1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
use crate::attributes;
use crate::builder::Builder;
use crate::common::Funclet;
use crate::context::CodegenCx;
use crate::llvm;
use crate::llvm_util;
use crate::type_::Type;
use crate::type_of::LayoutLlvmExt;
use crate::value::Value;

use rustc_ast::{InlineAsmOptions, InlineAsmTemplatePiece};
use rustc_codegen_ssa::mir::operand::OperandValue;
use rustc_codegen_ssa::traits::*;
use rustc_data_structures::fx::FxHashMap;
use rustc_middle::ty::layout::TyAndLayout;
use rustc_middle::{bug, span_bug, ty::Instance};
use rustc_span::{Pos, Span};
use rustc_target::abi::*;
use rustc_target::asm::*;

use libc::{c_char, c_uint};
use smallvec::SmallVec;
use tracing::debug;

impl<'ll, 'tcx> AsmBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
    fn codegen_inline_asm(
        &mut self,
        template: &[InlineAsmTemplatePiece],
        operands: &[InlineAsmOperandRef<'tcx, Self>],
        options: InlineAsmOptions,
        line_spans: &[Span],
        instance: Instance<'_>,
        dest_catch_funclet: Option<(Self::BasicBlock, Self::BasicBlock, Option<&Self::Funclet>)>,
    ) {
        let asm_arch = self.tcx.sess.asm_arch.unwrap();

        // Collect the types of output operands
        let mut constraints = vec![];
        let mut clobbers = vec![];
        let mut output_types = vec![];
        let mut op_idx = FxHashMap::default();
        let mut clobbered_x87 = false;
        for (idx, op) in operands.iter().enumerate() {
            match *op {
                InlineAsmOperandRef::Out { reg, late, place } => {
                    let is_target_supported = |reg_class: InlineAsmRegClass| {
                        for &(_, feature) in reg_class.supported_types(asm_arch) {
                            if let Some(feature) = feature {
                                let codegen_fn_attrs = self.tcx.codegen_fn_attrs(instance.def_id());
                                if self.tcx.sess.target_features.contains(&feature)
                                    || codegen_fn_attrs.target_features.contains(&feature)
                                {
                                    return true;
                                }
                            } else {
                                // Register class is unconditionally supported
                                return true;
                            }
                        }
                        false
                    };

                    let mut layout = None;
                    let ty = if let Some(ref place) = place {
                        layout = Some(&place.layout);
                        llvm_fixup_output_type(self.cx, reg.reg_class(), &place.layout)
                    } else if matches!(
                        reg.reg_class(),
                        InlineAsmRegClass::X86(
                            X86InlineAsmRegClass::mmx_reg | X86InlineAsmRegClass::x87_reg
                        )
                    ) {
                        // Special handling for x87/mmx registers: we always
                        // clobber the whole set if one register is marked as
                        // clobbered. This is due to the way LLVM handles the
                        // FP stack in inline assembly.
                        if !clobbered_x87 {
                            clobbered_x87 = true;
                            clobbers.push("~{st}".to_string());
                            for i in 1..=7 {
                                clobbers.push(format!("~{{st({})}}", i));
                            }
                        }
                        continue;
                    } else if !is_target_supported(reg.reg_class())
                        || reg.reg_class().is_clobber_only(asm_arch)
                    {
                        // We turn discarded outputs into clobber constraints
                        // if the target feature needed by the register class is
                        // disabled. This is necessary otherwise LLVM will try
                        // to actually allocate a register for the dummy output.
                        assert!(matches!(reg, InlineAsmRegOrRegClass::Reg(_)));
                        clobbers.push(format!("~{}", reg_to_llvm(reg, None)));
                        continue;
                    } else {
                        // If the output is discarded, we don't really care what
                        // type is used. We're just using this to tell LLVM to
                        // reserve the register.
                        dummy_output_type(self.cx, reg.reg_class())
                    };
                    output_types.push(ty);
                    op_idx.insert(idx, constraints.len());
                    let prefix = if late { "=" } else { "=&" };
                    constraints.push(format!("{}{}", prefix, reg_to_llvm(reg, layout)));
                }
                InlineAsmOperandRef::InOut { reg, late, in_value, out_place } => {
                    let layout = if let Some(ref out_place) = out_place {
                        &out_place.layout
                    } else {
                        // LLVM required tied operands to have the same type,
                        // so we just use the type of the input.
                        &in_value.layout
                    };
                    let ty = llvm_fixup_output_type(self.cx, reg.reg_class(), layout);
                    output_types.push(ty);
                    op_idx.insert(idx, constraints.len());
                    let prefix = if late { "=" } else { "=&" };
                    constraints.push(format!("{}{}", prefix, reg_to_llvm(reg, Some(layout))));
                }
                _ => {}
            }
        }

