2 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
3 ; RUN: opt -mtriple=amdgcn-amd-amdhsa -S -structurizecfg %s | FileCheck %s
5 ; StructurizeCFG::orderNodes used an arbitrary and nonsensical sorting
6 ; function which broke the basic backedge identification algorithm. It
7 ; would use RPO order, but then do a weird partial sort by the loop
8 ; depth assuming blocks are sorted by loop. However a block can appear
9 ; in between blocks of a loop that is not part of a loop, breaking the
10 ; assumption of the sort.
12 ; The collectInfos must be done in RPO order. The actual
13 ; structurization order I think is less important, but unless the loop
14 ; headers are identified in RPO order, it finds the wrong set of back
17 define amdgpu_kernel void @loop_backedge_misidentified(i32 addrspace(1)* %arg0) #0 {
18 ; CHECK-LABEL: @loop_backedge_misidentified(
20 ; CHECK-NEXT: [[TMP:%.*]] = load volatile <2 x i32>, <2 x i32> addrspace(1)* undef, align 16
21 ; CHECK-NEXT: [[LOAD1:%.*]] = load volatile <2 x float>, <2 x float> addrspace(1)* undef
22 ; CHECK-NEXT: [[TID:%.*]] = call i32 @llvm.amdgcn.workitem.id.x()
23 ; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds i32, i32 addrspace(1)* [[ARG0:%.*]], i32 [[TID]]
24 ; CHECK-NEXT: [[I_INITIAL:%.*]] = load volatile i32, i32 addrspace(1)* [[GEP]], align 4
25 ; CHECK-NEXT: br label [[LOOP_HEADER:%.*]]
27 ; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[I_INITIAL]], [[ENTRY:%.*]] ], [ [[TMP10:%.*]], [[FLOW4:%.*]] ]
28 ; CHECK-NEXT: call void asm sideeffect "s_nop 0x100b
29 ; CHECK-NEXT: [[TMP12:%.*]] = zext i32 [[I]] to i64
30 ; CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* null, i64 [[TMP12]]
31 ; CHECK-NEXT: [[TMP14:%.*]] = load <4 x i32>, <4 x i32> addrspace(1)* [[TMP13]], align 16
32 ; CHECK-NEXT: [[TMP15:%.*]] = extractelement <4 x i32> [[TMP14]], i64 0
33 ; CHECK-NEXT: [[TMP16:%.*]] = and i32 [[TMP15]], 65535
34 ; CHECK-NEXT: [[TMP17:%.*]] = icmp eq i32 [[TMP16]], 1
35 ; CHECK-NEXT: [[TMP0:%.*]] = xor i1 [[TMP17]], true
36 ; CHECK-NEXT: br i1 [[TMP0]], label [[BB62:%.*]], label [[FLOW:%.*]]
38 ; CHECK-NEXT: br label [[FLOW]]
40 ; CHECK-NEXT: [[TMP19:%.*]] = extractelement <2 x i32> [[TMP]], i64 0
41 ; CHECK-NEXT: [[TMP22:%.*]] = lshr i32 [[TMP19]], 16
42 ; CHECK-NEXT: [[TMP24:%.*]] = urem i32 [[TMP22]], 52
43 ; CHECK-NEXT: [[TMP25:%.*]] = mul nuw nsw i32 [[TMP24]], 52
44 ; CHECK-NEXT: br label [[INNER_LOOP:%.*]]
46 ; CHECK-NEXT: [[TMP1:%.*]] = phi i32 [ [[TMP59:%.*]], [[INNER_LOOP_BREAK:%.*]] ], [ [[TMP7:%.*]], [[FLOW]] ]
