1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2 ; RUN: opt -mtriple=amdgcn-amd-amdhsa -S -structurizecfg %s | FileCheck %s
4 ; StructurizeCFG::orderNodes basically uses a reverse post-order (RPO) traversal of the region
5 ; list to get the order. The only problem with it is that sometimes backedges
6 ; for outer loops will be visited before backedges for inner loops. To solve this problem,
7 ; a loop depth based approach has been used to make sure all blocks in this loop has been visited
8 ; before moving on to outer loop.
10 ; However, we found a problem for a SubRegion which is a loop itself:
14 ; --> BB1 --> BB2 --> BB3 -->
16 ; In this case, BB2 is a SubRegion (loop), and thus its loopdepth is different than that of
17 ; BB1 and BB3. This fact will lead BB2 to be placed in the wrong order.
19 ; In this work, we treat the SubRegion as a special case and use its exit block to determine
20 ; the loop and its depth to guard the sorting.
21 define amdgpu_kernel void @loop_subregion_misordered(ptr addrspace(1) %arg0) #0 {
22 ; CHECK-LABEL: @loop_subregion_misordered(
24 ; CHECK-NEXT: [[TMP:%.*]] = load volatile <2 x i32>, ptr addrspace(1) undef, align 16
25 ; CHECK-NEXT: [[LOAD1:%.*]] = load volatile <2 x float>, ptr addrspace(1) undef, align 8
26 ; CHECK-NEXT: [[TID:%.*]] = call i32 @llvm.amdgcn.workitem.id.x()
27 ; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds i32, ptr addrspace(1) [[ARG0:%.*]], i32 [[TID]]
28 ; CHECK-NEXT: [[I_INITIAL:%.*]] = load volatile i32, ptr addrspace(1) [[GEP]], align 4
29 ; CHECK-NEXT: br label [[LOOP_HEADER:%.*]]
31 ; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[I_INITIAL]], [[ENTRY:%.*]] ], [ [[TMP5:%.*]], [[FLOW3:%.*]] ]
32 ; CHECK-NEXT: call void asm sideeffect "s_nop 0x100b
33 ; CHECK-NEXT: [[TMP12:%.*]] = zext i32 [[I]] to i64
34 ; CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds <4 x i32>, ptr addrspace(1) null, i64 [[TMP12]]
35 ; CHECK-NEXT: [[TMP14:%.*]] = load <4 x i32>, ptr addrspace(1) [[TMP13]], align 16
36 ; CHECK-NEXT: [[TMP15:%.*]] = extractelement <4 x i32> [[TMP14]], i64 0
37 ; CHECK-NEXT: [[TMP16:%.*]] = and i32 [[TMP15]], 65535
38 ; CHECK-NEXT: [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 1
39 ; CHECK-NEXT: br i1 [[TMP17]], label [[BB62:%.*]], label [[FLOW:%.*]]
41 ; CHECK-NEXT: [[TMP0:%.*]] = phi i32 [ [[INC_I:%.*]], [[INCREMENT_I:%.*]] ], [ undef, [[BB62]] ]
42 ; CHECK-NEXT: [[TMP1:%.*]] = phi i1 [ false, [[INCREMENT_I]] ], [ true, [[BB62]] ]
43 ; CHECK-NEXT: [[TMP2:%.*]] = phi i1 [ true, [[INCREMENT_I]] ], [ false, [[BB62]] ]
