1 # RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx803 -run-pass=simple-register-coalescing -o - %s | FileCheck -check-prefix=GCN %s
3 # With one version of the D48102 fix, this test failed with
4 # Assertion failed: (Id != S.end() && T != S.end() && T->valno == Id->valno), function pruneSubRegValues, file ../lib/CodeGen/RegisterCoalescer.cpp, line 2875.
9 target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5"
10 target triple = "amdgcn--amdpal"
12 ; Function Attrs: nounwind
13 define amdgpu_ps void @_amdgpu_ps_main(float %arg) #0 {
17 attributes #0 = { nounwind "InitialPSInputAddr"="3842" "target-cpu"="gfx803" }
22 tracksRegLiveness: true
25 successors: %bb.2(0x40000000), %bb.1(0x40000000)
26 %0:sreg_64 = COPY $exec
27 %1:sgpr_32 = S_MOV_B32 0
28 undef %2.sub0:sgpr_128 = COPY %1
29 %2.sub1:sgpr_128 = COPY %1
30 %2.sub2:sgpr_128 = COPY %1
31 %2.sub3:sgpr_128 = COPY %1
32 $exec = S_WQM_B64 $exec, implicit-def dead $scc
33 S_CBRANCH_SCC0 %bb.2, implicit undef $scc
36 successors: %bb.3(0x80000000)
37 %3:sgpr_128 = COPY killed %2
38 %4:vreg_128 = COPY killed %3
39 %5:vreg_128 = COPY killed %4
43 successors: %bb.4(0x80000000)
44 %6:vgpr_32 = V_MUL_F32_e32 1031798784, undef %7:vgpr_32, implicit $exec
45 %8:vgpr_32 = V_FLOOR_F32_e32 killed %6, implicit $exec
46 %9:vgpr_32 = V_ADD_F32_e32 0, killed %8, implicit $exec
47 %10:vgpr_32 = V_CVT_U32_F32_e32 killed %9, implicit $exec
48 %11:vgpr_32 = V_LSHLREV_B32_e32 1, killed %10, implicit $exec
49 %12:sreg_64 = S_MOV_B64 0
50 %13:sgpr_128 = COPY killed %2
51 %14:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
52 %15:vreg_128 = COPY killed %13
53 %16:sreg_64 = COPY killed %12
54 %17:vreg_128 = IMPLICIT_DEF
55 %18:vgpr_32 = COPY killed %14
56 %19:vreg_128 = COPY killed %15
60 successors: %bb.17(0x80000000)
61 %20:vreg_128 = COPY killed %5
65 successors: %bb.8(0x40000000), %bb.9(0x40000000)
66 %21:vreg_128 = COPY killed %19
67 %22:vgpr_32 = COPY killed %18
68 %23:vreg_128 = COPY killed %17
69 %24:sreg_64 = COPY killed %16
70 %25:vgpr_32 = V_OR_B32_e32 %22, %11, implicit $exec
71 %26:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN killed %25, undef %27:sgpr_128, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from constant-pool, align 1, addrspace 4)
72 %28:vgpr_32 = V_LSHRREV_B32_e32 30, killed %26.sub0, implicit $exec
73 %29:vreg_128 = COPY killed %21
74 %29.sub0:vreg_128 = COPY %1
75 %30:sreg_64 = V_CMP_NE_U32_e64 0, %28, implicit $exec
76 %31:sreg_64_xexec = V_CMP_EQ_U32_e64 0, %28, implicit $exec
77 dead %32:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, -1, killed %31, implicit $exec
78 %33:vreg_128 = COPY %29
79 %33.sub1:vreg_128 = COPY undef %32
80 %34:vgpr_32 = V_MOV_B32_e32 -1, implicit $exec
81 %35:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
82 %36:sreg_64 = COPY %24
83 %37:vreg_128 = COPY %23
84 %38:vreg_128 = IMPLICIT_DEF
85 %39:vreg_128 = IMPLICIT_DEF
86 %40:vgpr_32 = IMPLICIT_DEF
87 %41:vgpr_32 = COPY killed %35
88 %42:vgpr_32 = COPY killed %34
89 %43:sreg_64 = COPY $exec, implicit-def $exec
90 %44:sreg_64 = S_AND_B64 %43, %30, implicit-def dead $scc
