|
| 1 | +; NOTE: Assertions have been autogenerated by utils/update_test_checks.py |
| 2 | +; RUN: opt < %s -slp-vectorizer -mtriple=x86_64-grtev4-linux-gnu -S | FileCheck %s |
| 3 | + |
| 4 | +; This checks that reorderBottomToTop() can handle reordering of a TreeEntry |
| 5 | +; which has a user TreeEntry that has already been reordered. |
| 6 | +; Here is when the crash occurs: |
| 7 | +; |
| 8 | +; (N4)OrderB |
| 9 | +; | |
| 10 | +; (N1)OrderA (N2)OrderA (N3)NoOrder |
| 11 | +; \ | / |
| 12 | +; (Phi)NoOrder |
| 13 | +; |
| 14 | +; 1. Phi is visited along with its operands (N1,N2,N3). BestOrder is "OrderA". |
| 15 | +; 2. Phi along with all its operands (N1,N2,N3) are reordered. The result is: |
| 16 | +; |
| 17 | +; (N4)OrderB |
| 18 | +; | |
| 19 | +; (N1)NoOrder (N2)NoOrder (N3)OrderA |
| 20 | +; \ | / |
| 21 | +; (Phi)OrderA |
| 22 | +; |
| 23 | +; 3. N3 is now visited along with its operand N4. BestOrder is "OrderB". |
| 24 | +; 4. N3 and N4 are reordered. The result is this: |
| 25 | +; |
| 26 | +; (N4)NoOrder |
| 27 | +; | |
| 28 | +; (N1)NoOrder (N2)NoOrder (N3)OrderB |
| 29 | +; \ | / |
| 30 | +; (Phi)OrderA |
| 31 | +; |
| 32 | +; At this point there is a discrepancy between Phi's Operand 2 which are |
| 33 | +; reordered based on OrderA and N3's OrderB. This results in a crash in |
| 34 | +; vectorizeTree() on its way from N3 back to the Phi. The reason is that |
| 35 | +; N3->isSame(Phi's operand 2) returns false and vectorizeTree() skips N3. |
| 36 | +; |
| 37 | +; This patch changes the order in which the nodes are visited to bottom-up, |
| 38 | +; which fixes the issue. |
| 39 | +; |
| 40 | +; NOTE: The crash shows up when reorderTopToBottom() does not reorder the tree, |
| 41 | +; so to simulate this we add external store users. Alternatively one can |
| 42 | +; comment out reorderTopToBottom() and remove the stores. |
| 43 | + |
| 44 | + |
| 45 | +define void @reorder_crash(float* %ptr) { |
| 46 | +; CHECK-LABEL: @reorder_crash( |
| 47 | +; CHECK-NEXT: entry: |
| 48 | +; CHECK-NEXT: [[GEP0:%.*]] = getelementptr inbounds float, float* [[PTR:%.*]], i64 0 |
| 49 | +; CHECK-NEXT: br i1 undef, label [[BB0:%.*]], label [[BB12:%.*]] |
| 50 | +; CHECK: bb0: |
| 51 | +; CHECK-NEXT: [[TMP0:%.*]] = bitcast float* [[GEP0]] to <4 x float>* |
| 52 | +; CHECK-NEXT: [[TMP1:%.*]] = load <4 x float>, <4 x float>* [[TMP0]], align 4 |
| 53 | +; CHECK-NEXT: [[TMP2:%.*]] = bitcast float* [[GEP0]] to <4 x float>* |
| 54 | +; CHECK-NEXT: store <4 x float> [[TMP1]], <4 x float>* [[TMP2]], align 4 |
| 55 | +; CHECK-NEXT: br label [[BB3:%.*]] |
| 56 | +; CHECK: bb12: |
| 57 | +; CHECK-NEXT: br i1 undef, label [[BB1:%.*]], label [[BB2:%.*]] |
| 58 | +; CHECK: bb1: |
| 59 | +; CHECK-NEXT: [[TMP3:%.*]] = bitcast float* [[GEP0]] to <4 x float>* |
| 60 | +; CHECK-NEXT: [[TMP4:%.*]] = load <4 x float>, <4 x float>* [[TMP3]], align 4 |
| 61 | +; CHECK-NEXT: [[TMP5:%.*]] = bitcast float* [[GEP0]] to <4 x float>* |
| 62 | +; CHECK-NEXT: store <4 x float> [[TMP4]], <4 x float>* [[TMP5]], align 4 |
| 63 | +; CHECK-NEXT: br label [[BB3]] |
