1 use crate::builder::Builder;
2 use crate::context::CodegenCx;
4 use crate::type_::Type;
5 use crate::type_of::LayoutLlvmExt;
6 use crate::value::Value;
8 use rustc_ast::LlvmAsmDialect;
9 use rustc_ast::{InlineAsmOptions, InlineAsmTemplatePiece};
10 use rustc_codegen_ssa::mir::operand::OperandValue;
11 use rustc_codegen_ssa::mir::place::PlaceRef;
12 use rustc_codegen_ssa::traits::*;
13 use rustc_data_structures::fx::FxHashMap;
15 use rustc_middle::ty::layout::TyAndLayout;
16 use rustc_middle::{bug, span_bug};
17 use rustc_span::{Pos, Span};
18 use rustc_target::abi::*;
19 use rustc_target::asm::*;
21 use libc::{c_char, c_uint};
24 impl AsmBuilderMethods<'tcx> for Builder<'a, 'll, 'tcx> {
25 fn codegen_llvm_inline_asm(
27 ia: &hir::LlvmInlineAsmInner,
28 outputs: Vec<PlaceRef<'tcx, &'ll Value>>,
29 mut inputs: Vec<&'ll Value>,
32 let mut ext_constraints = vec![];
33 let mut output_types = vec![];
35 // Prepare the output operands
36 let mut indirect_outputs = vec![];
37 for (i, (out, &place)) in ia.outputs.iter().zip(&outputs).enumerate() {
39 let operand = self.load_operand(place);
40 if let OperandValue::Immediate(_) = operand.val {
41 inputs.push(operand.immediate());
43 ext_constraints.push(i.to_string());
46 let operand = self.load_operand(place);
47 if let OperandValue::Immediate(_) = operand.val {
48 indirect_outputs.push(operand.immediate());
51 output_types.push(place.layout.llvm_type(self.cx));
54 if !indirect_outputs.is_empty() {
55 indirect_outputs.extend_from_slice(&inputs);
56 inputs = indirect_outputs;
59 let clobbers = ia.clobbers.iter().map(|s| format!("~{{{}}}", &s));
61 // Default per-arch clobbers
62 // Basically what clang does
63 let arch_clobbers = match &self.sess().target.arch[..] {
64 "x86" | "x86_64" => vec!["~{dirflag}", "~{fpsr}", "~{flags}"],
65 "mips" | "mips64" => vec!["~{$1}"],
69 let all_constraints = ia
72 .map(|out| out.constraint.to_string())
73 .chain(ia.inputs.iter().map(|s| s.to_string()))
74 .chain(ext_constraints)
76 .chain(arch_clobbers.iter().map(|s| (*s).to_string()))
77 .collect::<Vec<String>>()
80 debug!("Asm Constraints: {}", &all_constraints);
82 // Depending on how many outputs we have, the return type is different
83 let num_outputs = output_types.len();
84 let output_type = match num_outputs {
85 0 => self.type_void(),
87 _ => self.type_struct(&output_types, false),
90 let asm = ia.asm.as_str();
91 let r = inline_asm_call(
107 // Again, based on how many outputs we have
108 let outputs = ia.outputs.iter().zip(&outputs).filter(|&(ref o, _)| !o.is_indirect);
109 for (i, (_, &place)) in outputs.enumerate() {
110 let v = if num_outputs == 1 { r } else { self.extract_value(r, i as u64) };
111 OperandValue::Immediate(v).store(self, place);
117 fn codegen_inline_asm(
119 template: &[InlineAsmTemplatePiece],
120 operands: &[InlineAsmOperandRef<'tcx, Self>],
121 options: InlineAsmOptions,
124 let asm_arch = self.tcx.sess.asm_arch.unwrap();
126 // Collect the types of output operands
127 let mut constraints = vec![];
128 let mut output_types = vec![];
129 let mut op_idx = FxHashMap::default();
130 for (idx, op) in operands.iter().enumerate() {
132 InlineAsmOperandRef::Out { reg, late, place } => {
133 let mut layout = None;
134 let ty = if let Some(ref place) = place {
135 layout = Some(&place.layout);
136 llvm_fixup_output_type(self.cx, reg.reg_class(), &place.layout)
138 // If the output is discarded, we don't really care what
139 // type is used. We're just using this to tell LLVM to
140 // reserve the register.
