xref: /aosp_15_r20/external/llvm/test/CodeGen/X86/vzero-excess.ll (revision 9880d6810fe72a1726cb53787c6711e909410d58)
1*9880d681SAndroid Build Coastguard Worker; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2*9880d681SAndroid Build Coastguard Worker; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s
3*9880d681SAndroid Build Coastguard Worker
4*9880d681SAndroid Build Coastguard Worker; In the following 4 tests, the existing call to VZU/VZA ensures clean state before
5*9880d681SAndroid Build Coastguard Worker; the call to the unknown, so we don't need to insert a second VZU at that point.
6*9880d681SAndroid Build Coastguard Worker
7*9880d681SAndroid Build Coastguard Workerdefine <4 x float> @zeroupper_v4f32(<8 x float> *%x, <8 x float> %y) nounwind {
8*9880d681SAndroid Build Coastguard Worker; CHECK-LABEL: zeroupper_v4f32:
9*9880d681SAndroid Build Coastguard Worker; CHECK:       # BB#0:
10*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    pushq %rbx
11*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    subq $48, %rsp
12*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
13*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    movq %rdi, %rbx
14*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vzeroupper
15*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    callq the_unknown
16*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
17*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vaddps (%rbx), %ymm0, %ymm0
18*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
19*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vaddps %xmm1, %xmm0, %xmm0
20*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    addq $48, %rsp
21*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    popq %rbx
22*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vzeroupper
23*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    retq
24*9880d681SAndroid Build Coastguard Worker  call void @llvm.x86.avx.vzeroupper()
25*9880d681SAndroid Build Coastguard Worker  call void @the_unknown()
26*9880d681SAndroid Build Coastguard Worker  %loadx = load <8 x float>, <8 x float> *%x, align 32
27*9880d681SAndroid Build Coastguard Worker  %sum = fadd <8 x float> %loadx, %y
28*9880d681SAndroid Build Coastguard Worker  %lo = shufflevector <8 x float> %sum, <8 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
29*9880d681SAndroid Build Coastguard Worker  %hi = shufflevector <8 x float> %sum, <8 x float> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
30*9880d681SAndroid Build Coastguard Worker  %res = fadd <4 x float> %lo, %hi
31*9880d681SAndroid Build Coastguard Worker  ret <4 x float> %res
32*9880d681SAndroid Build Coastguard Worker}
33*9880d681SAndroid Build Coastguard Worker
34*9880d681SAndroid Build Coastguard Workerdefine <8 x float> @zeroupper_v8f32(<8 x float> %x) nounwind {
35*9880d681SAndroid Build Coastguard Worker; CHECK-LABEL: zeroupper_v8f32:
36*9880d681SAndroid Build Coastguard Worker; CHECK:       # BB#0:
37*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    subq $56, %rsp
38*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
39*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vzeroupper
40*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    callq the_unknown
41*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
42*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    addq $56, %rsp
43*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    retq
44*9880d681SAndroid Build Coastguard Worker  call void @llvm.x86.avx.vzeroupper()
45*9880d681SAndroid Build Coastguard Worker  call void @the_unknown()
46*9880d681SAndroid Build Coastguard Worker  ret <8 x float> %x
47*9880d681SAndroid Build Coastguard Worker}
48*9880d681SAndroid Build Coastguard Worker
49*9880d681SAndroid Build Coastguard Workerdefine <4 x float> @zeroall_v4f32(<8 x float> *%x, <8 x float> %y) nounwind {
50*9880d681SAndroid Build Coastguard Worker; CHECK-LABEL: zeroall_v4f32:
51*9880d681SAndroid Build Coastguard Worker; CHECK:       # BB#0:
52*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    pushq %rbx
53*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    subq $48, %rsp
54*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
55*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    movq %rdi, %rbx
56*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vzeroall
57*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    callq the_unknown
58*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
59*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vaddps (%rbx), %ymm0, %ymm0
60*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
61*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vaddps %xmm1, %xmm0, %xmm0
62*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    addq $48, %rsp
63*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    popq %rbx
64*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vzeroupper
65*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    retq
66*9880d681SAndroid Build Coastguard Worker  call void @llvm.x86.avx.vzeroall()
67*9880d681SAndroid Build Coastguard Worker  call void @the_unknown()
68*9880d681SAndroid Build Coastguard Worker  %loadx = load <8 x float>, <8 x float> *%x, align 32
69*9880d681SAndroid Build Coastguard Worker  %sum = fadd <8 x float> %loadx, %y
70*9880d681SAndroid Build Coastguard Worker  %lo = shufflevector <8 x float> %sum, <8 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
71*9880d681SAndroid Build Coastguard Worker  %hi = shufflevector <8 x float> %sum, <8 x float> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
72*9880d681SAndroid Build Coastguard Worker  %res = fadd <4 x float> %lo, %hi
73*9880d681SAndroid Build Coastguard Worker  ret <4 x float> %res
74*9880d681SAndroid Build Coastguard Worker}
75*9880d681SAndroid Build Coastguard Worker
76*9880d681SAndroid Build Coastguard Workerdefine <8 x float> @zeroall_v8f32(<8 x float> %x) nounwind {
77*9880d681SAndroid Build Coastguard Worker; CHECK-LABEL: zeroall_v8f32:
78*9880d681SAndroid Build Coastguard Worker; CHECK:       # BB#0:
79*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    subq $56, %rsp
80*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
81*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vzeroall
82*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    callq the_unknown
83*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
84*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    addq $56, %rsp
85*9880d681SAndroid Build Coastguard Worker; CHECK-NEXT:    retq
86*9880d681SAndroid Build Coastguard Worker  call void @llvm.x86.avx.vzeroall()
87*9880d681SAndroid Build Coastguard Worker  call void @the_unknown()
88*9880d681SAndroid Build Coastguard Worker  ret <8 x float> %x
89*9880d681SAndroid Build Coastguard Worker}
90*9880d681SAndroid Build Coastguard Worker
91*9880d681SAndroid Build Coastguard Workerdeclare void @llvm.x86.avx.vzeroupper() nounwind readnone
92*9880d681SAndroid Build Coastguard Workerdeclare void @llvm.x86.avx.vzeroall() nounwind readnone
93*9880d681SAndroid Build Coastguard Workerdeclare void @the_unknown() nounwind
94*9880d681SAndroid Build Coastguard Worker
95