        // Collect input operands
        let mut inputs = vec![];
        for (idx, op) in operands.iter().enumerate() {
            match *op {
                InlineAsmOperandRef::In { reg, value } => {
                    let llval =
                        llvm_fixup_input(self, value.immediate(), reg.reg_class(), &value.layout);
                    inputs.push(llval);
                    op_idx.insert(idx, constraints.len());
                    constraints.push(reg_to_llvm(reg, Some(&value.layout)));
                }
                InlineAsmOperandRef::InOut { reg, late: _, in_value, out_place: _ } => {
                    let value = llvm_fixup_input(
                        self,
                        in_value.immediate(),
                        reg.reg_class(),
                        &in_value.layout,
                    );
                    inputs.push(value);
                    constraints.push(format!("{}", op_idx[&idx]));
                }
                InlineAsmOperandRef::SymFn { instance } => {
                    inputs.push(self.cx.get_fn(instance));
                    op_idx.insert(idx, constraints.len());
                    constraints.push("s".to_string());
                }
                InlineAsmOperandRef::SymStatic { def_id } => {
                    inputs.push(self.cx.get_static(def_id));
                    op_idx.insert(idx, constraints.len());
                    constraints.push("s".to_string());
                }
                _ => {}
            }
        }

        // Build the template string
        let mut template_str = String::new();
        for piece in template {
            match *piece {
                InlineAsmTemplatePiece::String(ref s) => {
                    if s.contains('$') {
                        for c in s.chars() {
                            if c == '$' {
                                template_str.push_str("$$");
                            } else {
                                template_str.push(c);
                            }
                        }
                    } else {
                        template_str.push_str(s)
                    }
                }
                InlineAsmTemplatePiece::Placeholder { operand_idx, modifier, span: _ } => {
                    match operands[operand_idx] {
                        InlineAsmOperandRef::In { reg, .. }
                        | InlineAsmOperandRef::Out { reg, .. }
                        | InlineAsmOperandRef::InOut { reg, .. } => {
                            let modifier = modifier_to_llvm(asm_arch, reg.reg_class(), modifier);
                            if let Some(modifier) = modifier {
                                template_str.push_str(&format!(
                                    "${{{}:{}}}",
                                    op_idx[&operand_idx], modifier
                                ));
                            } else {
                                template_str.push_str(&format!("${{{}}}", op_idx[&operand_idx]));
                            }
                        }
                        InlineAsmOperandRef::Const { ref string } => {
                            // Const operands get injected directly into the template
                            template_str.push_str(string);
                        }
                        InlineAsmOperandRef::SymFn { .. }
                        | InlineAsmOperandRef::SymStatic { .. } => {
                            // Only emit the raw symbol name
                            template_str.push_str(&format!("${{{}:c}}", op_idx[&operand_idx]));
                        }
                    }
                }
            }
        }

        constraints.append(&mut clobbers);
        if !options.contains(InlineAsmOptions::PRESERVES_FLAGS) {
            match asm_arch {
                InlineAsmArch::AArch64 | InlineAsmArch::Arm => {
                    constraints.push("~{cc}".to_string());
                }
                InlineAsmArch::X86 | InlineAsmArch::X86_64 => {
                    constraints.extend_from_slice(&[
                        "~{dirflag}".to_string(),
                        "~{fpsr}".to_string(),
                        "~{flags}".to_string(),
                    ]);
                }
                InlineAsmArch::RiscV32 | InlineAsmArch::RiscV64 => {
                    constraints.extend_from_slice(&[
                        "~{vtype}".to_string(),
                        "~{vl}".to_string(),
                        "~{vxsat}".to_string(),
                        "~{vxrm}".to_string(),
                    ]);
                }
                InlineAsmArch::Avr => {
                    constraints.push("~{sreg}".to_string());
                }
                InlineAsmArch::Nvptx64 => {}
                InlineAsmArch::PowerPC | InlineAsmArch::PowerPC64 => {}
                InlineAsmArch::Hexagon => {}
                InlineAsmArch::Mips | InlineAsmArch::Mips64 => {}
                InlineAsmArch::S390x => {}
                InlineAsmArch::SpirV => {}
                InlineAsmArch::Wasm32 | InlineAsmArch::Wasm64 => {}
                InlineAsmArch::Bpf => {}
                InlineAsmArch::Msp430 => {
                    constraints.push("~{sr}".to_string());
                }
            }
        }
        if !options.contains(InlineAsmOptions::NOMEM) {
            // This is actually ignored by LLVM, but it's probably best to keep
            // it just in case. LLVM instead uses the ReadOnly/ReadNone
            // attributes on the call instruction to optimize.
            constraints.push("~{memory}".to_string());
        }
        let volatile = !options.contains(InlineAsmOptions::PURE);
        let alignstack = !options.contains(InlineAsmOptions::NOSTACK);
        let output_type = match &output_types[..] {
            [] => self.type_void(),
            [ty] => ty,
            tys => self.type_struct(tys, false),
        };
        let dialect = match asm_arch {
            InlineAsmArch::X86 | InlineAsmArch::X86_64
                if !options.contains(InlineAsmOptions::ATT_SYNTAX) =>
            {
                llvm::AsmDialect::Intel
            }
            _ => llvm::AsmDialect::Att,
        };
        let result = inline_asm_call(
            self,
            &template_str,
            &constraints.join(","),
            &inputs,
            output_type,
            volatile,
            alignstack,
            dialect,
            line_spans,
            options.contains(InlineAsmOptions::MAY_UNWIND),
            dest_catch_funclet,
        )
        .unwrap_or_else(|| span_bug!(line_spans[0], "LLVM asm constraint validation failed"));