141 dummy_output_type(self.cx, reg.reg_class())
143 output_types.push(ty);
144 op_idx.insert(idx, constraints.len());
145 let prefix = if late { "=" } else { "=&" };
146 constraints.push(format!("{}{}", prefix, reg_to_llvm(reg, layout)));
148 InlineAsmOperandRef::InOut { reg, late, in_value, out_place } => {
149 let layout = if let Some(ref out_place) = out_place {
152 // LLVM required tied operands to have the same type,
153 // so we just use the type of the input.
156 let ty = llvm_fixup_output_type(self.cx, reg.reg_class(), layout);
157 output_types.push(ty);
158 op_idx.insert(idx, constraints.len());
159 let prefix = if late { "=" } else { "=&" };
160 constraints.push(format!("{}{}", prefix, reg_to_llvm(reg, Some(layout))));
166 // Collect input operands
167 let mut inputs = vec![];
168 for (idx, op) in operands.iter().enumerate() {
170 InlineAsmOperandRef::In { reg, value } => {
172 llvm_fixup_input(self, value.immediate(), reg.reg_class(), &value.layout);
174 op_idx.insert(idx, constraints.len());
175 constraints.push(reg_to_llvm(reg, Some(&value.layout)));
177 InlineAsmOperandRef::InOut { reg, late: _, in_value, out_place: _ } => {
178 let value = llvm_fixup_input(
180 in_value.immediate(),
185 constraints.push(format!("{}", op_idx[&idx]));
187 InlineAsmOperandRef::SymFn { instance } => {
188 inputs.push(self.cx.get_fn(instance));
189 op_idx.insert(idx, constraints.len());
190 constraints.push("s".to_string());
192 InlineAsmOperandRef::SymStatic { def_id } => {
193 inputs.push(self.cx.get_static(def_id));
194 op_idx.insert(idx, constraints.len());
195 constraints.push("s".to_string());
201 // Build the template string
202 let mut template_str = String::new();
203 for piece in template {
205 InlineAsmTemplatePiece::String(ref s) => {
209 template_str.push_str("$$");
211 template_str.push(c);
215 template_str.push_str(s)
218 InlineAsmTemplatePiece::Placeholder { operand_idx, modifier, span: _ } => {
219 match operands[operand_idx] {
220 InlineAsmOperandRef::In { reg, .. }
221 | InlineAsmOperandRef::Out { reg, .. }
222 | InlineAsmOperandRef::InOut { reg, .. } => {
223 let modifier = modifier_to_llvm(asm_arch, reg.reg_class(), modifier);
224 if let Some(modifier) = modifier {
225 template_str.push_str(&format!(
227 op_idx[&operand_idx], modifier
230 template_str.push_str(&format!("${{{}}}", op_idx[&operand_idx]));
233 InlineAsmOperandRef::Const { ref string } => {
234 // Const operands get injected directly into the template
235 template_str.push_str(string);
237 InlineAsmOperandRef::SymFn { .. }
238 | InlineAsmOperandRef::SymStatic { .. } => {
239 // Only emit the raw symbol name
240 template_str.push_str(&format!("${{{}:c}}", op_idx[&operand_idx]));
247 if !options.contains(InlineAsmOptions::PRESERVES_FLAGS) {
249 InlineAsmArch::AArch64 | InlineAsmArch::Arm => {
250 constraints.push("~{cc}".to_string());
252 InlineAsmArch::X86 | InlineAsmArch::X86_64 => {
253 constraints.extend_from_slice(&[
254 "~{dirflag}".to_string(),
255 "~{fpsr}".to_string(),
256 "~{flags}".to_string(),
259 InlineAsmArch::RiscV32 | InlineAsmArch::RiscV64 => {}
260 InlineAsmArch::Nvptx64 => {}
261 InlineAsmArch::Hexagon => {}
262 InlineAsmArch::Mips | InlineAsmArch::Mips64 => {}
263 InlineAsmArch::SpirV => {}
264 InlineAsmArch::Wasm32 => {}
267 if !options.contains(InlineAsmOptions::NOMEM) {
268 // This is actually ignored by LLVM, but it's probably best to keep
269 // it just in case. LLVM instead uses the ReadOnly/ReadNone
270 // attributes on the call instruction to optimize.