47 ; CHECK-NEXT: [[TMP2:%.*]] = phi i1 [ true, [[INNER_LOOP_BREAK]] ], [ [[TMP8:%.*]], [[FLOW]] ]
48 ; CHECK-NEXT: br i1 [[TMP2]], label [[END_ELSE_BLOCK:%.*]], label [[FLOW4]]
50 ; CHECK-NEXT: [[INNER_LOOP_J:%.*]] = phi i32 [ [[INNER_LOOP_J_INC:%.*]], [[INNER_LOOP]] ], [ [[TMP25]], [[BB18:%.*]] ]
51 ; CHECK-NEXT: call void asm sideeffect "
52 ; CHECK-NEXT: [[INNER_LOOP_J_INC]] = add nsw i32 [[INNER_LOOP_J]], 1
53 ; CHECK-NEXT: [[INNER_LOOP_CMP:%.*]] = icmp eq i32 [[INNER_LOOP_J]], 0
54 ; CHECK-NEXT: br i1 [[INNER_LOOP_CMP]], label [[INNER_LOOP_BREAK]], label [[INNER_LOOP]]
55 ; CHECK: INNER_LOOP_BREAK:
56 ; CHECK-NEXT: [[TMP59]] = extractelement <4 x i32> [[TMP14]], i64 2
57 ; CHECK-NEXT: call void asm sideeffect "s_nop 23 ", "~{memory}"() #0
58 ; CHECK-NEXT: br label [[FLOW3:%.*]]
60 ; CHECK-NEXT: [[LOAD13:%.*]] = icmp ult i32 [[TMP16]], 271
61 ; CHECK-NEXT: [[TMP3:%.*]] = xor i1 [[LOAD13]], true
62 ; CHECK-NEXT: br i1 [[TMP3]], label [[INCREMENT_I:%.*]], label [[FLOW1:%.*]]
64 ; CHECK-NEXT: [[TMP4:%.*]] = phi i32 [ [[INC_I:%.*]], [[INCREMENT_I]] ], [ undef, [[BB62]] ]
65 ; CHECK-NEXT: [[TMP5:%.*]] = phi i1 [ true, [[INCREMENT_I]] ], [ false, [[BB62]] ]
66 ; CHECK-NEXT: [[TMP6:%.*]] = phi i1 [ false, [[INCREMENT_I]] ], [ true, [[BB62]] ]
67 ; CHECK-NEXT: br i1 [[TMP6]], label [[BB64:%.*]], label [[FLOW2:%.*]]
69 ; CHECK-NEXT: call void asm sideeffect "s_nop 42", "~{memory}"() #0
70 ; CHECK-NEXT: br label [[FLOW2]]
72 ; CHECK-NEXT: [[TMP7]] = phi i32 [ [[TMP4]], [[FLOW2]] ], [ undef, [[LOOP_HEADER]] ]
73 ; CHECK-NEXT: [[TMP8]] = phi i1 [ [[TMP5]], [[FLOW2]] ], [ false, [[LOOP_HEADER]] ]
74 ; CHECK-NEXT: [[TMP9:%.*]] = phi i1 [ false, [[FLOW2]] ], [ true, [[LOOP_HEADER]] ]
75 ; CHECK-NEXT: br i1 [[TMP9]], label [[BB18]], label [[FLOW3]]
77 ; CHECK-NEXT: [[INC_I]] = add i32 [[I]], 1
78 ; CHECK-NEXT: call void asm sideeffect "s_nop 0x1336
79 ; CHECK-NEXT: br label [[FLOW1]]
80 ; CHECK: END_ELSE_BLOCK:
81 ; CHECK-NEXT: [[I_FINAL:%.*]] = phi i32 [ [[TMP1]], [[FLOW3]] ]
82 ; CHECK-NEXT: call void asm sideeffect "s_nop 0x1337
83 ; CHECK-NEXT: [[CMP_END_ELSE_BLOCK:%.*]] = icmp eq i32 [[I_FINAL]], -1
84 ; CHECK-NEXT: br label [[FLOW4]]
86 ; CHECK-NEXT: [[TMP10]] = phi i32 [ [[I_FINAL]], [[END_ELSE_BLOCK]] ], [ undef, [[FLOW3]] ]
87 ; CHECK-NEXT: [[TMP11:%.*]] = phi i1 [ [[CMP_END_ELSE_BLOCK]], [[END_ELSE_BLOCK]] ], [ true, [[FLOW3]] ]
88 ; CHECK-NEXT: br i1 [[TMP11]], label [[RETURN:%.*]], label [[LOOP_HEADER]]
90 ; CHECK-NEXT: call void asm sideeffect "s_nop 0x99
91 ; CHECK-NEXT: store volatile <2 x float> [[LOAD1]], <2 x float> addrspace(1)* undef, align 8