        let mut attrs = SmallVec::<[_; 2]>::new();
        if options.contains(InlineAsmOptions::PURE) {
            if options.contains(InlineAsmOptions::NOMEM) {
                attrs.push(llvm::AttributeKind::ReadNone.create_attr(self.cx.llcx));
            } else if options.contains(InlineAsmOptions::READONLY) {
                attrs.push(llvm::AttributeKind::ReadOnly.create_attr(self.cx.llcx));
            }
            attrs.push(llvm::AttributeKind::WillReturn.create_attr(self.cx.llcx));
        } else if options.contains(InlineAsmOptions::NOMEM) {
            attrs.push(llvm::AttributeKind::InaccessibleMemOnly.create_attr(self.cx.llcx));
        } else {
            // LLVM doesn't have an attribute to represent ReadOnly + SideEffect
        }
        attributes::apply_to_callsite(result, llvm::AttributePlace::Function, &{ attrs });

        // Switch to the 'normal' basic block if we did an `invoke` instead of a `call`
        if let Some((dest, _, _)) = dest_catch_funclet {
            self.switch_to_block(dest);
        }

        // Write results to outputs
        for (idx, op) in operands.iter().enumerate() {
            if let InlineAsmOperandRef::Out { reg, place: Some(place), .. }
            | InlineAsmOperandRef::InOut { reg, out_place: Some(place), .. } = *op
            {
                let value = if output_types.len() == 1 {
                    result
                } else {
                    self.extract_value(result, op_idx[&idx] as u64)
                };
                let value = llvm_fixup_output(self, value, reg.reg_class(), &place.layout);
                OperandValue::Immediate(value).store(self, place);
            }
        }
    }
}

impl<'tcx> AsmMethods<'tcx> for CodegenCx<'_, 'tcx> {
    fn codegen_global_asm(
        &self,
        template: &[InlineAsmTemplatePiece],
        operands: &[GlobalAsmOperandRef<'tcx>],
        options: InlineAsmOptions,
        _line_spans: &[Span],
    ) {
        let asm_arch = self.tcx.sess.asm_arch.unwrap();

        // Default to Intel syntax on x86
        let intel_syntax = matches!(asm_arch, InlineAsmArch::X86 | InlineAsmArch::X86_64)
            && !options.contains(InlineAsmOptions::ATT_SYNTAX);

        // Build the template string
        let mut template_str = String::new();
        if intel_syntax {
            template_str.push_str(".intel_syntax\n");
        }
        for piece in template {
            match *piece {
                InlineAsmTemplatePiece::String(ref s) => template_str.push_str(s),
                InlineAsmTemplatePiece::Placeholder { operand_idx, modifier: _, span: _ } => {
                    match operands[operand_idx] {
                        GlobalAsmOperandRef::Const { ref string } => {
                            // Const operands get injected directly into the
                            // template. Note that we don't need to escape $
                            // here unlike normal inline assembly.
                            template_str.push_str(string);
                        }
                        GlobalAsmOperandRef::SymFn { instance } => {
                            let llval = self.get_fn(instance);
                            self.add_compiler_used_global(llval);
                            let symbol = llvm::build_string(|s| unsafe {
                                llvm::LLVMRustGetMangledName(llval, s);
                            })
                            .expect("symbol is not valid UTF-8");
                            template_str.push_str(&symbol);
                        }
                        GlobalAsmOperandRef::SymStatic { def_id } => {
                            let llval = self
                                .renamed_statics
                                .borrow()
                                .get(&def_id)
                                .copied()
                                .unwrap_or_else(|| self.get_static(def_id));
                            self.add_compiler_used_global(llval);
                            let symbol = llvm::build_string(|s| unsafe {
                                llvm::LLVMRustGetMangledName(llval, s);
                            })
                            .expect("symbol is not valid UTF-8");
                            template_str.push_str(&symbol);
                        }
                    }
                }
            }
        }
        if intel_syntax {
            template_str.push_str("\n.att_syntax\n");
        }

        unsafe {
            llvm::LLVMRustAppendModuleInlineAsm(
                self.llmod,
                template_str.as_ptr().cast(),
                template_str.len(),
            );
        }
    }
}