271 constraints.push("~{memory}".to_string());
273 let volatile = !options.contains(InlineAsmOptions::PURE);
274 let alignstack = !options.contains(InlineAsmOptions::NOSTACK);
275 let output_type = match &output_types[..] {
276 [] => self.type_void(),
278 tys => self.type_struct(&tys, false),
280 let dialect = match asm_arch {
281 InlineAsmArch::X86 | InlineAsmArch::X86_64
282 if !options.contains(InlineAsmOptions::ATT_SYNTAX) =>
284 LlvmAsmDialect::Intel
286 _ => LlvmAsmDialect::Att,
288 let result = inline_asm_call(
291 &constraints.join(","),
299 .unwrap_or_else(|| span_bug!(line_spans[0], "LLVM asm constraint validation failed"));
301 if options.contains(InlineAsmOptions::PURE) {
302 if options.contains(InlineAsmOptions::NOMEM) {
303 llvm::Attribute::ReadNone.apply_callsite(llvm::AttributePlace::Function, result);
304 } else if options.contains(InlineAsmOptions::READONLY) {
305 llvm::Attribute::ReadOnly.apply_callsite(llvm::AttributePlace::Function, result);
307 } else if options.contains(InlineAsmOptions::NOMEM) {
308 llvm::Attribute::InaccessibleMemOnly
309 .apply_callsite(llvm::AttributePlace::Function, result);
311 // LLVM doesn't have an attribute to represent ReadOnly + SideEffect
314 // Write results to outputs
315 for (idx, op) in operands.iter().enumerate() {
316 if let InlineAsmOperandRef::Out { reg, place: Some(place), .. }
317 | InlineAsmOperandRef::InOut { reg, out_place: Some(place), .. } = *op
319 let value = if output_types.len() == 1 {
322 self.extract_value(result, op_idx[&idx] as u64)
324 let value = llvm_fixup_output(self, value, reg.reg_class(), &place.layout);
325 OperandValue::Immediate(value).store(self, place);
331 impl AsmMethods for CodegenCx<'ll, 'tcx> {
332 fn codegen_global_asm(&self, ga: &hir::GlobalAsm) {
333 let asm = ga.asm.as_str();
335 llvm::LLVMRustAppendModuleInlineAsm(self.llmod, asm.as_ptr().cast(), asm.len());
341 bx: &mut Builder<'a, 'll, 'tcx>,
344 inputs: &[&'ll Value],
345 output: &'ll llvm::Type,
350 ) -> Option<&'ll Value> {
351 let volatile = if volatile { llvm::True } else { llvm::False };
352 let alignstack = if alignstack { llvm::True } else { llvm::False };
357 debug!("Asm Input Type: {:?}", *v);
360 .collect::<Vec<_>>();
362 debug!("Asm Output Type: {:?}", output);
363 let fty = bx.cx.type_func(&argtys[..], output);
365 // Ask LLVM to verify that the constraints are well-formed.
366 let constraints_ok = llvm::LLVMRustInlineAsmVerify(fty, cons.as_ptr().cast(), cons.len());
367 debug!("constraint verification result: {:?}", constraints_ok);
369 let v = llvm::LLVMRustInlineAsm(
373 cons.as_ptr().cast(),
377 llvm::AsmDialect::from_generic(dia),
379 let call = bx.call(v, inputs, None);
381 // Store mark in a metadata node so we can map LLVM errors
382 // back to source locations. See #17552.
384 let kind = llvm::LLVMGetMDKindIDInContext(
386 key.as_ptr() as *const c_char,
390 // srcloc contains one integer for each line of assembly code.
