This is an archive of the discontinued LLVM Phabricator instance.

Differential D72101

[InstCombine] replace undef elements in vector constant when doing icmp folds (PR44383)
ClosedPublic

Authored by spatel on Jan 2 2020, 12:55 PM.

Details

Reviewers
nlopes
lebedev.ri
regehr
Commits
rG164058274364: [InstCombine] replace undef elements in vector constant when doing icmp folds…
Summary

As shown in P44383:
https://bugs.llvm.org/show_bug.cgi?id=44383
...we can't safely propagate a vector constant through this icmp fold if that vector constant contains undefined elements.

We know that each defined element of the constant is safe though, so find the first of those and replicate it into the formerly undef lanes.

Diff Detail

Event Timeline

spatel created this revision.Jan 2 2020, 12:55 PM
Herald added a project: Restricted Project. · View Herald TranscriptJan 2 2020, 12:56 PM
Herald added subscribers: hiraditya, mcrosier. · View Herald Transcript
nlopes accepted this revision.Jan 2 2020, 1:54 PM

LGTM, thanks!

This revision is now accepted and ready to land.Jan 2 2020, 1:54 PM
Closed by commit rG164058274364: [InstCombine] replace undef elements in vector constant when doing icmp folds… (authored by spatel). · Explain WhyJan 3 2020, 6:17 AM
This revision was automatically updated to reflect the committed changes.

Revision Contents

Diff 236047

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

Show First 20 Lines • Show All 3,358 Lines • ▼ Show 20 Lines case ICmpInst::Predicate::ICMP_SLE:
if (!match(M, m_NonNegative())) // Must not have any -1 vector elements. if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
return nullptr; return nullptr;
DstPred = ICmpInst::Predicate::ICMP_SLE; DstPred = ICmpInst::Predicate::ICMP_SLE;
break; break;
default: default:
llvm_unreachable("All possible folds are handled."); llvm_unreachable("All possible folds are handled.");
} }
// The mask value may be a vector constant that has undefined elements. But it
// may not be safe to propagate those undefs into the new compare, so replace
// those elements by copying an existing, defined, and safe scalar constant.
Type *OpTy = M->getType();
auto *VecC = dyn_cast<Constant>(M);
if (OpTy->isVectorTy() && VecC && VecC->containsUndefElement()) {
Constant *SafeReplacementConstant = nullptr;
for (unsigned i = 0, e = OpTy->getVectorNumElements(); i != e; ++i) {
if (!isa<UndefValue>(VecC->getAggregateElement(i))) {
SafeReplacementConstant = VecC->getAggregateElement(i);
break;
}
}
assert(SafeReplacementConstant && "Failed to find undef replacement");
M = Constant::replaceUndefsWith(VecC, SafeReplacementConstant);
}
return Builder.CreateICmp(DstPred, X, M); return Builder.CreateICmp(DstPred, X, M);
} }
/// Some comparisons can be simplified. /// Some comparisons can be simplified.
/// In this case, we are looking for comparisons that look like /// In this case, we are looking for comparisons that look like
/// a check for a lossy signed truncation. /// a check for a lossy signed truncation.
/// Folds: (MaskedBits is a constant.) /// Folds: (MaskedBits is a constant.)
/// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
▲ Show 20 Lines • Show All 2,784 Lines • Show Last 20 Lines

llvm/test/Transforms/InstCombine/canonicalize-constant-low-bit-mask-and-icmp-ne-to-icmp-ugt.ll

Show First 20 Lines • Show All 76 Lines • ▼ Show 20 Lines
; ;
%tmp0 = and <2 x i8> %x, <i8 3, i8 -1> %tmp0 = and <2 x i8> %x, <i8 3, i8 -1>
%ret = icmp ne <2 x i8> %tmp0, %x %ret = icmp ne <2 x i8> %tmp0, %x
ret <2 x i1> %ret ret <2 x i1> %ret
} }
define <3 x i1> @p3_vec_splat_undef(<3 x i8> %x) { define <3 x i1> @p3_vec_splat_undef(<3 x i8> %x) {
; CHECK-LABEL: @p3_vec_splat_undef( ; CHECK-LABEL: @p3_vec_splat_undef(
; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X:%.*]], <i8 3, i8 undef, i8 3> ; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X:%.*]], <i8 3, i8 3, i8 3>
; CHECK-NEXT: ret <3 x i1> [[TMP1]] ; CHECK-NEXT: ret <3 x i1> [[TMP1]]
; ;
%tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 3> %tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 3>
%ret = icmp ne <3 x i8> %tmp0, %x %ret = icmp ne <3 x i8> %tmp0, %x
ret <3 x i1> %ret ret <3 x i1> %ret
} }
define <3 x i1> @p3_vec_nonsplat_undef(<3 x i8> %x) { define <3 x i1> @p3_vec_nonsplat_undef(<3 x i8> %x) {
; CHECK-LABEL: @p3_vec_nonsplat_undef( ; CHECK-LABEL: @p3_vec_nonsplat_undef(
; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X:%.*]], <i8 -1, i8 undef, i8 3> ; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X:%.*]], <i8 -1, i8 -1, i8 3>
; CHECK-NEXT: ret <3 x i1> [[TMP1]] ; CHECK-NEXT: ret <3 x i1> [[TMP1]]
; ;
%tmp0 = and <3 x i8> %x, <i8 -1, i8 undef, i8 3> %tmp0 = and <3 x i8> %x, <i8 -1, i8 undef, i8 3>
%ret = icmp ne <3 x i8> %tmp0, %x %ret = icmp ne <3 x i8> %tmp0, %x
ret <3 x i1> %ret ret <3 x i1> %ret
} }
; ============================================================================ ; ; ============================================================================ ;
▲ Show 20 Lines • Show All 118 Lines • Show Last 20 Lines

llvm/test/Transforms/InstCombine/canonicalize-constant-low-bit-mask-and-icmp-sgt-to-icmp-sgt.ll