pub(crate) fn inline_asm_call<'ll>(
    bx: &mut Builder<'_, 'll, '_>,
    asm: &str,
    cons: &str,
    inputs: &[&'ll Value],
    output: &'ll llvm::Type,
    volatile: bool,
    alignstack: bool,
    dia: llvm::AsmDialect,
    line_spans: &[Span],
    unwind: bool,
    dest_catch_funclet: Option<(
        &'ll llvm::BasicBlock,
        &'ll llvm::BasicBlock,
        Option<&Funclet<'ll>>,
    )>,
) -> Option<&'ll Value> {
    let volatile = if volatile { llvm::True } else { llvm::False };
    let alignstack = if alignstack { llvm::True } else { llvm::False };
    let can_throw = if unwind { llvm::True } else { llvm::False };

    let argtys = inputs
        .iter()
        .map(|v| {
            debug!("Asm Input Type: {:?}", *v);
            bx.cx.val_ty(*v)
        })
        .collect::<Vec<_>>();

    debug!("Asm Output Type: {:?}", output);
    let fty = bx.cx.type_func(&argtys, output);
    unsafe {
        // Ask LLVM to verify that the constraints are well-formed.
        let constraints_ok = llvm::LLVMRustInlineAsmVerify(fty, cons.as_ptr().cast(), cons.len());
        debug!("constraint verification result: {:?}", constraints_ok);
        if constraints_ok {
            if unwind && llvm_util::get_version() < (13, 0, 0) {
                bx.cx.sess().span_fatal(
                    line_spans[0],
                    "unwinding from inline assembly is only supported on llvm >= 13.",
                );
            }

            let v = llvm::LLVMRustInlineAsm(
                fty,
                asm.as_ptr().cast(),
                asm.len(),
                cons.as_ptr().cast(),
                cons.len(),
                volatile,
                alignstack,
                dia,
                can_throw,
            );

            let call = if let Some((dest, catch, funclet)) = dest_catch_funclet {
                bx.invoke(fty, v, inputs, dest, catch, funclet)
            } else {
                bx.call(fty, v, inputs, None)
            };

            // Store mark in a metadata node so we can map LLVM errors
            // back to source locations.  See #17552.
            let key = "srcloc";
            let kind = llvm::LLVMGetMDKindIDInContext(
                bx.llcx,
                key.as_ptr() as *const c_char,
                key.len() as c_uint,
            );

            // srcloc contains one integer for each line of assembly code.
            // Unfortunately this isn't enough to encode a full span so instead
            // we just encode the start position of each line.
            // FIXME: Figure out a way to pass the entire line spans.
            let mut srcloc = vec![];
            if dia == llvm::AsmDialect::Intel && line_spans.len() > 1 {
                // LLVM inserts an extra line to add the ".intel_syntax", so add
                // a dummy srcloc entry for it.
                //
                // Don't do this if we only have 1 line span since that may be
                // due to the asm template string coming from a macro. LLVM will
                // default to the first srcloc for lines that don't have an
                // associated srcloc.
                srcloc.push(bx.const_i32(0));
            }
            srcloc.extend(line_spans.iter().map(|span| bx.const_i32(span.lo().to_u32() as i32)));
            let md = llvm::LLVMMDNodeInContext(bx.llcx, srcloc.as_ptr(), srcloc.len() as u32);
            llvm::LLVMSetMetadata(call, kind, md);

            Some(call)
        } else {
            // LLVM has detected an issue with our constraints, bail out
            None
        }
    }
}

/// If the register is an xmm/ymm/zmm register then return its index.
fn xmm_reg_index(reg: InlineAsmReg) -> Option<u32> {
    match reg {
        InlineAsmReg::X86(reg)
            if reg as u32 >= X86InlineAsmReg::xmm0 as u32
                && reg as u32 <= X86InlineAsmReg::xmm15 as u32 =>
        {
            Some(reg as u32 - X86InlineAsmReg::xmm0 as u32)
        }
        InlineAsmReg::X86(reg)
            if reg as u32 >= X86InlineAsmReg::ymm0 as u32
                && reg as u32 <= X86InlineAsmReg::ymm15 as u32 =>
        {
            Some(reg as u32 - X86InlineAsmReg::ymm0 as u32)
        }
        InlineAsmReg::X86(reg)
            if reg as u32 >= X86InlineAsmReg::zmm0 as u32
                && reg as u32 <= X86InlineAsmReg::zmm31 as u32 =>
        {
            Some(reg as u32 - X86InlineAsmReg::zmm0 as u32)
        }
        _ => None,
    }
}

/// If the register is an AArch64 vector register then return its index.
fn a64_vreg_index(reg: InlineAsmReg) -> Option<u32> {
    match reg {
        InlineAsmReg::AArch64(reg)
            if reg as u32 >= AArch64InlineAsmReg::v0 as u32
                && reg as u32 <= AArch64InlineAsmReg::v31 as u32 =>
        {
            Some(reg as u32 - AArch64InlineAsmReg::v0 as u32)
        }
        _ => None,
    }
}