391 // Unfortunately this isn't enough to encode a full span so instead
392 // we just encode the start position of each line.
393 // FIXME: Figure out a way to pass the entire line spans.
394 let mut srcloc = vec![];
395 if dia == LlvmAsmDialect::Intel && line_spans.len() > 1 {
396 // LLVM inserts an extra line to add the ".intel_syntax", so add
397 // a dummy srcloc entry for it.
399 // Don't do this if we only have 1 line span since that may be
400 // due to the asm template string coming from a macro. LLVM will
401 // default to the first srcloc for lines that don't have an
402 // associated srcloc.
403 srcloc.push(bx.const_i32(0));
405 srcloc.extend(line_spans.iter().map(|span| bx.const_i32(span.lo().to_u32() as i32)));
406 let md = llvm::LLVMMDNodeInContext(bx.llcx, srcloc.as_ptr(), srcloc.len() as u32);
407 llvm::LLVMSetMetadata(call, kind, md);
411 // LLVM has detected an issue with our constraints, bail out
417 /// If the register is an xmm/ymm/zmm register then return its index.
418 fn xmm_reg_index(reg: InlineAsmReg) -> Option<u32> {
420 InlineAsmReg::X86(reg)
421 if reg as u32 >= X86InlineAsmReg::xmm0 as u32
422 && reg as u32 <= X86InlineAsmReg::xmm15 as u32 =>
424 Some(reg as u32 - X86InlineAsmReg::xmm0 as u32)
426 InlineAsmReg::X86(reg)
427 if reg as u32 >= X86InlineAsmReg::ymm0 as u32
428 && reg as u32 <= X86InlineAsmReg::ymm15 as u32 =>
430 Some(reg as u32 - X86InlineAsmReg::ymm0 as u32)
432 InlineAsmReg::X86(reg)
433 if reg as u32 >= X86InlineAsmReg::zmm0 as u32
434 && reg as u32 <= X86InlineAsmReg::zmm31 as u32 =>
436 Some(reg as u32 - X86InlineAsmReg::zmm0 as u32)
442 /// If the register is an AArch64 vector register then return its index.
443 fn a64_vreg_index(reg: InlineAsmReg) -> Option<u32> {
445 InlineAsmReg::AArch64(reg)
446 if reg as u32 >= AArch64InlineAsmReg::v0 as u32
447 && reg as u32 <= AArch64InlineAsmReg::v31 as u32 =>
449 Some(reg as u32 - AArch64InlineAsmReg::v0 as u32)
455 /// Converts a register class to an LLVM constraint code.
456 fn reg_to_llvm(reg: InlineAsmRegOrRegClass, layout: Option<&TyAndLayout<'tcx>>) -> String {
458 // For vector registers LLVM wants the register name to match the type size.
459 InlineAsmRegOrRegClass::Reg(reg) => {
460 if let Some(idx) = xmm_reg_index(reg) {
461 let class = if let Some(layout) = layout {
462 match layout.size.bytes() {
468 // We use f32 as the type for discarded outputs
471 format!("{{{}mm{}}}", class, idx)
472 } else if let Some(idx) = a64_vreg_index(reg) {
473 let class = if let Some(layout) = layout {
474 match layout.size.bytes() {
479 1 => 'd', // We fixup i8 to i8x8
483 // We use i64x2 as the type for discarded outputs
486 format!("{{{}{}}}", class, idx)
487 } else if reg == InlineAsmReg::AArch64(AArch64InlineAsmReg::x30) {
488 // LLVM doesn't recognize x30
491 format!("{{{}}}", reg.name())
494 InlineAsmRegOrRegClass::RegClass(reg) => match reg {
495 InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => "r",
496 InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg) => "w",
497 InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => "x",
498 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg) => "r",
499 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg_thumb) => "l",
500 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg)
501 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16)
502 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8) => "t",
503 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16)
504 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8)
505 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => "x",
506 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg)
507 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg) => "w",
508 InlineAsmRegClass::Hexagon(HexagonInlineAsmRegClass::reg) => "r",
509 InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg) => "r",
510 InlineAsmRegClass::Mips(MipsInlineAsmRegClass::freg) => "f",
511 InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg16) => "h",
512 InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg32) => "r",
513 InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg64) => "l",
514 InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) => "r",
515 InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => "f",
516 InlineAsmRegClass::X86(X86InlineAsmRegClass::reg) => "r",
517 InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => "Q",
518 InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => "q",
519 InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg)
520 | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg) => "x",
521 InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => "v",
522 InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => "^Yk",
523 InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => "r",
524 InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => {
525 bug!("LLVM backend does not support SPIR-V")
532 /// Converts a modifier into LLVM's equivalent modifier.