44 ; CHECK-NEXT: br label [[FLOW]]
46 ; CHECK-NEXT: [[TMP19:%.*]] = extractelement <2 x i32> [[TMP]], i64 0
47 ; CHECK-NEXT: [[TMP22:%.*]] = lshr i32 [[TMP19]], 16
48 ; CHECK-NEXT: [[TMP24:%.*]] = urem i32 [[TMP22]], 52
49 ; CHECK-NEXT: [[TMP25:%.*]] = mul nuw nsw i32 [[TMP24]], 52
50 ; CHECK-NEXT: br label [[INNER_LOOP:%.*]]
52 ; CHECK-NEXT: [[TMP3:%.*]] = phi i32 [ [[TMP59:%.*]], [[INNER_LOOP_BREAK:%.*]] ], [ [[TMP7:%.*]], [[FLOW]] ]
53 ; CHECK-NEXT: [[TMP4:%.*]] = phi i1 [ true, [[INNER_LOOP_BREAK]] ], [ [[TMP9:%.*]], [[FLOW]] ]
54 ; CHECK-NEXT: br i1 [[TMP4]], label [[END_ELSE_BLOCK:%.*]], label [[FLOW3]]
56 ; CHECK-NEXT: [[INNER_LOOP_J:%.*]] = phi i32 [ [[INNER_LOOP_J_INC:%.*]], [[INNER_LOOP]] ], [ [[TMP25]], [[BB18:%.*]] ]
57 ; CHECK-NEXT: call void asm sideeffect "
58 ; CHECK-NEXT: [[INNER_LOOP_J_INC]] = add nsw i32 [[INNER_LOOP_J]], 1
59 ; CHECK-NEXT: [[INNER_LOOP_CMP:%.*]] = icmp eq i32 [[INNER_LOOP_J]], 0
60 ; CHECK-NEXT: br i1 [[INNER_LOOP_CMP]], label [[INNER_LOOP_BREAK]], label [[INNER_LOOP]]
61 ; CHECK: INNER_LOOP_BREAK:
62 ; CHECK-NEXT: [[TMP59]] = extractelement <4 x i32> [[TMP14]], i64 2
63 ; CHECK-NEXT: call void asm sideeffect "s_nop 23 ", "~{memory}"() #[[ATTR0:[0-9]+]]
64 ; CHECK-NEXT: br label [[FLOW2:%.*]]
66 ; CHECK-NEXT: [[LOAD13:%.*]] = icmp uge i32 [[TMP16]], 271
67 ; CHECK-NEXT: br i1 [[LOAD13]], label [[INCREMENT_I]], label [[FLOW1:%.*]]
69 ; CHECK-NEXT: [[TMP5]] = phi i32 [ [[TMP3]], [[END_ELSE_BLOCK]] ], [ undef, [[FLOW2]] ]
70 ; CHECK-NEXT: [[TMP6:%.*]] = phi i1 [ [[CMP_END_ELSE_BLOCK:%.*]], [[END_ELSE_BLOCK]] ], [ true, [[FLOW2]] ]
71 ; CHECK-NEXT: br i1 [[TMP6]], label [[FLOW4:%.*]], label [[LOOP_HEADER]]
73 ; CHECK-NEXT: br i1 [[TMP8:%.*]], label [[BB64:%.*]], label [[RETURN:%.*]]
75 ; CHECK-NEXT: call void asm sideeffect "s_nop 42", "~{memory}"() #[[ATTR0]]
76 ; CHECK-NEXT: br label [[RETURN]]
78 ; CHECK-NEXT: [[TMP7]] = phi i32 [ [[TMP0]], [[FLOW1]] ], [ undef, [[LOOP_HEADER]] ]
79 ; CHECK-NEXT: [[TMP8]] = phi i1 [ [[TMP1]], [[FLOW1]] ], [ false, [[LOOP_HEADER]] ]
80 ; CHECK-NEXT: [[TMP9]] = phi i1 [ [[TMP2]], [[FLOW1]] ], [ false, [[LOOP_HEADER]] ]
81 ; CHECK-NEXT: [[TMP10:%.*]] = phi i1 [ false, [[FLOW1]] ], [ true, [[LOOP_HEADER]] ]
82 ; CHECK-NEXT: br i1 [[TMP10]], label [[BB18]], label [[FLOW2]]
84 ; CHECK-NEXT: [[INC_I]] = add i32 [[I]], 1
85 ; CHECK-NEXT: call void asm sideeffect "s_nop 0x1336
86 ; CHECK-NEXT: br label [[FLOW1]]
87 ; CHECK: END_ELSE_BLOCK:
88 ; CHECK-NEXT: call void asm sideeffect "s_nop 0x1337
89 ; CHECK-NEXT: [[CMP_END_ELSE_BLOCK]] = icmp eq i32 [[TMP3]], -1
