|
| 1 | +; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 3 |
| 2 | +; RUN: opt -passes=slp-vectorizer -mtriple=x86_64 -S -mcpu=core-avx2 -pass-remarks-output=%t < %s | FileCheck %s |
| 3 | +; RUN: FileCheck --input-file=%t --check-prefix=YAML %s |
| 4 | + |
| 5 | +; YAML: --- !Passed |
| 6 | +; YAML: Pass: slp-vectorizer |
| 7 | +; YAML: Name: StoresVectorized |
| 8 | +; YAML: Function: test |
| 9 | +; YAML: Args: |
| 10 | +; YAML: - String: 'Stores SLP vectorized with cost ' |
| 11 | +; YAML: - Cost: '-3' |
| 12 | +; YAML: - String: ' and with tree size ' |
| 13 | +; YAML: - TreeSize: '14' |
| 14 | +; YAML: ... |
| 15 | + |
| 16 | +; Test that SLP cost modeling is able to match gathering tree |
| 17 | +; entries: two instances of { call0 .. call3 }. Both used by |
| 18 | +; vectorized PHI entries: { phi0 .. phi3 } and { phi4 .. phi7 }. |
| 19 | +; Check that cost is not added twice for gathered calls. |
| 20 | + |
| 21 | +define void @test(ptr %dst, float %a, float %b, float %c, float %d) { |
| 22 | +; CHECK-LABEL: define void @test( |
| 23 | +; CHECK-SAME: ptr [[DST:%.*]], float [[A:%.*]], float [[B:%.*]], float [[C:%.*]], float [[D:%.*]]) #[[ATTR0:[0-9]+]] { |
| 24 | +; CHECK-NEXT: Entry: |
| 25 | +; CHECK-NEXT: br i1 poison, label [[LOOP0:%.*]], label [[EXIT:%.*]] |
| 26 | +; CHECK: loop0: |
| 27 | +; CHECK-NEXT: [[TMP0:%.*]] = phi <4 x float> [ <float 5.000000e-01, float 5.000000e-01, float 5.000000e-01, float 5.000000e-01>, [[ENTRY:%.*]] ], [ [[TMP7:%.*]], [[USERBLOCK1:%.*]] ] |
| 28 | +; CHECK-NEXT: br i1 poison, label [[USERBLOCK0:%.*]], label [[BLKX:%.*]] |
| 29 | +; CHECK: UserBlock0: |
| 30 | +; CHECK-NEXT: [[TMP1:%.*]] = phi <4 x float> [ zeroinitializer, [[LOOP0]] ], [ [[TMP5:%.*]], [[BLKX]] ] |
| 31 | +; CHECK-NEXT: br label [[LOOP_INNER:%.*]] |
| 32 | +; CHECK: blkx: |
| 33 | +; CHECK-NEXT: [[CALL0:%.*]] = call fast float @foo(float [[A]]) |
| 34 | +; CHECK-NEXT: [[CALL1:%.*]] = call fast float @foo(float [[B]]) |
| 35 | +; CHECK-NEXT: [[CALL2:%.*]] = call fast float @foo(float [[C]]) |
| 36 | +; CHECK-NEXT: [[CALL3:%.*]] = call fast float @foo(float [[D]]) |
| 37 | +; CHECK-NEXT: [[TMP2:%.*]] = insertelement <4 x float> poison, float [[CALL0]], i32 0 |
| 38 | +; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x float> [[TMP2]], float [[CALL1]], i32 1 |
| 39 | +; CHECK-NEXT: [[TMP4:%.*]] = insertelement <4 x float> [[TMP3]], float [[CALL2]], i32 2 |
| 40 | +; CHECK-NEXT: [[TMP5]] = insertelement <4 x float> [[TMP4]], float [[CALL3]], i32 3 |
| 41 | +; CHECK-NEXT: br i1 poison, label [[USERBLOCK0]], label [[USERBLOCK1]] |
| 42 | +; CHECK: UserBlock1: |
| 43 | +; CHECK-NEXT: [[TMP6:%.*]] = phi <4 x float> [ [[TMP5]], [[BLKX]] ], [ [[TMP9:%.*]], [[LOOP_INNER]] ] |
| 44 | +; CHECK-NEXT: [[TMP7]] = fadd fast <4 x float> [[TMP0]], [[TMP6]] |
| 45 | +; CHECK-NEXT: br i1 poison, label [[EXIT]], label [[LOOP0]] |
| 46 | +; CHECK: loop.inner: |
| 47 | +; CHECK-NEXT: [[TMP8:%.*]] = phi <4 x float> [ [[TMP1]], [[USERBLOCK0]] ], [ [[TMP9]], [[LOOP_INNER]] ] |
| 48 | +; CHECK-NEXT: [[TMP9]] = fadd fast <4 x float> [[TMP8]], poison |
| 49 | +; CHECK-NEXT: br i1 poison, label [[USERBLOCK1]], label [[LOOP_INNER]] |
| 50 | +; CHECK: Exit: |
| 51 | +; CHECK-NEXT: [[TMP10:%.*]] = phi <4 x float> [ <float 5.000000e-01, float 5.000000e-01, float 5.000000e-01, float 5.000000e-01>, [[ENTRY]] ], [ [[TMP7]], [[USERBLOCK1]] ] |
| 52 | +; CHECK-NEXT: [[IDX0:%.*]] = add i64 0, poison |
| 53 | +; CHECK-NEXT: [[GEP0:%.*]] = getelementptr inbounds float, ptr [[DST]], i64 [[IDX0]] |
