Reapply this transformation now that it is passing the external test which it previously failed.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110987 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
index a58124d..fea8509 100644
--- a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -448,10 +448,8 @@
if (isa<UndefValue>(SVI.getOperand(2)))
return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
- unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
-
- if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
- return 0;
+ unsigned VWidth = Mask.size();
+ unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements();
APInt UndefElts(VWidth, 0);
APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
@@ -464,14 +462,12 @@
// Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
// Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
if (LHS == RHS || isa<UndefValue>(LHS)) {
- if (isa<UndefValue>(LHS) && LHS == RHS) {
- // shuffle(undef,undef,mask) -> undef.
- return ReplaceInstUsesWith(SVI, LHS);
- }
+ if (isa<UndefValue>(LHS) && LHS == RHS)
+ return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
// Remap any references to RHS to use LHS.
std::vector<Constant*> Elts;
- for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
+ for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
if (Mask[i] >= 2*e)
Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
else {
@@ -495,67 +491,130 @@
}
// Analyze the shuffle, are the LHS or RHS and identity shuffles?
- bool isLHSID = true, isRHSID = true;
-
- for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
- if (Mask[i] >= e*2) continue; // Ignore undef values.
- // Is this an identity shuffle of the LHS value?
- isLHSID &= (Mask[i] == i);
+ if (VWidth == LHSWidth) {
+ bool isLHSID = true, isRHSID = true;
- // Is this an identity shuffle of the RHS value?
- isRHSID &= (Mask[i]-e == i);
+ for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
+ if (Mask[i] >= e*2) continue; // Ignore undef values.
+ // Is this an identity shuffle of the LHS value?
+ isLHSID &= (Mask[i] == i);
+
+ // Is this an identity shuffle of the RHS value?
+ isRHSID &= (Mask[i]-e == i);
+ }
+
+ // Eliminate identity shuffles.
+ if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
+ if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
}
- // Eliminate identity shuffles.
- if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
- if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
+ // Check for a handful of important shuffle(shuffle()) combinations.
+ ShuffleVectorInst *LSVI = dyn_cast<ShuffleVectorInst>(LHS);
+ if (!LSVI)
+ return MadeChange ? &SVI : 0;
+
+ LHS = LSVI->getOperand(0);
+ std::vector<unsigned> LHSMask = getShuffleMask(LSVI);
+ unsigned LHSInNElts = cast<VectorType>(LHS->getType())->getNumElements();
- // If the LHS is a shufflevector itself, see if we can combine it with this
- // one without producing an unusual shuffle. Here we are really conservative:
- // we are absolutely afraid of producing a shuffle mask not in the input
- // program, because the code gen may not be smart enough to turn a merged
- // shuffle into two specific shuffles: it may produce worse code. As such,
- // we only merge two shuffles if the result is one of the two input shuffle
- // masks. In this case, merging the shuffles just removes one instruction,
- // which we know is safe. This is good for things like turning:
- // (splat(splat)) -> splat.
- if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
- if (isa<UndefValue>(RHS)) {
- std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
-
- if (LHSMask.size() == Mask.size()) {
- std::vector<unsigned> NewMask;
- for (unsigned i = 0, e = Mask.size(); i != e; ++i)
- if (Mask[i] >= e)
- NewMask.push_back(2*e);
- else
- NewMask.push_back(LHSMask[Mask[i]]);
-
- // If the result mask is equal to the src shuffle or this
- // shuffle mask, do the replacement.
- if (NewMask == LHSMask || NewMask == Mask) {
- unsigned LHSInNElts =
- cast<VectorType>(LHSSVI->getOperand(0)->getType())->
- getNumElements();
- std::vector<Constant*> Elts;
- for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
- if (NewMask[i] >= LHSInNElts*2) {
- Elts.push_back(UndefValue::get(
- Type::getInt32Ty(SVI.getContext())));
- } else {
- Elts.push_back(ConstantInt::get(
- Type::getInt32Ty(SVI.getContext()),
- NewMask[i]));
- }
- }
- return new ShuffleVectorInst(LHSSVI->getOperand(0),
- LHSSVI->getOperand(1),
- ConstantVector::get(Elts));
- }
+ // If the LHS is an identity shuffle, or if SVI + LHS form a full unpack
+ // operation, merge the LHS and SVI shuffles. This allows llvm to emit
+ // efficient code for matrix transposes written with generic vector ops.