/// Converts a register class to an LLVM constraint code.
fn reg_to_llvm(reg: InlineAsmRegOrRegClass, layout: Option<&TyAndLayout<'_>>) -> String {
    match reg {
        // For vector registers LLVM wants the register name to match the type size.
        InlineAsmRegOrRegClass::Reg(reg) => {
            if let Some(idx) = xmm_reg_index(reg) {
                let class = if let Some(layout) = layout {
                    match layout.size.bytes() {
                        64 => 'z',
                        32 => 'y',
                        _ => 'x',
                    }
                } else {
                    // We use f32 as the type for discarded outputs
                    'x'
                };
                format!("{{{}mm{}}}", class, idx)
            } else if let Some(idx) = a64_vreg_index(reg) {
                let class = if let Some(layout) = layout {
                    match layout.size.bytes() {
                        16 => 'q',
                        8 => 'd',
                        4 => 's',
                        2 => 'h',
                        1 => 'd', // We fixup i8 to i8x8
                        _ => unreachable!(),
                    }
                } else {
                    // We use i64x2 as the type for discarded outputs
                    'q'
                };
                format!("{{{}{}}}", class, idx)
            } else if reg == InlineAsmReg::AArch64(AArch64InlineAsmReg::x30) {
                // LLVM doesn't recognize x30
                "{lr}".to_string()
            } else if reg == InlineAsmReg::Arm(ArmInlineAsmReg::r14) {
                // LLVM doesn't recognize r14
                "{lr}".to_string()
            } else {
                format!("{{{}}}", reg.name())
            }
        }
        InlineAsmRegOrRegClass::RegClass(reg) => match reg {
            InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => "r",
            InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg) => "w",
            InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => "x",
            InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::preg) => {
                unreachable!("clobber-only")
            }
            InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg) => "r",
            InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg)
            | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16)
            | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8) => "t",
            InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16)
            | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8)
            | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => "x",
            InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg)
            | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg) => "w",
            InlineAsmRegClass::Hexagon(HexagonInlineAsmRegClass::reg) => "r",
            InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg) => "r",
            InlineAsmRegClass::Mips(MipsInlineAsmRegClass::freg) => "f",
            InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg16) => "h",
            InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg32) => "r",
            InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg64) => "l",
            InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg) => "r",
            InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg_nonzero) => "b",
            InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::freg) => "f",
            InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::cr)
            | InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::xer) => {
                unreachable!("clobber-only")
            }
            InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) => "r",
            InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => "f",
            InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::vreg) => {
                unreachable!("clobber-only")
            }
            InlineAsmRegClass::X86(X86InlineAsmRegClass::reg) => "r",
            InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => "Q",
            InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => "q",
            InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg)
            | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg) => "x",
            InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => "v",
            InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => "^Yk",
            InlineAsmRegClass::X86(
                X86InlineAsmRegClass::x87_reg
                | X86InlineAsmRegClass::mmx_reg
                | X86InlineAsmRegClass::kreg0
                | X86InlineAsmRegClass::tmm_reg,
            ) => unreachable!("clobber-only"),
            InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => "r",
            InlineAsmRegClass::Bpf(BpfInlineAsmRegClass::reg) => "r",
            InlineAsmRegClass::Bpf(BpfInlineAsmRegClass::wreg) => "w",
            InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg) => "r",
            InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg_upper) => "d",
            InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg_pair) => "r",
            InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg_iw) => "w",
            InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg_ptr) => "e",
            InlineAsmRegClass::S390x(S390xInlineAsmRegClass::reg) => "r",
            InlineAsmRegClass::S390x(S390xInlineAsmRegClass::freg) => "f",
            InlineAsmRegClass::Msp430(Msp430InlineAsmRegClass::reg) => "r",
            InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => {
                bug!("LLVM backend does not support SPIR-V")
            }
            InlineAsmRegClass::Err => unreachable!(),
        }
        .to_string(),
    }
}

