Skip to content

Commit

Permalink
[InstCombine] refine UB-handling in shuffle-binop transform
Browse files Browse the repository at this point in the history 
  
As noted in rL333782, we can be both better for optimization and
safer with this transform:
BinOp (shuffle V1, Mask), C --> shuffle (BinOp V1, NewC), Mask

The only potentially unsafe-to-speculate binops are integer div/rem.
All other binops are always safe (although I don't see a way to
assert that in code here).

For opcodes like shifts that can produce poison, it can't matter
here because we know the lanes with undef are dropped by the
subsequent shuffle.

Differential Revision: https://reviews.llvm.org/D47686

llvm-svn: 333962
  • Loading branch information
@rotateright
rotateright committed Jun 4, 2018
1 parent 800ac42 commit dcb8d30
Show file tree
Hide file tree
Showing 2 changed files with 33 additions and 33 deletions.
28 changes: 14 additions & 14 deletions llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
View file
Expand Up @@ -1417,21 +1417,21 @@ Instruction *InstCombiner::foldShuffledBinop(BinaryOperator &Inst) {
}
}
if (MayChange) {
// It's not safe to use a vector with undef elements because the entire
// instruction can be folded to undef (for example, div/rem divisors).
// Replace undef lanes with the first non-undef element. Vector demanded
// elements can change those back to undef values if that is safe.
Constant *SafeDummyConstant = nullptr;
for (unsigned i = 0; i < VWidth; ++i) {
if (!isa<UndefValue>(NewVecC[i])) {
SafeDummyConstant = NewVecC[i];
break;
}
// With integer div/rem instructions, it is not safe to use a vector with
// undef elements because the entire instruction can be folded to undef.
// So replace undef elements with '1' because that can never induce
// undefined behavior. All other binop opcodes are always safe to
// speculate, and therefore, it is fine to include undef elements for
// unused lanes (and using undefs may help optimization).
BinaryOperator::BinaryOps Opcode = Inst.getOpcode();
if (Opcode == Instruction::UDiv || Opcode == Instruction::URem ||
Opcode == Instruction::SDiv || Opcode == Instruction::SRem) {
assert(C->getType()->getScalarType()->isIntegerTy() &&
"Not expecting FP opcodes/operands/constants here");
for (unsigned i = 0; i < VWidth; ++i)
if (isa<UndefValue>(NewVecC[i]))
NewVecC[i] = ConstantInt::get(NewVecC[i]->getType(), 1);
}
assert(SafeDummyConstant && "Undef constant vector was not simplified?");
for (unsigned i = 0; i < VWidth; ++i)
if (isa<UndefValue>(NewVecC[i]))
NewVecC[i] = SafeDummyConstant;

// Op(shuffle(V1, Mask), C) -> shuffle(Op(V1, NewC), Mask)
// Op(C, shuffle(V1, Mask)) -> shuffle(Op(NewC, V1), Mask)
Expand Down
38 changes: 19 additions & 19 deletions llvm/test/Transforms/InstCombine/vec_shuffle.ll
View file
Expand Up @@ -452,7 +452,7 @@ define <4 x i32> @mul_const_splat(<4 x i32> %v) {

define <4 x i32> @lshr_const_half_splat(<4 x i32> %v) {
; CHECK-LABEL: @lshr_const_half_splat(
; CHECK-NEXT: [[TMP1:%.*]] = lshr <4 x i32> <i32 8, i32 8, i32 9, i32 8>, [[V:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = lshr <4 x i32> <i32 undef, i32 8, i32 9, i32 undef>, [[V:%.*]]
; CHECK-NEXT: [[R:%.*]] = shufflevector <4 x i32> [[TMP1]], <4 x i32> undef, <4 x i32> <i32 1, i32 1, i32 2, i32 2>
; CHECK-NEXT: ret <4 x i32> [[R]]
;
Expand Down Expand Up @@ -643,7 +643,7 @@ define <2 x i32> @mul_splat_constant(<2 x i32> %x) {

define <2 x i32> @shl_splat_constant0(<2 x i32> %x) {
; CHECK-LABEL: @shl_splat_constant0(
; CHECK-NEXT: [[TMP1:%.*]] = shl <2 x i32> <i32 5, i32 5>, [[X:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = shl <2 x i32> <i32 5, i32 undef>, [[X:%.*]]
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x i32> [[R]]
;
Expand All @@ -654,7 +654,7 @@ define <2 x i32> @shl_splat_constant0(<2 x i32> %x) {

define <2 x i32> @shl_splat_constant1(<2 x i32> %x) {
; CHECK-LABEL: @shl_splat_constant1(
; CHECK-NEXT: [[TMP1:%.*]] = shl <2 x i32> [[X:%.*]], <i32 5, i32 5>
; CHECK-NEXT: [[TMP1:%.*]] = shl <2 x i32> [[X:%.*]], <i32 5, i32 undef>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x i32> [[R]]
;
Expand All @@ -665,7 +665,7 @@ define <2 x i32> @shl_splat_constant1(<2 x i32> %x) {