+ bool isLHSLoExtract = true, isLHSHiExtract = true;
+ bool isUnpackLo = isPowerOf2_32(VWidth);
+ bool isUnpackHi = isPowerOf2_32(VWidth);
+ for (unsigned i = 0, e = LHSMask.size(); i != e; ++i) {
+ if (LHSMask[i] >= LHSInNElts*2) continue; // Ignore undef values;
+ isLHSLoExtract &= (LHSMask[i] == i);
+ isLHSHiExtract &= (LHSMask[i] == i+(LHSInNElts/2));
+ isUnpackLo &= (LHSMask[i] == (i/2));
+ isUnpackHi &= (LHSMask[i] == (i/2) + (e/2));
+ }
+ for (unsigned i = 0, e = Mask.size(); i != e && (isUnpackLo || isUnpackHi);
+ i += 2) {
+ isUnpackLo &= (Mask[i] == i) && (Mask[i+1] == (i/2)+e);
+ isUnpackHi &= (Mask[i] == i) && (Mask[i+1] == (i/2)+e+(e/2));
+ }
+ if ((isLHSLoExtract || isLHSHiExtract || isUnpackLo || isUnpackHi) &&
+ (isa<UndefValue>(RHS) || (LHSWidth == LHSInNElts))) {
+ std::vector<Constant*> Elts;
+ for (unsigned i = 0, e = VWidth; i != e; ++i) {
+ if (Mask[i] >= 2*LHSWidth)
+ Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
+ else if (Mask[i] >= e)
+ Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()),
+ Mask[i]));
+ else
+ Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()),
+ LHSMask[Mask[i]]));
+ }
+ if (isa<UndefValue>(RHS))
+ RHS = UndefValue::get(LHS->getType());
+ return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Elts));
+ }
+
+ // If rhs is shuffle + identity, propagate.
+ if (ShuffleVectorInst *RSVI = dyn_cast<ShuffleVectorInst>(RHS)) {
+ std::vector<unsigned> RHSMask = getShuffleMask(RSVI);
+ unsigned RHSInNElts =
+ cast<VectorType>(RSVI->getOperand(0)->getType())->getNumElements();
+
+ // If rhs is identity, propagate
+ bool isRHSLoExtract = true, isRHSHiExtract = true;
+ for (unsigned i = 0, e = RHSMask.size(); i != e; ++i) {
+ if (RHSMask[i] >= RHSInNElts*2) continue; // Ignore undef values;
+ isRHSLoExtract &= (RHSMask[i] == i);
+ isRHSHiExtract &= (RHSMask[i] == i+(RHSInNElts/2));
+ }
+ if ((isRHSLoExtract || isRHSHiExtract) && (LHSWidth == RHSInNElts)) {
+ std::vector<Constant*> Elts;
+ for (unsigned i = 0, e = VWidth; i != e; ++i) {
+ if (Mask[i] >= 2*LHSWidth)
+ Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
+ else if (Mask[i] < LHSWidth)
+ Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()),
+ Mask[i]));
+ else
+ Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()),
+ RHSMask[Mask[i]-LHSWidth]+LHSWidth));
}
+ SVI.setOperand(1, RSVI->getOperand(0));
+ SVI.setOperand(2, ConstantVector::get(Elts));
+ return &SVI;
}
}
+ // Be extremely conservative when merging shufflevector instructions. It is
+ // difficult for the code generator to recognize a merged shuffle, which
+ // usually leads to worse code from merging a shuffle.
+ if (!isa<UndefValue>(RHS))
+ return MadeChange ? &SVI : 0;
+
+ // If the merged shuffle mask is one of the two input shuffle masks, which
+ // just removes one instruction. This should handle splat(splat) -> splat.
+ if (LHSMask.size() == Mask.size()) {
+ std::vector<unsigned> NewMask;
+ for (unsigned i = 0, e = Mask.size(); i != e; ++i)
+ if (Mask[i] >= e)
+ NewMask.push_back(2*e);
+ else
+ NewMask.push_back(LHSMask[Mask[i]]);
+
+ // If the result mask is equal to the src shuffle or this shuffle mask,
+ // do the replacement.