535 reg: InlineAsmRegClass,
536 modifier: Option<char>,
539 InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => modifier,
540 InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg)
541 | InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => {
542 if modifier == Some('v') { None } else { modifier }
544 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg)
545 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg_thumb) => None,
546 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg)
547 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16) => None,
548 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg)
549 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16)
550 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8) => Some('P'),
551 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg)
552 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8)
553 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => {
554 if modifier.is_none() {
560 InlineAsmRegClass::Hexagon(_) => None,
561 InlineAsmRegClass::Mips(_) => None,
562 InlineAsmRegClass::Nvptx(_) => None,
563 InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg)
564 | InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => None,
565 InlineAsmRegClass::X86(X86InlineAsmRegClass::reg)
566 | InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => match modifier {
567 None if arch == InlineAsmArch::X86_64 => Some('q'),
569 Some('l') => Some('b'),
570 Some('h') => Some('h'),
571 Some('x') => Some('w'),
572 Some('e') => Some('k'),
573 Some('r') => Some('q'),
576 InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => None,
577 InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::xmm_reg)
578 | InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::ymm_reg)
579 | InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::zmm_reg) => match (reg, modifier) {
580 (X86InlineAsmRegClass::xmm_reg, None) => Some('x'),
581 (X86InlineAsmRegClass::ymm_reg, None) => Some('t'),
582 (X86InlineAsmRegClass::zmm_reg, None) => Some('g'),
583 (_, Some('x')) => Some('x'),
584 (_, Some('y')) => Some('t'),
585 (_, Some('z')) => Some('g'),
588 InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => None,
589 InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => None,
590 InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => {
591 bug!("LLVM backend does not support SPIR-V")
596 /// Type to use for outputs that are discarded. It doesn't really matter what
597 /// the type is, as long as it is valid for the constraint code.
598 fn dummy_output_type(cx: &CodegenCx<'ll, 'tcx>, reg: InlineAsmRegClass) -> &'ll Type {
600 InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => cx.type_i32(),
601 InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg)
602 | InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => {
603 cx.type_vector(cx.type_i64(), 2)
605 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg)
606 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg_thumb) => cx.type_i32(),
607 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg)
608 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16) => cx.type_f32(),
609 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg)
610 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16)
611 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8) => cx.type_f64(),
612 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg)
613 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8)
614 | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => {
615 cx.type_vector(cx.type_i64(), 2)
617 InlineAsmRegClass::Hexagon(HexagonInlineAsmRegClass::reg) => cx.type_i32(),
618 InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg) => cx.type_i32(),
619 InlineAsmRegClass::Mips(MipsInlineAsmRegClass::freg) => cx.type_f32(),
620 InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg16) => cx.type_i16(),
621 InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg32) => cx.type_i32(),
622 InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg64) => cx.type_i64(),
623 InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) => cx.type_i32(),
624 InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => cx.type_f32(),
625 InlineAsmRegClass::X86(X86InlineAsmRegClass::reg)
626 | InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => cx.type_i32(),
627 InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => cx.type_i8(),
628 InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg)
629 | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg)
630 | InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => cx.type_f32(),
631 InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => cx.type_i16(),
632 InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => cx.type_i32(),
633 InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => {
634 bug!("LLVM backend does not support SPIR-V")
639 /// Helper function to get the LLVM type for a Scalar. Pointers are returned as
640 /// the equivalent integer type.