| 64 | +; CHECK: bb2: |
| 65 | +; CHECK-NEXT: [[TMP6:%.*]] = bitcast float* [[GEP0]] to <4 x float>* |
| 66 | +; CHECK-NEXT: [[TMP7:%.*]] = load <4 x float>, <4 x float>* [[TMP6]], align 4 |
| 67 | +; CHECK-NEXT: [[TMP8:%.*]] = fadd <4 x float> [[TMP7]], zeroinitializer |
| 68 | +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x float> [[TMP8]], <4 x float> poison, <4 x i32> <i32 3, i32 2, i32 0, i32 1> |
| 69 | +; CHECK-NEXT: br label [[BB3]] |
| 70 | +; CHECK: bb3: |
| 71 | +; CHECK-NEXT: [[TMP9:%.*]] = phi <4 x float> [ [[TMP1]], [[BB0]] ], [ [[TMP4]], [[BB1]] ], [ [[SHUFFLE]], [[BB2]] ] |
| 72 | +; CHECK-NEXT: ret void |
| 73 | +; |
| 74 | +entry: |
| 75 | + %gep0 = getelementptr inbounds float, float* %ptr, i64 0 |
| 76 | + %gep1 = getelementptr inbounds float, float* %ptr, i64 1 |
| 77 | + %gep2 = getelementptr inbounds float, float* %ptr, i64 2 |
| 78 | + %gep3 = getelementptr inbounds float, float* %ptr, i64 3 |
| 79 | + br i1 undef, label %bb0, label %bb12 |
| 80 | + |
| 81 | +bb0: |
| 82 | + ; Used by phi in this order: 1, 0, 2, 3 |
| 83 | + %ld00 = load float, float* %gep0 |
| 84 | + %ld01 = load float, float* %gep1 |
| 85 | + %ld02 = load float, float* %gep2 |
| 86 | + %ld03 = load float, float* %gep3 |
| 87 | + |
| 88 | + ; External store users in natural order 0, 1, 2, 3 |
| 89 | + store float %ld00, float *%gep0 |
| 90 | + store float %ld01, float *%gep1 |
| 91 | + store float %ld02, float *%gep2 |
| 92 | + store float %ld03, float *%gep3 |
| 93 | + br label %bb3 |
| 94 | + |
| 95 | +bb12: |
| 96 | + br i1 undef, label %bb1, label %bb2 |
| 97 | + |
| 98 | +bb1: |
| 99 | + ; Used by phi in this order: 1, 0, 2, 3 |
| 100 | + %ld10 = load float, float* %gep0 |
| 101 | + %ld11 = load float, float* %gep1 |
| 102 | + %ld12 = load float, float* %gep2 |
| 103 | + %ld13 = load float, float* %gep3 |
| 104 | + |
| 105 | + ; External store users in natural order 0, 1, 2, 3 |
| 106 | + store float %ld10, float *%gep0 |
| 107 | + store float %ld11, float *%gep1 |
| 108 | + store float %ld12, float *%gep2 |
| 109 | + store float %ld13, float *%gep3 |
| 110 | + |
| 111 | + br label %bb3 |
| 112 | + |
| 113 | +bb2: |
| 114 | + ; Used by fadd in this order: 2, 3, 0, 1 |
| 115 | + %ld20 = load float, float* %gep0 |
| 116 | + %ld21 = load float, float* %gep1 |
| 117 | + %ld22 = load float, float* %gep2 |
| 118 | + %ld23 = load float, float* %gep3 |
| 119 | + |
| 120 | + ; Used by phi in this order: 0, 1, 2, 3 |
| 121 | + %add20 = fadd float %ld22, 0.0 |
| 122 | + %add21 = fadd float %ld23, 0.0 |
| 123 | + %add22 = fadd float %ld20, 0.0 |
| 124 | + %add23 = fadd float %ld21, 0.0 |
| 125 | + br label %bb3 |
| 126 | + |
| 127 | +bb3: |
| 128 | + %phi0 = phi float [ %ld01, %bb0 ], [ %ld11, %bb1 ], [ %add20, %bb2 ] |
| 129 | + %phi1 = phi float [ %ld00, %bb0 ], [ %ld10, %bb1 ], [ %add21, %bb2 ] |
| 130 | + %phi2 = phi float [ %ld02, %bb0 ], [ %ld12, %bb1 ], [ %add22, %bb2 ] |
| 131 | + %phi3 = phi float [ %ld03, %bb0 ], [ %ld13, %bb1 ], [ %add23, %bb2 ] |
| 132 | + ret void |
| 133 | +} |
0 commit comments