| 54 | +; CHECK-NEXT: store <4 x float> [[TMP10]], ptr [[GEP0]], align 4 |
| 55 | +; CHECK-NEXT: ret void |
| 56 | +; |
| 57 | +Entry: |
| 58 | + br i1 poison, label %loop0, label %Exit |
| 59 | + |
| 60 | +loop0: |
| 61 | + %i0 = phi float [ 5.000000e-01, %Entry ], [ %add3, %UserBlock1 ] |
| 62 | + %i1 = phi float [ 5.000000e-01, %Entry ], [ %add2, %UserBlock1 ] |
| 63 | + %i2 = phi float [ 5.000000e-01, %Entry ], [ %add1, %UserBlock1 ] |
| 64 | + %i3 = phi float [ 5.000000e-01, %Entry ], [ %add0, %UserBlock1 ] |
| 65 | + br i1 poison, label %UserBlock0, label %blkx |
| 66 | + |
| 67 | +UserBlock0: |
| 68 | + %phi0 = phi float [ 0.000000e+00, %loop0 ], [ %call3, %blkx ] |
| 69 | + %phi1 = phi float [ 0.000000e+00, %loop0 ], [ %call2, %blkx ] |
| 70 | + %phi2 = phi float [ 0.000000e+00, %loop0 ], [ %call1, %blkx ] |
| 71 | + %phi3 = phi float [ 0.000000e+00, %loop0 ], [ %call0, %blkx ] |
| 72 | + br label %loop.inner |
| 73 | + |
| 74 | +blkx: |
| 75 | + %call0 = call fast float @foo(float %a) |
| 76 | + %call1 = call fast float @foo(float %b) |
| 77 | + %call2 = call fast float @foo(float %c) |
| 78 | + %call3 = call fast float @foo(float %d) |
| 79 | + br i1 poison, label %UserBlock0, label %UserBlock1 |
| 80 | + |
| 81 | +UserBlock1: |
| 82 | + %phi4 = phi float [ %call0, %blkx ], [ %add4, %loop.inner ] |
| 83 | + %phi5 = phi float [ %call1, %blkx ], [ %add5, %loop.inner ] |
| 84 | + %phi6 = phi float [ %call2, %blkx ], [ %add6, %loop.inner ] |
| 85 | + %phi7 = phi float [ %call3, %blkx ], [ %add7, %loop.inner ] |
| 86 | + %add0 = fadd fast float %i3, %phi4 |
| 87 | + %add1 = fadd fast float %i2, %phi5 |
| 88 | + %add2 = fadd fast float %i1, %phi6 |
| 89 | + %add3 = fadd fast float %i0, %phi7 |
| 90 | + br i1 poison, label %Exit, label %loop0 |
| 91 | + |
| 92 | +loop.inner: |
| 93 | + %t20 = phi float [ %phi3, %UserBlock0 ], [ %add4, %loop.inner ] |
| 94 | + %t19 = phi float [ %phi2, %UserBlock0 ], [ %add5, %loop.inner ] |
| 95 | + %t18 = phi float [ %phi1, %UserBlock0 ], [ %add6, %loop.inner ] |
| 96 | + %t17 = phi float [ %phi0, %UserBlock0 ], [ %add7, %loop.inner ] |
| 97 | + %add4 = fadd fast float %t20, poison |
| 98 | + %add5 = fadd fast float %t19, poison |
| 99 | + %add6 = fadd fast float %t18, poison |
| 100 | + %add7 = fadd fast float %t17, poison |
| 101 | + |
| 102 | + br i1 poison, label %UserBlock1, label %loop.inner |
| 103 | + |
| 104 | +Exit: |
| 105 | + %t48 = phi float [ 5.000000e-01, %Entry ], [ %add3, %UserBlock1 ] |
| 106 | + %t47 = phi float [ 5.000000e-01, %Entry ], [ %add2, %UserBlock1 ] |
| 107 | + %t46 = phi float [ 5.000000e-01, %Entry ], [ %add1, %UserBlock1 ] |
| 108 | + %t45 = phi float [ 5.000000e-01, %Entry ], [ %add0, %UserBlock1 ] |
| 109 | + %idx0 = add i64 0, poison |
| 110 | + %gep0 = getelementptr inbounds float, ptr %dst, i64 %idx0 |
| 111 | + store float %t45, ptr %gep0, align 4 |
| 112 | + %idx1 = add i64 1, poison |
| 113 | + %gep1 = getelementptr inbounds float, ptr %dst, i64 %idx1 |
| 114 | + store float %t46, ptr %gep1, align 4 |
| 115 | + %idx2 = add i64 2, poison |
| 116 | + %gep2 = getelementptr inbounds float, ptr %dst, i64 %idx2 |
| 117 | + store float %t47, ptr %gep2, align 4 |
| 118 | + %idx3 = add i64 3, poison |
| 119 | + %gep3 = getelementptr inbounds float, ptr %dst, i64 %idx3 |
| 120 | + store float %t48, ptr %gep3, align 4 |
| 121 | + ret void |
| 122 | +} |
| 123 | + |
| 124 | +declare float @foo(float) |
0 commit comments