• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown | FileCheck %s --check-prefixes=CHECK,SSE
3; RUN: llc < %s -mtriple=x86_64-unknown -mattr=avx512f | FileCheck %s --check-prefixes=CHECK,AVX512
4
5; This test makes sure that a vector that needs to be promoted that is bitcasted to fp16 is legalized correctly without causing a width mismatch.
6define void @constant_fold_vector_to_half() {
7; CHECK-LABEL: constant_fold_vector_to_half:
8; CHECK:       # %bb.0:
9; CHECK-NEXT:    movw $16384, (%rax) # imm = 0x4000
10; CHECK-NEXT:    retq
11  store volatile half bitcast (<4 x i4> <i4 0, i4 0, i4 0, i4 4> to half), half* undef
12  ret void
13}
14
15; Similarly this makes sure that the opposite bitcast of the above is also legalized without crashing.
16define void @pr38533_2(half %x) {
17; SSE-LABEL: pr38533_2:
18; SSE:       # %bb.0:
19; SSE-NEXT:    pushq %rax
20; SSE-NEXT:    .cfi_def_cfa_offset 16
21; SSE-NEXT:    callq __gnu_f2h_ieee
22; SSE-NEXT:    movw %ax, {{[0-9]+}}(%rsp)
23; SSE-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
24; SSE-NEXT:    movw %ax, (%rax)
25; SSE-NEXT:    popq %rax
26; SSE-NEXT:    .cfi_def_cfa_offset 8
27; SSE-NEXT:    retq
28;
29; AVX512-LABEL: pr38533_2:
30; AVX512:       # %bb.0:
31; AVX512-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
32; AVX512-NEXT:    vmovd %xmm0, %eax
33; AVX512-NEXT:    movw %ax, -{{[0-9]+}}(%rsp)
34; AVX512-NEXT:    movzwl -{{[0-9]+}}(%rsp), %eax
35; AVX512-NEXT:    movw %ax, (%rax)
36; AVX512-NEXT:    retq
37  %a = bitcast half %x to <4 x i4>
38  store volatile <4 x i4> %a, <4 x i4>* undef
39  ret void
40}
41
42; This case is a bitcast from fp16 to a 16-bit wide legal vector type. In this case the result type is legal when the bitcast gets type legalized.
43define void @pr38533_3(half %x) {
44; SSE-LABEL: pr38533_3:
45; SSE:       # %bb.0:
46; SSE-NEXT:    pushq %rax
47; SSE-NEXT:    .cfi_def_cfa_offset 16
48; SSE-NEXT:    callq __gnu_f2h_ieee
49; SSE-NEXT:    movw %ax, (%rsp)
50; SSE-NEXT:    movzwl (%rsp), %eax
51; SSE-NEXT:    movw %ax, (%rax)
52; SSE-NEXT:    popq %rax
53; SSE-NEXT:    .cfi_def_cfa_offset 8
54; SSE-NEXT:    retq
55;
56; AVX512-LABEL: pr38533_3:
57; AVX512:       # %bb.0:
58; AVX512-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
59; AVX512-NEXT:    vmovd %xmm0, %eax
60; AVX512-NEXT:    movw %ax, (%rax)
61; AVX512-NEXT:    retq
62  %a = bitcast half %x to <16 x i1>
63  store volatile <16 x i1> %a, <16 x i1>* undef
64  ret void
65}
66