120
|
1 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck %s
|
|
2 ; We may have subregister live ranges that are undefined on some paths. The
|
|
3 ; verifier should not complain about this.
|
|
4
|
|
5
|
|
6 ; CHECK-LABEL: {{^}}func:
|
121
|
7 define amdgpu_kernel void @func() #0 {
|
120
|
8 B0:
|
|
9 br i1 undef, label %B1, label %B2
|
|
10
|
|
11 B1:
|
|
12 br label %B2
|
|
13
|
|
14 B2:
|
|
15 %v0 = phi <4 x float> [ zeroinitializer, %B1 ], [ <float 0.0, float 0.0, float 0.0, float undef>, %B0 ]
|
|
16 br i1 undef, label %B30.1, label %B30.2
|
|
17
|
|
18 B30.1:
|
|
19 %sub = fsub <4 x float> %v0, undef
|
|
20 br label %B30.2
|
|
21
|
|
22 B30.2:
|
|
23 %v3 = phi <4 x float> [ %sub, %B30.1 ], [ %v0, %B2 ]
|
|
24 %ve0 = extractelement <4 x float> %v3, i32 0
|
|
25 store float %ve0, float addrspace(3)* undef, align 4
|
|
26 ret void
|
|
27 }
|
|
28
|
|
29 ; FIXME: Extra undef subregister copy should be removed before
|
|
30 ; overwritten with defined copy
|
|
31 ; CHECK-LABEL: {{^}}valley_partially_undef_copy:
|
|
32 define amdgpu_ps float @valley_partially_undef_copy() #0 {
|
|
33 bb:
|
|
34 %tmp = load volatile i32, i32 addrspace(1)* undef, align 4
|
|
35 %tmp1 = load volatile i32, i32 addrspace(1)* undef, align 4
|
|
36 %tmp2 = insertelement <4 x i32> undef, i32 %tmp1, i32 0
|
|
37 %tmp3 = insertelement <4 x i32> %tmp2, i32 %tmp1, i32 1
|
121
|
38 %tmp3.cast = bitcast <4 x i32> %tmp3 to <4 x float>
|
|
39 %tmp4 = call <4 x float> @llvm.amdgcn.image.sample.v4f32.v4f32.v8i32(<4 x float> %tmp3.cast, <8 x i32> undef, <4 x i32> undef, i32 15, i1 false, i1 false, i1 false, i1 false, i1 false)
|
120
|
40 %tmp5 = extractelement <4 x float> %tmp4, i32 0
|
|
41 %tmp6 = fmul float %tmp5, undef
|
|
42 %tmp7 = fadd float %tmp6, %tmp6
|
|
43 %tmp8 = insertelement <4 x i32> %tmp2, i32 %tmp, i32 1
|
|
44 store <4 x i32> %tmp8, <4 x i32> addrspace(1)* undef, align 16
|
|
45 store float %tmp7, float addrspace(1)* undef, align 4
|
|
46 br label %bb9
|
|
47
|
|
48 bb9: ; preds = %bb9, %bb
|
|
49 %tmp10 = icmp eq i32 %tmp, 0
|
|
50 br i1 %tmp10, label %bb9, label %bb11
|
|
51
|
|
52 bb11: ; preds = %bb9
|
|
53 store <4 x i32> %tmp2, <4 x i32> addrspace(1)* undef, align 16
|
|
54 ret float undef
|
|
55 }
|
|
56
|
|
57 ; FIXME: Should be able to remove the undef copies
|
|
58
|
|
59 ; CHECK-LABEL: {{^}}partially_undef_copy:
|
|
60 ; CHECK: v_mov_b32_e32 v5, 5
|
|
61 ; CHECK: v_mov_b32_e32 v6, 6
|
|
62
|
|
63 ; CHECK: v_mov_b32_e32 v[[OUTPUT_LO:[0-9]+]], v5
|
|
64
|
|
65 ; Undef copy
|
|
66 ; CHECK: v_mov_b32_e32 v1, v6
|
|
67
|
|
68 ; undef copy
|
|
69 ; CHECK: v_mov_b32_e32 v2, v7
|
|
70
|
|
71 ; CHECK: v_mov_b32_e32 v[[OUTPUT_HI:[0-9]+]], v8
|
|
72 ; CHECK: v_mov_b32_e32 v[[OUTPUT_LO]], v6
|
|
73
|
|
74 ; CHECK: buffer_store_dwordx4 v{{\[}}[[OUTPUT_LO]]:[[OUTPUT_HI]]{{\]}}
|
121
|
75 define amdgpu_kernel void @partially_undef_copy() #0 {
|
|
76 %tmp0 = call i32 asm sideeffect "v_mov_b32_e32 v5, 5", "={v5}"()
|
|
77 %tmp1 = call i32 asm sideeffect "v_mov_b32_e32 v6, 6", "={v6}"()
|
120
|
78
|
|
79 %partially.undef.0 = insertelement <4 x i32> undef, i32 %tmp0, i32 0
|
|
80 %partially.undef.1 = insertelement <4 x i32> %partially.undef.0, i32 %tmp1, i32 0
|
|
81
|
|
82 store volatile <4 x i32> %partially.undef.1, <4 x i32> addrspace(1)* undef, align 16
|
121
|
83 tail call void asm sideeffect "v_nop", "v={v[5:8]}"(<4 x i32> %partially.undef.0)
|
120
|
84 ret void
|
|
85 }
|
|
86
|
121
|
87 declare <4 x float> @llvm.amdgcn.image.sample.v4f32.v4f32.v8i32(<4 x float>, <8 x i32>, <4 x i32>, i32, i1, i1, i1, i1, i1) #1
|
120
|
88
|
|
89 attributes #0 = { nounwind }
|
121
|
90 attributes #1 = { nounwind readonly }
|