91 %45:sreg_64 = S_XOR_B64 %44, %43, implicit-def dead $scc
92 $exec = S_MOV_B64_term killed %44
93 SI_MASK_BRANCH %bb.9, implicit $exec
97 successors: %bb.9(0x80000000)
98 $exec = S_OR_B64 $exec, %46, implicit-def $scc
99 %47:vgpr_32 = COPY killed %48
100 %49:vgpr_32 = COPY killed %50
101 %51:vreg_128 = COPY killed %52
102 %53:vreg_128 = COPY killed %54
103 %55:sreg_64 = COPY killed %56
104 %57:sreg_64 = S_AND_B64 $exec, %46, implicit-def $scc
105 %57:sreg_64 = S_OR_B64 %57, killed %55, implicit-def $scc
106 %58:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
107 %36:sreg_64 = COPY killed %57
108 %37:vreg_128 = COPY killed %53
109 %38:vreg_128 = COPY killed %59
110 %39:vreg_128 = COPY killed %51
111 %40:vgpr_32 = COPY killed %49
112 %41:vgpr_32 = COPY killed %47
113 %42:vgpr_32 = COPY killed %58
117 successors: %bb.7(0x40000000), %bb.13(0x40000000)
118 $exec = S_OR_B64 $exec, killed %60, implicit-def $scc
119 %61:sreg_64 = V_CMP_NE_U32_e64 0, killed %62, implicit $exec
120 %63:vreg_128 = COPY killed %64
121 %65:vgpr_32 = COPY killed %66
122 %67:sreg_64 = COPY $exec, implicit-def $exec
123 %68:sreg_64 = S_AND_B64 %67, %61, implicit-def dead $scc
124 $exec = S_MOV_B64_term killed %68
125 SI_MASK_BRANCH %bb.13, implicit $exec
129 successors: %bb.13(0x80000000)
130 %69:vgpr_32 = V_MOV_B32_e32 -1, implicit $exec
131 %70:vreg_128 = COPY killed %33
132 %63:vreg_128 = COPY killed %70
133 %65:vgpr_32 = COPY killed %69
137 successors: %bb.10(0x80000000)
138 %71:sreg_64 = S_MOV_B64 0
139 %72:vreg_128 = COPY %33
140 %73:sreg_64 = COPY killed %71
141 %74:vreg_128 = COPY killed %72
142 %75:vreg_128 = COPY killed %29
146 successors: %bb.6(0x04000000), %bb.4(0x7c000000)
147 $exec = S_OR_B64 $exec, %45, implicit-def $scc
148 %62:vgpr_32 = COPY killed %42
149 %66:vgpr_32 = COPY killed %41
150 %76:vgpr_32 = COPY killed %40
151 %77:vreg_128 = COPY killed %39
152 %64:vreg_128 = COPY killed %38
153 %78:vreg_128 = COPY killed %37
154 %79:sreg_64 = COPY killed %36
155 %60:sreg_64 = S_AND_B64 $exec, %45, implicit-def $scc
156 %60:sreg_64 = S_OR_B64 %60, killed %79, implicit-def $scc
157 %80:vreg_128 = COPY %78
158 %16:sreg_64 = COPY %60
159 %17:vreg_128 = COPY killed %80
160 %18:vgpr_32 = COPY killed %76
161 %19:vreg_128 = COPY killed %77
162 $exec = S_ANDN2_B64_term $exec, %60, implicit-def $scc
163 S_CBRANCH_EXECNZ %bb.4, implicit $exec
167 successors: %bb.11(0x80000000)
168 %81:vreg_128 = COPY killed %75
169 %82:vreg_128 = COPY killed %74
170 %83:sreg_64 = COPY killed %73
173 successors: %bb.12(0x04000000), %bb.10(0x7c000000)
174 undef %59.sub0:vreg_128 = COPY %81.sub0
175 %59.sub2:vreg_128 = COPY %82.sub2
176 %59.sub3:vreg_128 = COPY killed %82.sub3
177 %84:sreg_64 = V_CMP_GE_U32_e64 killed %81.sub0, %28, implicit $exec
178 %85:sreg_64 = S_OR_B64 killed %84, killed %83, implicit-def $scc
179 %86:vreg_128 = COPY %59
180 %73:sreg_64 = COPY %85
181 %74:vreg_128 = COPY %59
182 %75:vreg_128 = COPY killed %86
183 $exec = S_ANDN2_B64_term $exec, %85, implicit-def $scc
184 S_CBRANCH_EXECNZ %bb.10, implicit $exec
188 successors: %bb.15(0x40000000), %bb.5(0x40000000)
189 $exec = S_OR_B64 $exec, killed %85, implicit-def $scc
