039 Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), false))