define <2 x i32> @ashr_splat_constant0(<2 x i32> %x) {
; CHECK-LABEL: @ashr_splat_constant0(
; CHECK-NEXT: [[TMP1:%.*]] = lshr <2 x i32> <i32 5, i32 5>, [[X:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = ashr <2 x i32> <i32 5, i32 undef>, [[X:%.*]]
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x i32> [[R]]
;
Expand All @@ -676,7 +676,7 @@ define <2 x i32> @ashr_splat_constant0(<2 x i32> %x) {

define <2 x i32> @ashr_splat_constant1(<2 x i32> %x) {
; CHECK-LABEL: @ashr_splat_constant1(
; CHECK-NEXT: [[TMP1:%.*]] = ashr <2 x i32> [[X:%.*]], <i32 5, i32 5>
; CHECK-NEXT: [[TMP1:%.*]] = ashr <2 x i32> [[X:%.*]], <i32 5, i32 undef>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x i32> [[R]]
;
Expand All @@ -687,7 +687,7 @@ define <2 x i32> @ashr_splat_constant1(<2 x i32> %x) {

define <2 x i32> @lshr_splat_constant0(<2 x i32> %x) {
; CHECK-LABEL: @lshr_splat_constant0(
; CHECK-NEXT: [[TMP1:%.*]] = lshr <2 x i32> <i32 5, i32 5>, [[X:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = lshr <2 x i32> <i32 5, i32 undef>, [[X:%.*]]
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x i32> [[R]]
;
Expand All @@ -698,7 +698,7 @@ define <2 x i32> @lshr_splat_constant0(<2 x i32> %x) {

define <2 x i32> @lshr_splat_constant1(<2 x i32> %x) {
; CHECK-LABEL: @lshr_splat_constant1(
; CHECK-NEXT: [[TMP1:%.*]] = lshr <2 x i32> [[X:%.*]], <i32 5, i32 5>
; CHECK-NEXT: [[TMP1:%.*]] = lshr <2 x i32> [[X:%.*]], <i32 5, i32 undef>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x i32> [[R]]
;
Expand All @@ -720,7 +720,7 @@ define <2 x i32> @urem_splat_constant0(<2 x i32> %x) {

define <2 x i32> @urem_splat_constant1(<2 x i32> %x) {
; CHECK-LABEL: @urem_splat_constant1(
; CHECK-NEXT: [[TMP1:%.*]] = urem <2 x i32> [[X:%.*]], <i32 42, i32 42>
; CHECK-NEXT: [[TMP1:%.*]] = urem <2 x i32> [[X:%.*]], <i32 42, i32 1>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x i32> [[R]]
;
Expand All @@ -742,7 +742,7 @@ define <2 x i32> @srem_splat_constant0(<2 x i32> %x) {

define <2 x i32> @srem_splat_constant1(<2 x i32> %x) {
; CHECK-LABEL: @srem_splat_constant1(
; CHECK-NEXT: [[TMP1:%.*]] = srem <2 x i32> [[X:%.*]], <i32 42, i32 42>
; CHECK-NEXT: [[TMP1:%.*]] = srem <2 x i32> [[X:%.*]], <i32 42, i32 1>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x i32> [[R]]
;
Expand All @@ -764,7 +764,7 @@ define <2 x i32> @udiv_splat_constant0(<2 x i32> %x) {

define <2 x i32> @udiv_splat_constant1(<2 x i32> %x) {
; CHECK-LABEL: @udiv_splat_constant1(
; CHECK-NEXT: [[TMP1:%.*]] = udiv <2 x i32> [[X:%.*]], <i32 42, i32 42>
; CHECK-NEXT: [[TMP1:%.*]] = udiv <2 x i32> [[X:%.*]], <i32 42, i32 1>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x i32> [[R]]
;
Expand All @@ -786,7 +786,7 @@ define <2 x i32> @sdiv_splat_constant0(<2 x i32> %x) {

define <2 x i32> @sdiv_splat_constant1(<2 x i32> %x) {
; CHECK-LABEL: @sdiv_splat_constant1(
; CHECK-NEXT: [[TMP1:%.*]] = sdiv <2 x i32> [[X:%.*]], <i32 42, i32 42>
; CHECK-NEXT: [[TMP1:%.*]] = sdiv <2 x i32> [[X:%.*]], <i32 42, i32 1>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x i32> [[R]]
;
Expand Down Expand Up @@ -830,7 +830,7 @@ define <2 x i32> @xor_splat_constant(<2 x i32> %x) {