/// Converts a modifier into LLVM's equivalent modifier.
fn modifier_to_llvm(
    arch: InlineAsmArch,
    reg: InlineAsmRegClass,
    modifier: Option<char>,
) -> Option<char> {
    match reg {
        InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => modifier,
        InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg)
        | InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => {
            if modifier == Some('v') { None } else { modifier }
        }
        InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::preg) => {
            unreachable!("clobber-only")
        }
        InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg) => None,
        InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg)
        | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16) => None,
        InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg)
        | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16)
        | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8) => Some('P'),
        InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg)
        | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8)
        | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => {
            if modifier.is_none() {
                Some('q')
            } else {
                modifier
            }
        }
        InlineAsmRegClass::Hexagon(_) => None,
        InlineAsmRegClass::Mips(_) => None,
        InlineAsmRegClass::Nvptx(_) => None,
        InlineAsmRegClass::PowerPC(_) => None,
        InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg)
        | InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => None,
        InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::vreg) => {
            unreachable!("clobber-only")
        }
        InlineAsmRegClass::X86(X86InlineAsmRegClass::reg)
        | InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => match modifier {
            None if arch == InlineAsmArch::X86_64 => Some('q'),
            None => Some('k'),
            Some('l') => Some('b'),
            Some('h') => Some('h'),
            Some('x') => Some('w'),
            Some('e') => Some('k'),
            Some('r') => Some('q'),
            _ => unreachable!(),
        },
        InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => None,
        InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::xmm_reg)
        | InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::ymm_reg)
        | InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::zmm_reg) => match (reg, modifier) {
            (X86InlineAsmRegClass::xmm_reg, None) => Some('x'),
            (X86InlineAsmRegClass::ymm_reg, None) => Some('t'),
            (X86InlineAsmRegClass::zmm_reg, None) => Some('g'),
            (_, Some('x')) => Some('x'),
            (_, Some('y')) => Some('t'),
            (_, Some('z')) => Some('g'),
            _ => unreachable!(),
        },
        InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => None,
        InlineAsmRegClass::X86(
            X86InlineAsmRegClass::x87_reg
            | X86InlineAsmRegClass::mmx_reg
            | X86InlineAsmRegClass::kreg0
            | X86InlineAsmRegClass::tmm_reg,
        ) => {
            unreachable!("clobber-only")
        }
        InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => None,
        InlineAsmRegClass::Bpf(_) => None,
        InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg_pair)
        | InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg_iw)
        | InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg_ptr) => match modifier {
            Some('h') => Some('B'),
            Some('l') => Some('A'),
            _ => None,
        },
        InlineAsmRegClass::Avr(_) => None,
        InlineAsmRegClass::S390x(_) => None,
        InlineAsmRegClass::Msp430(_) => None,
        InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => {
            bug!("LLVM backend does not support SPIR-V")
        }
        InlineAsmRegClass::Err => unreachable!(),
    }
}

/// Type to use for outputs that are discarded. It doesn't really matter what
/// the type is, as long as it is valid for the constraint code.
fn dummy_output_type<'ll>(cx: &CodegenCx<'ll, '_>, reg: InlineAsmRegClass) -> &'ll Type {
    match reg {
        InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::reg) => cx.type_i32(),
        InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg)
        | InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16) => {
            cx.type_vector(cx.type_i64(), 2)
        }
        InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::preg) => {
            unreachable!("clobber-only")
        }
        InlineAsmRegClass::Arm(ArmInlineAsmRegClass::reg) => cx.type_i32(),
        InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg)
        | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg_low16) => cx.type_f32(),
        InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg)
        | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low16)
        | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg_low8) => cx.type_f64(),
        InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg)
        | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low8)
        | InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => {
            cx.type_vector(cx.type_i64(), 2)
        }
        InlineAsmRegClass::Hexagon(HexagonInlineAsmRegClass::reg) => cx.type_i32(),
        InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg) => cx.type_i32(),
        InlineAsmRegClass::Mips(MipsInlineAsmRegClass::freg) => cx.type_f32(),
        InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg16) => cx.type_i16(),
        InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg32) => cx.type_i32(),
        InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg64) => cx.type_i64(),
        InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg) => cx.type_i32(),
        InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::reg_nonzero) => cx.type_i32(),
        InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::freg) => cx.type_f64(),
        InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::cr)
        | InlineAsmRegClass::PowerPC(PowerPCInlineAsmRegClass::xer) => {
            unreachable!("clobber-only")
        }
        InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) => cx.type_i32(),
        InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => cx.type_f32(),
        InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::vreg) => {
            unreachable!("clobber-only")
        }
        InlineAsmRegClass::X86(X86InlineAsmRegClass::reg)
        | InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => cx.type_i32(),
        InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => cx.type_i8(),
        InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg)
        | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg)
        | InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => cx.type_f32(),
        InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => cx.type_i16(),
        InlineAsmRegClass::X86(
            X86InlineAsmRegClass::x87_reg
            | X86InlineAsmRegClass::mmx_reg
            | X86InlineAsmRegClass::kreg0
            | X86InlineAsmRegClass::tmm_reg,
        ) => {
            unreachable!("clobber-only")
        }
        InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => cx.type_i32(),
        InlineAsmRegClass::Bpf(BpfInlineAsmRegClass::reg) => cx.type_i64(),
        InlineAsmRegClass::Bpf(BpfInlineAsmRegClass::wreg) => cx.type_i32(),
        InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg) => cx.type_i8(),
        InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg_upper) => cx.type_i8(),
        InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg_pair) => cx.type_i16(),
        InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg_iw) => cx.type_i16(),
        InlineAsmRegClass::Avr(AvrInlineAsmRegClass::reg_ptr) => cx.type_i16(),
        InlineAsmRegClass::S390x(S390xInlineAsmRegClass::reg) => cx.type_i32(),
        InlineAsmRegClass::S390x(S390xInlineAsmRegClass::freg) => cx.type_f64(),
        InlineAsmRegClass::Msp430(Msp430InlineAsmRegClass::reg) => cx.type_i16(),
        InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => {
            bug!("LLVM backend does not support SPIR-V")
        }
        InlineAsmRegClass::Err => unreachable!(),
    }
}