92 ; CHECK-NEXT: ret void
95 %tmp = load volatile <2 x i32>, <2 x i32> addrspace(1)* undef, align 16
96 %load1 = load volatile <2 x float>, <2 x float> addrspace(1)* undef
97 %tid = call i32 @llvm.amdgcn.workitem.id.x()
98 %gep = getelementptr inbounds i32, i32 addrspace(1)* %arg0, i32 %tid
99 %i.initial = load volatile i32, i32 addrspace(1)* %gep, align 4
100 br label %LOOP.HEADER
103 %i = phi i32 [ %i.final, %END_ELSE_BLOCK ], [ %i.initial, %entry ]
104 call void asm sideeffect "s_nop 0x100b ; loop $0 ", "r,~{memory}"(i32 %i) #0
105 %tmp12 = zext i32 %i to i64
106 %tmp13 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* null, i64 %tmp12
107 %tmp14 = load <4 x i32>, <4 x i32> addrspace(1)* %tmp13, align 16
108 %tmp15 = extractelement <4 x i32> %tmp14, i64 0
109 %tmp16 = and i32 %tmp15, 65535
110 %tmp17 = icmp eq i32 %tmp16, 1
111 br i1 %tmp17, label %bb18, label %bb62
114 %tmp19 = extractelement <2 x i32> %tmp, i64 0
115 %tmp22 = lshr i32 %tmp19, 16
116 %tmp24 = urem i32 %tmp22, 52
117 %tmp25 = mul nuw nsw i32 %tmp24, 52
121 %inner.loop.j = phi i32 [ %tmp25, %bb18 ], [ %inner.loop.j.inc, %INNER_LOOP ]
122 call void asm sideeffect "; inner loop body", ""() #0
123 %inner.loop.j.inc = add nsw i32 %inner.loop.j, 1
124 %inner.loop.cmp = icmp eq i32 %inner.loop.j, 0
125 br i1 %inner.loop.cmp, label %INNER_LOOP_BREAK, label %INNER_LOOP
128 %tmp59 = extractelement <4 x i32> %tmp14, i64 2
129 call void asm sideeffect "s_nop 23 ", "~{memory}"() #0
130 br label %END_ELSE_BLOCK
133 %load13 = icmp ult i32 %tmp16, 271
134 br i1 %load13, label %bb64, label %INCREMENT_I
137 call void asm sideeffect "s_nop 42", "~{memory}"() #0
141 %inc.i = add i32 %i, 1
142 call void asm sideeffect "s_nop 0x1336 ; increment $0", "v,~{memory}"(i32 %inc.i) #0
143 br label %END_ELSE_BLOCK
146 %i.final = phi i32 [ %tmp59, %INNER_LOOP_BREAK ], [ %inc.i, %INCREMENT_I ]
147 call void asm sideeffect "s_nop 0x1337 ; end else block $0", "v,~{memory}"(i32 %i.final) #0
148 %cmp.end.else.block = icmp eq i32 %i.final, -1
149 br i1 %cmp.end.else.block, label %RETURN, label %LOOP.HEADER
152 call void asm sideeffect "s_nop 0x99 ; ClosureEval return", "~{memory}"() #0
153 store volatile <2 x float> %load1, <2 x float> addrspace(1)* undef, align 8
157 ; The same function, except break to return block goes directly to the
158 ; return, which managed to hide the bug.
159 ; FIXME: Merge variant from backedge-id-bug-xfail
161 declare i32 @llvm.amdgcn.workitem.id.x() #1
163 attributes #0 = { convergent nounwind }
164 attributes #1 = { convergent nounwind readnone }