Closed nunoplopes closed 2 years ago
SLPVectorizer incorrectly replaces add nsw undef with add poison. However, add nsw undef, 0 is undef, not poison, so the transformation is not correct. Example:
add nsw undef
add poison
add nsw undef, 0
; Test: Transforms/SLPVectorizer/X86/vectorize-reorder-alt-shuffle.ll define void @foo(ptr %c, ptr %d) { %entry: %arrayidx1 = gep inbounds ptr %c, 1 x i64 4 %0 = load i8, ptr %arrayidx1, align 1 %conv2 = zext i8 %0 to i32 %and = and i32 %conv2, 3 %arrayidx4 = gep inbounds ptr %c, 1 x i64 1 %1 = load i8, ptr %arrayidx4, align 1 %conv5 = zext i8 %1 to i32 %shl6 = shl nsw nuw i32 %conv5, 2 %arrayidx12 = gep inbounds ptr %c, 1 x i64 2 %2 = load i8, ptr %arrayidx12, align 1 %conv13 = zext i8 %2 to i32 %shl14 = shl nsw nuw i32 %conv13, 2 %arrayidx17 = gep inbounds ptr %c, 1 x i64 3 %3 = load i8, ptr %arrayidx17, align 1 %conv18 = zext i8 %3 to i32 %shl19 = shl nsw nuw i32 %conv18, 2 %sub = add nsw i32 undef, %shl6 %conv27 = sitofp i32 %sub to float, exceptions=ignore %div = fdiv float %conv27, undef, exceptions=ignore %add.ptr = gep inbounds ptr %d, 4 x i64 -1 store float %div, ptr %add.ptr, align 4 %sub32 = add nsw i32 undef, %and %conv33 = sitofp i32 %sub32 to float, exceptions=ignore %div36 = fdiv float %conv33, undef, exceptions=ignore %add.ptr37 = gep inbounds ptr %d, 4 x i64 -2 store float %div36, ptr %add.ptr37, align 4 ; HERE: add nsw undef %sub40 = add nsw i32 undef, %shl19 %conv41 = sitofp i32 %sub40 to float, exceptions=ignore %div44 = fdiv float %conv41, undef, exceptions=ignore %add.ptr45 = gep inbounds ptr %d, 4 x i64 -3 store float %div44, ptr %add.ptr45, align 4 %sub48 = add nsw i32 undef, %shl14 %conv49 = sitofp i32 %sub48 to float, exceptions=ignore %div52 = fdiv float %conv49, undef, exceptions=ignore %add.ptr53 = gep inbounds ptr %d, 4 x i64 -4 store float %div52, ptr %add.ptr53, align 4 ret void } => define void @foo(ptr %c, ptr %d) { %entry: %arrayidx4 = gep inbounds ptr %c, 1 x i64 1 %add.ptr53 = gep inbounds ptr %d, 4 x i64 -4 %0 = bitcast ptr %arrayidx4 to ptr %1 = load <4 x i8>, ptr %0, align 1 %2 = zext <4 x i8> %1 to <4 x i32> %3 = shl nsw nuw <4 x i32> %2, { 2, 2, 2, 3 } %4 = and <4 x i32> %2, { 2, 2, 2, 3 } %5 = shufflevector <4 x i32> %3, <4 x i32> %4, 1, 2, 7, 0 ; And here we have add poison %6 = add nsw <4 x i32> poison, %5 %7 = sitofp <4 x i32> %6 to <4 x float>, exceptions=ignore %8 = fdiv <4 x float> %7, poison, exceptions=ignore %9 = bitcast ptr %add.ptr53 to ptr store <4 x float> %8, ptr %9, align 4 ret void } Transformation doesn't verify! ERROR: Mismatch in memory Example: ptr %c = pointer(non-local, block_id=1, offset=2) ptr %d = pointer(non-local, block_id=1, offset=32) Source: ptr %arrayidx1 = pointer(non-local, block_id=1, offset=6) i8 %0 = #x00 (0) i32 %conv2 = #x00000000 (0) i32 %and = #x00000000 (0) ptr %arrayidx4 = pointer(non-local, block_id=1, offset=3) i8 %1 = poison i32 %conv5 = poison i32 %shl6 = poison ptr %arrayidx12 = pointer(non-local, block_id=1, offset=4) i8 %2 = poison i32 %conv13 = poison i32 %shl14 = poison ptr %arrayidx17 = pointer(non-local, block_id=1, offset=5) i8 %3 = #x00 (0) i32 %conv18 = #x00000000 (0) i32 %shl19 = #x00000000 (0) i32 %sub = poison float %conv27 = poison float %div = poison ptr %add.ptr = pointer(non-local, block_id=1, offset=28) i32 %sub32 = #x00000000 (0) [based on undef value] float %conv33 = #x00000000 (+0.0) float %div36 = NaN [based on undef value] ptr %add.ptr37 = pointer(non-local, block_id=1, offset=24) i32 %sub40 = #x00000000 (0) [based on undef value] float %conv41 = #x00000000 (+0.0) float %div44 = NaN [based on undef value] ptr %add.ptr45 = pointer(non-local, block_id=1, offset=20) i32 %sub48 = poison float %conv49 = poison float %div52 = poison ptr %add.ptr53 = pointer(non-local, block_id=1, offset=16) SOURCE MEMORY STATE =================== NON-LOCAL BLOCKS: Block 0 > size: 0 align: 1 alloc type: 0 Block 1 > size: 135 align: 256 alloc type: 0 Block 2 > size: 254 align: 65536 alloc type: 0 Target: ptr %arrayidx4 = pointer(non-local, block_id=1, offset=3) ptr %add.ptr53 = pointer(non-local, block_id=1, offset=16) ptr %0 = pointer(non-local, block_id=1, offset=3) <4 x i8> %1 = < poison, poison, #x00 (0), #x00 (0) > <4 x i32> %2 = < poison, poison, #x00000000 (0), #x00000000 (0) > <4 x i32> %3 = < poison, poison, #x00000000 (0), #x00000000 (0) > <4 x i32> %4 = < poison, poison, #x00000000 (0), #x00000000 (0) > <4 x i32> %5 = < poison, #x00000000 (0), #x00000000 (0), poison > <4 x i32> %6 = < poison, poison, poison, poison > <4 x float> %7 = < poison, poison, poison, poison > <4 x float> %8 = < poison, poison, poison, poison > ptr %9 = pointer(non-local, block_id=1, offset=16) Mismatch in pointer(non-local, block_id=1, offset=22) Source value: #x80 Target value: poison
cc @regehr @cilkplus @davemgreen @RKSimon
Also cc @alexey-bataev
Fixed in 120d52b0ef8b0d41d5ce5898422d2d2f42047d0b
SLPVectorizer incorrectly replaces
add nsw undef
withadd poison
. However,add nsw undef, 0
is undef, not poison, so the transformation is not correct. Example:cc @regehr @cilkplus @davemgreen @RKSimon