90 ; CHECK-NEXT: br label [[FLOW3]]
92 ; CHECK-NEXT: call void asm sideeffect "s_nop 0x99
93 ; CHECK-NEXT: store volatile <2 x float> [[LOAD1]], ptr addrspace(1) undef, align 8
94 ; CHECK-NEXT: ret void
97 %tmp = load volatile <2 x i32>, ptr addrspace(1) undef, align 16
98 %load1 = load volatile <2 x float>, ptr addrspace(1) undef
99 %tid = call i32 @llvm.amdgcn.workitem.id.x()
100 %gep = getelementptr inbounds i32, ptr addrspace(1) %arg0, i32 %tid
101 %i.initial = load volatile i32, ptr addrspace(1) %gep, align 4
102 br label %LOOP.HEADER
105 %i = phi i32 [ %i.final, %END_ELSE_BLOCK ], [ %i.initial, %entry ]
106 call void asm sideeffect "s_nop 0x100b ; loop $0 ", "r,~{memory}"(i32 %i) #0
107 %tmp12 = zext i32 %i to i64
108 %tmp13 = getelementptr inbounds <4 x i32>, ptr addrspace(1) null, i64 %tmp12
109 %tmp14 = load <4 x i32>, ptr addrspace(1) %tmp13, align 16
110 %tmp15 = extractelement <4 x i32> %tmp14, i64 0
111 %tmp16 = and i32 %tmp15, 65535
112 %tmp17 = icmp eq i32 %tmp16, 1
113 br i1 %tmp17, label %bb18, label %bb62
116 %tmp19 = extractelement <2 x i32> %tmp, i64 0
117 %tmp22 = lshr i32 %tmp19, 16
118 %tmp24 = urem i32 %tmp22, 52
119 %tmp25 = mul nuw nsw i32 %tmp24, 52
123 %inner.loop.j = phi i32 [ %tmp25, %bb18 ], [ %inner.loop.j.inc, %INNER_LOOP ]
124 call void asm sideeffect "; inner loop body", ""() #0
125 %inner.loop.j.inc = add nsw i32 %inner.loop.j, 1
126 %inner.loop.cmp = icmp eq i32 %inner.loop.j, 0
127 br i1 %inner.loop.cmp, label %INNER_LOOP_BREAK, label %INNER_LOOP
130 %tmp59 = extractelement <4 x i32> %tmp14, i64 2
131 call void asm sideeffect "s_nop 23 ", "~{memory}"() #0
132 br label %END_ELSE_BLOCK
135 %load13 = icmp ult i32 %tmp16, 271
136 br i1 %load13, label %bb64, label %INCREMENT_I
139 call void asm sideeffect "s_nop 42", "~{memory}"() #0
143 %inc.i = add i32 %i, 1
144 call void asm sideeffect "s_nop 0x1336 ; increment $0", "v,~{memory}"(i32 %inc.i) #0
145 br label %END_ELSE_BLOCK
148 %i.final = phi i32 [ %tmp59, %INNER_LOOP_BREAK ], [ %inc.i, %INCREMENT_I ]
149 call void asm sideeffect "s_nop 0x1337 ; end else block $0", "v,~{memory}"(i32 %i.final) #0
150 %cmp.end.else.block = icmp eq i32 %i.final, -1
151 br i1 %cmp.end.else.block, label %RETURN, label %LOOP.HEADER
154 call void asm sideeffect "s_nop 0x99 ; ClosureEval return", "~{memory}"() #0
155 store volatile <2 x float> %load1, ptr addrspace(1) undef, align 8
159 declare i32 @llvm.amdgcn.workitem.id.x() #1
161 attributes #0 = { convergent nounwind }
162 attributes #1 = { convergent nounwind readnone }