190 %87:sreg_64 = V_CMP_LT_U32_e64 11, killed %28, implicit $exec
191 %88:vgpr_32 = V_MOV_B32_e32 -1, implicit $exec
192 %56:sreg_64 = COPY %24
193 %54:vreg_128 = COPY killed %23
194 %52:vreg_128 = IMPLICIT_DEF
195 %50:vgpr_32 = IMPLICIT_DEF
196 %48:vgpr_32 = COPY killed %88
197 %89:sreg_64 = COPY $exec, implicit-def $exec
198 %90:sreg_64 = S_AND_B64 %89, %87, implicit-def dead $scc
199 %46:sreg_64 = S_XOR_B64 %90, %89, implicit-def dead $scc
200 $exec = S_MOV_B64_term killed %90
201 SI_MASK_BRANCH %bb.5, implicit $exec
205 successors: %bb.14(0x40000000), %bb.16(0x40000000)
206 $exec = S_OR_B64 $exec, killed %67, implicit-def $scc
207 %91:vgpr_32 = COPY killed %65
208 %92:vreg_128 = COPY killed %63
209 %93:sreg_64 = V_CMP_NE_U32_e64 0, killed %91, implicit $exec
210 %94:vreg_128 = COPY killed %78
211 %95:sreg_64 = COPY $exec, implicit-def $exec
212 %96:sreg_64 = S_AND_B64 %95, %93, implicit-def dead $scc
213 $exec = S_MOV_B64_term killed %96
214 SI_MASK_BRANCH %bb.16, implicit $exec
218 successors: %bb.16(0x80000000)
219 %97:vreg_128 = COPY killed %92
220 %94:vreg_128 = COPY killed %97
224 successors: %bb.5(0x80000000)
225 %98:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
226 %99:sreg_64 = V_CMP_NE_U32_e64 0, killed %22, implicit $exec
227 %100:sreg_64 = S_OR_B64 killed %99, killed %24, implicit-def $scc
228 %101:vreg_128 = COPY %59
229 %102:vgpr_32 = V_MOV_B32_e32 1, implicit $exec
230 %56:sreg_64 = COPY killed %100
231 %54:vreg_128 = COPY killed %101
232 %52:vreg_128 = COPY %59
233 %50:vgpr_32 = COPY killed %102
234 %48:vgpr_32 = COPY killed %98
238 successors: %bb.3(0x80000000)
239 $exec = S_OR_B64 $exec, killed %95, implicit-def $scc
240 %103:vreg_128 = COPY killed %94
241 %104:vreg_128 = COPY killed %103
242 %5:vreg_128 = COPY killed %104
246 %105:vgpr_32 = V_ADD_F32_e32 target-flags(amdgpu-rel32-lo) 0, %20.sub3, implicit $exec
247 %106:vgpr_32 = V_ADD_F32_e32 target-flags(amdgpu-gotprel32-hi) 0, killed %20.sub2, implicit $exec
248 undef %107.sub0:vreg_64 = COPY killed %106
249 %107.sub1:vreg_64 = COPY killed %105
250 $exec = S_AND_B64 $exec, killed %0, implicit-def dead $scc
251 %108:sreg_32_xm0 = S_MOV_B32 0
252 undef %109.sub0:sreg_256 = COPY %108
253 %109.sub1:sreg_256 = COPY %108
254 %109.sub2:sreg_256 = COPY %108
255 %109.sub3:sreg_256 = COPY %108
256 %109.sub4:sreg_256 = COPY %108
257 %109.sub5:sreg_256 = COPY %108
258 %109.sub6:sreg_256 = COPY %108
259 %109.sub7:sreg_256 = COPY killed %108
260 %110:vgpr_32 = IMAGE_SAMPLE_V1_V2 killed %107, killed %109, undef %111:sgpr_128, 8, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from constant-pool, addrspace 4)
261 %112:vgpr_32 = V_MUL_F32_e32 0, killed %110, implicit $exec
262 %113:vgpr_32 = V_MUL_F32_e32 0, killed %112, implicit $exec
263 %114:vgpr_32 = V_MAD_F32 0, killed %113, 0, 0, 0, 0, 0, 0, implicit $exec
264 %115:vgpr_32 = V_MAX_F32_e32 0, killed %114, implicit $exec
265 %116:vgpr_32 = V_CVT_PKRTZ_F16_F32_e64 0, killed %115, 0, 1065353216, 0, 0, implicit $exec
266 EXP 0, undef %117:vgpr_32, killed %116, undef %118:vgpr_32, undef %119:vgpr_32, -1, -1, 15, implicit $exec