/// Helper function to get the LLVM type for a Scalar. Pointers are returned as
/// the equivalent integer type.
fn llvm_asm_scalar_type<'ll>(cx: &CodegenCx<'ll, '_>, scalar: Scalar) -> &'ll Type {
    match scalar.primitive() {
        Primitive::Int(Integer::I8, _) => cx.type_i8(),
        Primitive::Int(Integer::I16, _) => cx.type_i16(),
        Primitive::Int(Integer::I32, _) => cx.type_i32(),
        Primitive::Int(Integer::I64, _) => cx.type_i64(),
        Primitive::F32 => cx.type_f32(),
        Primitive::F64 => cx.type_f64(),
        Primitive::Pointer => cx.type_isize(),
        _ => unreachable!(),
    }
}

/// Fix up an input value to work around LLVM bugs.
fn llvm_fixup_input<'ll, 'tcx>(
    bx: &mut Builder<'_, 'll, 'tcx>,
    mut value: &'ll Value,
    reg: InlineAsmRegClass,
    layout: &TyAndLayout<'tcx>,
) -> &'ll Value {
    match (reg, layout.abi) {
        (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg), Abi::Scalar(s)) => {
            if let Primitive::Int(Integer::I8, _) = s.primitive() {
                let vec_ty = bx.cx.type_vector(bx.cx.type_i8(), 8);
                bx.insert_element(bx.const_undef(vec_ty), value, bx.const_i32(0))
            } else {
                value
            }
        }
        (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16), Abi::Scalar(s)) => {
            let elem_ty = llvm_asm_scalar_type(bx.cx, s);
            let count = 16 / layout.size.bytes();
            let vec_ty = bx.cx.type_vector(elem_ty, count);
            if let Primitive::Pointer = s.primitive() {
                value = bx.ptrtoint(value, bx.cx.type_isize());
            }
            bx.insert_element(bx.const_undef(vec_ty), value, bx.const_i32(0))
        }
        (
            InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16),
            Abi::Vector { element, count },
        ) if layout.size.bytes() == 8 => {
            let elem_ty = llvm_asm_scalar_type(bx.cx, element);
            let vec_ty = bx.cx.type_vector(elem_ty, count);
            let indices: Vec<_> = (0..count * 2).map(|x| bx.const_i32(x as i32)).collect();
            bx.shuffle_vector(value, bx.const_undef(vec_ty), bx.const_vector(&indices))
        }
        (InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd), Abi::Scalar(s))
            if s.primitive() == Primitive::F64 =>
        {
            bx.bitcast(value, bx.cx.type_i64())
        }
        (
            InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg | X86InlineAsmRegClass::zmm_reg),
            Abi::Vector { .. },
        ) if layout.size.bytes() == 64 => bx.bitcast(value, bx.cx.type_vector(bx.cx.type_f64(), 8)),
        (
            InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg | ArmInlineAsmRegClass::sreg_low16),
            Abi::Scalar(s),
        ) => {
            if let Primitive::Int(Integer::I32, _) = s.primitive() {
                bx.bitcast(value, bx.cx.type_f32())
            } else {
                value
            }
        }
        (
            InlineAsmRegClass::Arm(
                ArmInlineAsmRegClass::dreg
                | ArmInlineAsmRegClass::dreg_low8
                | ArmInlineAsmRegClass::dreg_low16,
            ),
            Abi::Scalar(s),
        ) => {
            if let Primitive::Int(Integer::I64, _) = s.primitive() {
                bx.bitcast(value, bx.cx.type_f64())
            } else {
                value
            }
        }
        (InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg), Abi::Scalar(s)) => {
            match s.primitive() {
                // MIPS only supports register-length arithmetics.
                Primitive::Int(Integer::I8 | Integer::I16, _) => bx.zext(value, bx.cx.type_i32()),
                Primitive::F32 => bx.bitcast(value, bx.cx.type_i32()),
                Primitive::F64 => bx.bitcast(value, bx.cx.type_i64()),
                _ => value,
            }
        }
        _ => value,
    }
}