641 fn llvm_asm_scalar_type(cx: &CodegenCx<'ll, 'tcx>, scalar: &Scalar) -> &'ll Type {
643 Primitive::Int(Integer::I8, _) => cx.type_i8(),
644 Primitive::Int(Integer::I16, _) => cx.type_i16(),
645 Primitive::Int(Integer::I32, _) => cx.type_i32(),
646 Primitive::Int(Integer::I64, _) => cx.type_i64(),
647 Primitive::F32 => cx.type_f32(),
648 Primitive::F64 => cx.type_f64(),
649 Primitive::Pointer => cx.type_isize(),
654 /// Fix up an input value to work around LLVM bugs.
656 bx: &mut Builder<'a, 'll, 'tcx>,
657 mut value: &'ll Value,
658 reg: InlineAsmRegClass,
659 layout: &TyAndLayout<'tcx>,
661 match (reg, &layout.abi) {
662 (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg), Abi::Scalar(s)) => {
663 if let Primitive::Int(Integer::I8, _) = s.value {
664 let vec_ty = bx.cx.type_vector(bx.cx.type_i8(), 8);
665 bx.insert_element(bx.const_undef(vec_ty), value, bx.const_i32(0))
670 (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16), Abi::Scalar(s)) => {
671 let elem_ty = llvm_asm_scalar_type(bx.cx, s);
672 let count = 16 / layout.size.bytes();
673 let vec_ty = bx.cx.type_vector(elem_ty, count);
674 if let Primitive::Pointer = s.value {
675 value = bx.ptrtoint(value, bx.cx.type_isize());
677 bx.insert_element(bx.const_undef(vec_ty), value, bx.const_i32(0))
680 InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16),
681 Abi::Vector { element, count },
682 ) if layout.size.bytes() == 8 => {
683 let elem_ty = llvm_asm_scalar_type(bx.cx, element);
684 let vec_ty = bx.cx.type_vector(elem_ty, *count);
685 let indices: Vec<_> = (0..count * 2).map(|x| bx.const_i32(x as i32)).collect();
686 bx.shuffle_vector(value, bx.const_undef(vec_ty), bx.const_vector(&indices))
688 (InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd), Abi::Scalar(s))
689 if s.value == Primitive::F64 =>
691 bx.bitcast(value, bx.cx.type_i64())
694 InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg | X86InlineAsmRegClass::zmm_reg),
696 ) if layout.size.bytes() == 64 => bx.bitcast(value, bx.cx.type_vector(bx.cx.type_f64(), 8)),
698 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg | ArmInlineAsmRegClass::sreg_low16),
701 if let Primitive::Int(Integer::I32, _) = s.value {
702 bx.bitcast(value, bx.cx.type_f32())
708 InlineAsmRegClass::Arm(
709 ArmInlineAsmRegClass::dreg
710 | ArmInlineAsmRegClass::dreg_low8
711 | ArmInlineAsmRegClass::dreg_low16,
715 if let Primitive::Int(Integer::I64, _) = s.value {
716 bx.bitcast(value, bx.cx.type_f64())
721 (InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg), Abi::Scalar(s)) => match s.value {
722 // MIPS only supports register-length arithmetics.
723 Primitive::Int(Integer::I8 | Integer::I16, _) => bx.zext(value, bx.cx.type_i32()),
724 Primitive::F32 => bx.bitcast(value, bx.cx.type_i32()),
725 Primitive::F64 => bx.bitcast(value, bx.cx.type_i64()),
732 /// Fix up an output value to work around LLVM bugs.
