1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
2 ; RUN: llc -global-isel=0 -mcpu=generic -mtriple=x86_64-apple-darwin %s -o - | FileCheck %s
4 ; For all these tests we disable optimizations through function attributes
5 ; because the code we are exercising here needs phis and we want to keep the
8 ; This code used to crash in SDISel because bf16 was promoted to f32 through
9 ; a `f32 = vector_extract_elt <1 x bf16>, i32 0`, which is illegal.
10 ; The invalid SDNode and thus, the crash was only exposed by the constant
12 define void @phi_vec1bf16_to_f32_with_const_folding(ptr %dst) #0 {
13 ; CHECK-LABEL: phi_vec1bf16_to_f32_with_const_folding:
14 ; CHECK: ## %bb.0: ## %entry
15 ; CHECK-NEXT: pushq %rbx
16 ; CHECK-NEXT: .cfi_def_cfa_offset 16
17 ; CHECK-NEXT: .cfi_offset %rbx, -16
18 ; CHECK-NEXT: movq %rdi, %rbx
19 ; CHECK-NEXT: jmp LBB0_1
20 ; CHECK-NEXT: LBB0_1: ## %bb
21 ; CHECK-NEXT: xorps %xmm0, %xmm0
22 ; CHECK-NEXT: callq ___truncsfbf2
23 ; CHECK-NEXT: pextrw $0, %xmm0, %eax
24 ; CHECK-NEXT: movw %ax, 2(%rbx)
25 ; CHECK-NEXT: popq %rbx
31 %phi = phi <1 x bfloat> [ zeroinitializer, %entry ]
32 %res = shufflevector <1 x bfloat> poison, <1 x bfloat> %phi, <2 x i32> <i32 0, i32 1>
33 store <2 x bfloat> %res, ptr %dst
36 ; Same as phi_vec1bf16_to_f32_with_const_folding but without the constant
38 ; This test exercises the same invalid SDNode, but it happened to work by
39 ; accident before. Here we make sure the fix also work as expected in the
40 ; non-constant folding case.
41 define void @phi_vec1bf16_to_f32(ptr %src, ptr %dst) #0 {
42 ; CHECK-LABEL: phi_vec1bf16_to_f32:
43 ; CHECK: ## %bb.0: ## %entry
44 ; CHECK-NEXT: pushq %rbx
45 ; CHECK-NEXT: .cfi_def_cfa_offset 16
46 ; CHECK-NEXT: .cfi_offset %rbx, -16
47 ; CHECK-NEXT: movq %rsi, %rbx
48 ; CHECK-NEXT: movzwl (%rdi), %eax
49 ; CHECK-NEXT: shll $16, %eax
50 ; CHECK-NEXT: movd %eax, %xmm0
51 ; CHECK-NEXT: jmp LBB1_1
52 ; CHECK-NEXT: LBB1_1: ## %bb
53 ; CHECK-NEXT: callq ___truncsfbf2
54 ; CHECK-NEXT: pextrw $0, %xmm0, %eax
55 ; CHECK-NEXT: movw %ax, 2(%rbx)
56 ; CHECK-NEXT: popq %rbx
59 %input = load <1 x bfloat>, ptr %src
63 %phi = phi <1 x bfloat> [ %input, %entry ]
64 %res = shufflevector <1 x bfloat> poison, <1 x bfloat> %phi, <2 x i32> <i32 0, i32 1>
65 store <2 x bfloat> %res, ptr %dst
70 ; Half type is legal on x86 so nothing special here, but it
71 ; doesn't hurt to be thorough.
72 define void @phi_vec1half_with_const_folding(ptr %dst) #0 {
73 ; CHECK-LABEL: phi_vec1half_with_const_folding:
74 ; CHECK: ## %bb.0: ## %entry
75 ; CHECK-NEXT: xorps %xmm0, %xmm0
76 ; CHECK-NEXT: jmp LBB2_1
77 ; CHECK-NEXT: LBB2_1: ## %bb
78 ; CHECK-NEXT: pextrw $0, %xmm0, %eax
79 ; CHECK-NEXT: movw %ax, 2(%rdi)
85 %phi = phi <1 x half> [ zeroinitializer, %entry ]
86 %res = shufflevector <1 x half> poison, <1 x half> %phi, <2 x i32> <i32 0, i32 1>
87 store <2 x half> %res, ptr %dst
91 ; Half type is legal on x86 so nothing special here, but it
92 ; doesn't hurt to be thorough.
93 ; Same as phi_vec1half_with_constant_folding but without the constant folding.
94 define void @phi_vec1half(ptr %src, ptr %dst) #0 {
95 ; CHECK-LABEL: phi_vec1half:
96 ; CHECK: ## %bb.0: ## %entry
97 ; CHECK-NEXT: pinsrw $0, (%rdi), %xmm0
98 ; CHECK-NEXT: jmp LBB3_1
99 ; CHECK-NEXT: LBB3_1: ## %bb
100 ; CHECK-NEXT: pextrw $0, %xmm0, %eax
101 ; CHECK-NEXT: movw %ax, 2(%rsi)
104 %input = load <1 x half>, ptr %src
108 %phi = phi <1 x half> [ %input, %entry ]
109 %res = shufflevector <1 x half> poison, <1 x half> %phi, <2 x i32> <i32 0, i32 1>
110 store <2 x half> %res, ptr %dst
114 attributes #0 = { noinline optnone }