/// Fix up an output value to work around LLVM bugs.
fn llvm_fixup_output<'ll, 'tcx>(
    bx: &mut Builder<'_, 'll, 'tcx>,
    mut value: &'ll Value,
    reg: InlineAsmRegClass,
    layout: &TyAndLayout<'tcx>,
) -> &'ll Value {
    match (reg, layout.abi) {
        (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg), Abi::Scalar(s)) => {
            if let Primitive::Int(Integer::I8, _) = s.primitive() {
                bx.extract_element(value, bx.const_i32(0))
            } else {
                value
            }
        }
        (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16), Abi::Scalar(s)) => {
            value = bx.extract_element(value, bx.const_i32(0));
            if let Primitive::Pointer = s.primitive() {
                value = bx.inttoptr(value, layout.llvm_type(bx.cx));
            }
            value
        }
        (
            InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16),
            Abi::Vector { element, count },
        ) if layout.size.bytes() == 8 => {
            let elem_ty = llvm_asm_scalar_type(bx.cx, element);
            let vec_ty = bx.cx.type_vector(elem_ty, count * 2);
            let indices: Vec<_> = (0..count).map(|x| bx.const_i32(x as i32)).collect();
            bx.shuffle_vector(value, bx.const_undef(vec_ty), bx.const_vector(&indices))
        }
        (InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd), Abi::Scalar(s))
            if s.primitive() == Primitive::F64 =>
        {
            bx.bitcast(value, bx.cx.type_f64())
        }
        (
            InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg | X86InlineAsmRegClass::zmm_reg),
            Abi::Vector { .. },
        ) if layout.size.bytes() == 64 => bx.bitcast(value, layout.llvm_type(bx.cx)),
        (
            InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg | ArmInlineAsmRegClass::sreg_low16),
            Abi::Scalar(s),
        ) => {
            if let Primitive::Int(Integer::I32, _) = s.primitive() {
                bx.bitcast(value, bx.cx.type_i32())
            } else {
                value
            }
        }
        (
            InlineAsmRegClass::Arm(
                ArmInlineAsmRegClass::dreg
                | ArmInlineAsmRegClass::dreg_low8
                | ArmInlineAsmRegClass::dreg_low16,
            ),
            Abi::Scalar(s),
        ) => {
            if let Primitive::Int(Integer::I64, _) = s.primitive() {
                bx.bitcast(value, bx.cx.type_i64())
            } else {
                value
            }
        }
        (InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg), Abi::Scalar(s)) => {
            match s.primitive() {
                // MIPS only supports register-length arithmetics.
                Primitive::Int(Integer::I8, _) => bx.trunc(value, bx.cx.type_i8()),
                Primitive::Int(Integer::I16, _) => bx.trunc(value, bx.cx.type_i16()),
                Primitive::F32 => bx.bitcast(value, bx.cx.type_f32()),
                Primitive::F64 => bx.bitcast(value, bx.cx.type_f64()),
                _ => value,
            }
        }
        _ => value,
    }
}

/// Output type to use for llvm_fixup_output.
fn llvm_fixup_output_type<'ll, 'tcx>(
    cx: &CodegenCx<'ll, 'tcx>,
    reg: InlineAsmRegClass,
    layout: &TyAndLayout<'tcx>,
) -> &'ll Type {
    match (reg, layout.abi) {
        (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg), Abi::Scalar(s)) => {
            if let Primitive::Int(Integer::I8, _) = s.primitive() {
                cx.type_vector(cx.type_i8(), 8)
            } else {
                layout.llvm_type(cx)
            }
        }
        (InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16), Abi::Scalar(s)) => {
            let elem_ty = llvm_asm_scalar_type(cx, s);
            let count = 16 / layout.size.bytes();
            cx.type_vector(elem_ty, count)
        }
        (
            InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::vreg_low16),
            Abi::Vector { element, count },
        ) if layout.size.bytes() == 8 => {
            let elem_ty = llvm_asm_scalar_type(cx, element);
            cx.type_vector(elem_ty, count * 2)
        }
        (InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd), Abi::Scalar(s))
            if s.primitive() == Primitive::F64 =>
        {
            cx.type_i64()
        }
        (
            InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg | X86InlineAsmRegClass::zmm_reg),
            Abi::Vector { .. },
        ) if layout.size.bytes() == 64 => cx.type_vector(cx.type_f64(), 8),
        (
            InlineAsmRegClass::Arm(ArmInlineAsmRegClass::sreg | ArmInlineAsmRegClass::sreg_low16),
            Abi::Scalar(s),
        ) => {
            if let Primitive::Int(Integer::I32, _) = s.primitive() {
                cx.type_f32()
            } else {
                layout.llvm_type(cx)
            }
        }
        (
            InlineAsmRegClass::Arm(
                ArmInlineAsmRegClass::dreg
                | ArmInlineAsmRegClass::dreg_low8
                | ArmInlineAsmRegClass::dreg_low16,
            ),
            Abi::Scalar(s),
        ) => {
            if let Primitive::Int(Integer::I64, _) = s.primitive() {
                cx.type_f64()
            } else {
                layout.llvm_type(cx)
            }
        }
        (InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg), Abi::Scalar(s)) => {
            match s.primitive() {
                // MIPS only supports register-length arithmetics.
                Primitive::Int(Integer::I8 | Integer::I16, _) => cx.type_i32(),
                Primitive::F32 => cx.type_i32(),
                Primitive::F64 => cx.type_i64(),
                _ => layout.llvm_type(cx),
            }
        }
        _ => layout.llvm_type(cx),
    }
}