+ if (NewMask == LHSMask || NewMask == Mask) {
+ std::vector<Constant*> Elts;
+ for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
+ if (NewMask[i] >= LHSInNElts*2) {
+ Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
+ } else {
+ Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()),
+ NewMask[i]));
+ }
+ }
+ return new ShuffleVectorInst(LHS, LSVI->getOperand(1),
+ ConstantVector::get(Elts));
+ }
+ }
return MadeChange ? &SVI : 0;
}
-
diff --git a/test/Transforms/InstCombine/vec_shuffle.ll b/test/Transforms/InstCombine/vec_shuffle.ll
index 5132a8f..21fe08a 100644
--- a/test/Transforms/InstCombine/vec_shuffle.ll
+++ b/test/Transforms/InstCombine/vec_shuffle.ll
@@ -87,3 +87,32 @@
%tmp9 = shufflevector <4 x i8> %tmp7, <4 x i8> undef, <4 x i32> < i32 3, i32 1, i32 2, i32 0 > ; <<4 x i8>> [#uses=1]
ret <4 x i8> %tmp9
}
+
+; Test fold of hi/lo vector halves
+; Test fold of unpack operation
+define void @test10(<16 x i8>* %out, <16 x i8> %r, <16 x i8> %g, <16 x i8> %b, <16 x i8> %a) nounwind ssp {
+; CHECK: @test10
+; CHECK-NEXT: shufflevector
+; CHECK-NEXT: shufflevector
+; CHECK-NEXT: store
+; CHECK-NEXT: getelementptr
+; CHECK-NEXT: store
+; CHECK-NEXT: ret
+ %tmp1 = shufflevector <16 x i8> %r, <16 x i8> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7> ; <<8 x i8>> [#uses=1]
+ %tmp3 = shufflevector <8 x i8> %tmp1, <8 x i8> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> ; <<16 x i8>> [#uses=1]
+ %tmp4 = shufflevector <16 x i8> undef, <16 x i8> %tmp3, <16 x i32> <i32 16, i32 1, i32 17, i32 3, i32 18, i32 5, i32 19, i32 7, i32 20, i32 9, i32 21, i32 11, i32 22, i32 13, i32 23, i32 15> ; <<16 x i8>> [#uses=1]
+ %tmp6 = shufflevector <16 x i8> %b, <16 x i8> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7> ; <<8 x i8>> [#uses=1]
+ %tmp8 = shufflevector <8 x i8> %tmp6, <8 x i8> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> ; <<16 x i8>> [#uses=1]
+ %tmp9 = shufflevector <16 x i8> %tmp4, <16 x i8> %tmp8, <16 x i32> <i32 0, i32 16, i32 2, i32 17, i32 4, i32 18, i32 6, i32 19, i32 8, i32 20, i32 10, i32 21, i32 12, i32 22, i32 14, i32 23> ; <<16 x i8>> [#uses=1]
+ %tmp11 = shufflevector <16 x i8> %r, <16 x i8> undef, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15> ; <<8 x i8>> [#uses=1]
+ %tmp13 = shufflevector <8 x i8> %tmp11, <8 x i8> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> ; <<16 x i8>> [#uses=1]
+ %tmp14 = shufflevector <16 x i8> undef, <16 x i8> %tmp13, <16 x i32> <i32 16, i32 1, i32 17, i32 3, i32 18, i32 5, i32 19, i32 7, i32 20, i32 9, i32 21, i32 11, i32 22, i32 13, i32 23, i32 15> ; <<16 x i8>> [#uses=1]
+ %tmp16 = shufflevector <16 x i8> %b, <16 x i8> undef, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15> ; <<8 x i8>> [#uses=1]
+ %tmp18 = shufflevector <8 x i8> %tmp16, <8 x i8> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> ; <<16 x i8>> [#uses=1]
+ %tmp19 = shufflevector <16 x i8> %tmp14, <16 x i8> %tmp18, <16 x i32> <i32 0, i32 16, i32 2, i32 17, i32 4, i32 18, i32 6, i32 19, i32 8, i32 20, i32 10, i32 21, i32 12, i32 22, i32 14, i32 23> ; <<16 x i8>> [#uses=1]
+ %arrayidx = getelementptr inbounds <16 x i8>* %out, i64 0 ; <<16 x i8>*> [#uses=1]
+ store <16 x i8> %tmp9, <16 x i8>* %arrayidx
+ %arrayidx24 = getelementptr inbounds <16 x i8>* %out, i64 1 ; <<16 x i8>*> [#uses=1]
+ store <16 x i8> %tmp19, <16 x i8>* %arrayidx24
+ ret void
+}