define <2 x float> @fadd_splat_constant(<2 x float> %x) {
; CHECK-LABEL: @fadd_splat_constant(
; CHECK-NEXT: [[TMP1:%.*]] = fadd <2 x float> [[X:%.*]], <float 4.200000e+01, float 4.200000e+01>
; CHECK-NEXT: [[TMP1:%.*]] = fadd <2 x float> [[X:%.*]], <float 4.200000e+01, float undef>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x float> [[TMP1]], <2 x float> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x float> [[R]]
;
Expand All @@ -841,7 +841,7 @@ define <2 x float> @fadd_splat_constant(<2 x float> %x) {

define <2 x float> @fsub_splat_constant0(<2 x float> %x) {
; CHECK-LABEL: @fsub_splat_constant0(
; CHECK-NEXT: [[TMP1:%.*]] = fsub <2 x float> <float 4.200000e+01, float 4.200000e+01>, [[X:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = fsub <2 x float> <float 4.200000e+01, float undef>, [[X:%.*]]
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x float> [[TMP1]], <2 x float> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x float> [[R]]
;
Expand All @@ -852,7 +852,7 @@ define <2 x float> @fsub_splat_constant0(<2 x float> %x) {

define <2 x float> @fsub_splat_constant1(<2 x float> %x) {
; CHECK-LABEL: @fsub_splat_constant1(
; CHECK-NEXT: [[TMP1:%.*]] = fadd <2 x float> [[X:%.*]], <float -4.200000e+01, float -4.200000e+01>
; CHECK-NEXT: [[TMP1:%.*]] = fadd <2 x float> [[X:%.*]], <float -4.200000e+01, float 0x7FF8000000000000>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x float> [[TMP1]], <2 x float> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x float> [[R]]
;
Expand All @@ -863,7 +863,7 @@ define <2 x float> @fsub_splat_constant1(<2 x float> %x) {

define <2 x float> @fmul_splat_constant(<2 x float> %x) {
; CHECK-LABEL: @fmul_splat_constant(
; CHECK-NEXT: [[TMP1:%.*]] = fmul <2 x float> [[X:%.*]], <float 4.200000e+01, float 4.200000e+01>
; CHECK-NEXT: [[TMP1:%.*]] = fmul <2 x float> [[X:%.*]], <float 4.200000e+01, float undef>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x float> [[TMP1]], <2 x float> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x float> [[R]]
;
Expand All @@ -874,7 +874,7 @@ define <2 x float> @fmul_splat_constant(<2 x float> %x) {

define <2 x float> @fdiv_splat_constant0(<2 x float> %x) {
; CHECK-LABEL: @fdiv_splat_constant0(
; CHECK-NEXT: [[TMP1:%.*]] = fdiv <2 x float> <float 4.200000e+01, float 4.200000e+01>, [[X:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = fdiv <2 x float> <float 4.200000e+01, float undef>, [[X:%.*]]
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x float> [[TMP1]], <2 x float> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x float> [[R]]
;
Expand All @@ -885,7 +885,7 @@ define <2 x float> @fdiv_splat_constant0(<2 x float> %x) {

define <2 x float> @fdiv_splat_constant1(<2 x float> %x) {
; CHECK-LABEL: @fdiv_splat_constant1(
; CHECK-NEXT: [[TMP1:%.*]] = fdiv <2 x float> [[X:%.*]], <float 4.200000e+01, float 4.200000e+01>
; CHECK-NEXT: [[TMP1:%.*]] = fdiv <2 x float> [[X:%.*]], <float 4.200000e+01, float undef>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x float> [[TMP1]], <2 x float> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x float> [[R]]
;
Expand All @@ -896,7 +896,7 @@ define <2 x float> @fdiv_splat_constant1(<2 x float> %x) {

define <2 x float> @frem_splat_constant0(<2 x float> %x) {
; CHECK-LABEL: @frem_splat_constant0(
; CHECK-NEXT: [[TMP1:%.*]] = frem <2 x float> <float 4.200000e+01, float 4.200000e+01>, [[X:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = frem <2 x float> <float 4.200000e+01, float undef>, [[X:%.*]]
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x float> [[TMP1]], <2 x float> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x float> [[R]]
;
Expand All @@ -907,7 +907,7 @@ define <2 x float> @frem_splat_constant0(<2 x float> %x) {

define <2 x float> @frem_splat_constant1(<2 x float> %x) {
; CHECK-LABEL: @frem_splat_constant1(
; CHECK-NEXT: [[TMP1:%.*]] = frem <2 x float> [[X:%.*]], <float 4.200000e+01, float 4.200000e+01>
; CHECK-NEXT: [[TMP1:%.*]] = frem <2 x float> [[X:%.*]], <float 4.200000e+01, float undef>
; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x float> [[TMP1]], <2 x float> undef, <2 x i32> zeroinitializer
; CHECK-NEXT: ret <2 x float> [[R]]
;
Expand Down

0 comments on commit dcb8d30

Please sign in to comment.