733 fn llvm_fixup_output(
734 bx: &mut Builder<'a, 'll, 'tcx>,
735 mut value: &'ll Value,
736 reg: InlineAsmRegClass,
737 layout: &TyAndLayout<'tcx>,
739 match (reg, &layout.abi) {
740 (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg), Abi::Scalar(s)) => {
741 if let Primitive::Int(Integer::I8, _) = s.value {
742 bx.extract_element(value, bx.const_i32(0))
747 (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16), Abi::Scalar(s)) => {
748 value = bx.extract_element(value, bx.const_i32(0));
749 if let Primitive::Pointer = s.value {
750 value = bx.inttoptr(value, layout.llvm_type(bx.cx));
755 InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16),
756 Abi::Vector { element, count },
757 ) if layout.size.bytes() == 8 => {
758 let elem_ty = llvm_asm_scalar_type(bx.cx, element);
759 let vec_ty = bx.cx.type_vector(elem_ty, *count * 2);
760 let indices: Vec<_> = (0..*count).map(|x| bx.const_i32(x as i32)).collect();
761 bx.shuffle_vector(value, bx.const_undef(vec_ty), bx.const_vector(&indices))
763 (InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd), Abi::Scalar(s))
764 if s.value == Primitive::F64 =>
766 bx.bitcast(value, bx.cx.type_f64())
769 InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg | X86InlineAsmRegClass::zmm_reg),
771 ) if layout.size.bytes() == 64 => bx.bitcast(value, layout.llvm_type(bx.cx)),
773 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg | ArmInlineAsmRegClass::sreg_low16),
776 if let Primitive::Int(Integer::I32, _) = s.value {
777 bx.bitcast(value, bx.cx.type_i32())
783 InlineAsmRegClass::Arm(
784 ArmInlineAsmRegClass::dreg
785 | ArmInlineAsmRegClass::dreg_low8
786 | ArmInlineAsmRegClass::dreg_low16,
790 if let Primitive::Int(Integer::I64, _) = s.value {
791 bx.bitcast(value, bx.cx.type_i64())
796 (InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg), Abi::Scalar(s)) => match s.value {
797 // MIPS only supports register-length arithmetics.
798 Primitive::Int(Integer::I8, _) => bx.trunc(value, bx.cx.type_i8()),
799 Primitive::Int(Integer::I16, _) => bx.trunc(value, bx.cx.type_i16()),
800 Primitive::F32 => bx.bitcast(value, bx.cx.type_f32()),
801 Primitive::F64 => bx.bitcast(value, bx.cx.type_f64()),
808 /// Output type to use for llvm_fixup_output.
809 fn llvm_fixup_output_type(
810 cx: &CodegenCx<'ll, 'tcx>,
811 reg: InlineAsmRegClass,
812 layout: &TyAndLayout<'tcx>,
814 match (reg, &layout.abi) {
815 (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg), Abi::Scalar(s)) => {
816 if let Primitive::Int(Integer::I8, _) = s.value {
817 cx.type_vector(cx.type_i8(), 8)
822 (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16), Abi::Scalar(s)) => {
823 let elem_ty = llvm_asm_scalar_type(cx, s);
824 let count = 16 / layout.size.bytes();
825 cx.type_vector(elem_ty, count)
828 InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16),
829 Abi::Vector { element, count },
830 ) if layout.size.bytes() == 8 => {
831 let elem_ty = llvm_asm_scalar_type(cx, element);
832 cx.type_vector(elem_ty, count * 2)
834 (InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd), Abi::Scalar(s))
835 if s.value == Primitive::F64 =>
840 InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg | X86InlineAsmRegClass::zmm_reg),
842 ) if layout.size.bytes() == 64 => cx.type_vector(cx.type_f64(), 8),
844 InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg | ArmInlineAsmRegClass::sreg_low16),
847 if let Primitive::Int(Integer::I32, _) = s.value {
854 InlineAsmRegClass::Arm(
855 ArmInlineAsmRegClass::dreg
856 | ArmInlineAsmRegClass::dreg_low8
857 | ArmInlineAsmRegClass::dreg_low16,
861 if let Primitive::Int(Integer::I64, _) = s.value {
867 (InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg), Abi::Scalar(s)) => match s.value {
868 // MIPS only supports register-length arithmetics.
869 Primitive::Int(Integer::I8 | Integer::I16, _) => cx.type_i32(),
870 Primitive::F32 => cx.type_i32(),
871 Primitive::F64 => cx.type_i64(),
872 _ => layout.llvm_type(cx),
874 _ => layout.llvm_type(cx),