Show First 20 Lines • Show All 70 Lines • ▼ Show 20 Lines;
%tmp0 = and <2 x i8> %x, <i8 3, i8 0> %tmp0 = and <2 x i8> %x, <i8 3, i8 0>
%ret = icmp sgt <2 x i8> %x, %tmp0 %ret = icmp sgt <2 x i8> %x, %tmp0
ret <2 x i1> %ret ret <2 x i1> %ret
} }
define <3 x i1> @p3_vec_splat_undef() { define <3 x i1> @p3_vec_splat_undef() {
; CHECK-LABEL: @p3_vec_splat_undef( ; CHECK-LABEL: @p3_vec_splat_undef(
; CHECK-NEXT: [[X:%.*]] = call <3 x i8> @gen3x8() ; CHECK-NEXT: [[X:%.*]] = call <3 x i8> @gen3x8()
; CHECK-NEXT: [[TMP1:%.*]] = icmp sgt <3 x i8> [[X]], <i8 3, i8 undef, i8 3> ; CHECK-NEXT: [[TMP1:%.*]] = icmp sgt <3 x i8> [[X]], <i8 3, i8 3, i8 3>
; CHECK-NEXT: ret <3 x i1> [[TMP1]] ; CHECK-NEXT: ret <3 x i1> [[TMP1]]
; ;
%x = call <3 x i8> @gen3x8() %x = call <3 x i8> @gen3x8()
%tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 3> %tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 3>
%ret = icmp sgt <3 x i8> %x, %tmp0 %ret = icmp sgt <3 x i8> %x, %tmp0
ret <3 x i1> %ret ret <3 x i1> %ret
} }
define <3 x i1> @p3_vec_nonsplat_undef() { define <3 x i1> @p3_vec_nonsplat_undef() {
; CHECK-LABEL: @p3_vec_nonsplat_undef( ; CHECK-LABEL: @p3_vec_nonsplat_undef(
; CHECK-NEXT: [[X:%.*]] = call <3 x i8> @gen3x8() ; CHECK-NEXT: [[X:%.*]] = call <3 x i8> @gen3x8()
; CHECK-NEXT: [[TMP1:%.*]] = icmp sgt <3 x i8> [[X]], <i8 15, i8 3, i8 undef> ; CHECK-NEXT: [[TMP1:%.*]] = icmp sgt <3 x i8> [[X]], <i8 15, i8 3, i8 15>
; CHECK-NEXT: ret <3 x i1> [[TMP1]] ; CHECK-NEXT: ret <3 x i1> [[TMP1]]
; ;
%x = call <3 x i8> @gen3x8() %x = call <3 x i8> @gen3x8()
%tmp0 = and <3 x i8> %x, <i8 15, i8 3, i8 undef> %tmp0 = and <3 x i8> %x, <i8 15, i8 3, i8 undef>
%ret = icmp sgt <3 x i8> %x, %tmp0 %ret = icmp sgt <3 x i8> %x, %tmp0
ret <3 x i1> %ret ret <3 x i1> %ret
} }
▲ Show 20 Lines • Show All 171 Lines • Show Last 20 Lines

llvm/test/Transforms/InstCombine/canonicalize-constant-low-bit-mask-and-icmp-slt-to-icmp-sgt.ll

Show First 20 Lines • Show All 55 Lines • ▼ Show 20 Lines
; ;
%tmp0 = and <2 x i8> %x, <i8 3, i8 0> %tmp0 = and <2 x i8> %x, <i8 3, i8 0>
%ret = icmp slt <2 x i8> %tmp0, %x %ret = icmp slt <2 x i8> %tmp0, %x
ret <2 x i1> %ret ret <2 x i1> %ret
} }
define <3 x i1> @p3_vec_splat_undef(<3 x i8> %x) { define <3 x i1> @p3_vec_splat_undef(<3 x i8> %x) {
; CHECK-LABEL: @p3_vec_splat_undef( ; CHECK-LABEL: @p3_vec_splat_undef(
; CHECK-NEXT: [[TMP1:%.*]] = icmp sgt <3 x i8> [[X:%.*]], <i8 3, i8 undef, i8 3> ; CHECK-NEXT: [[TMP1:%.*]] = icmp sgt <3 x i8> [[X:%.*]], <i8 3, i8 3, i8 3>
; CHECK-NEXT: ret <3 x i1> [[TMP1]] ; CHECK-NEXT: ret <3 x i1> [[TMP1]]
; ;
%tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 3> %tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 3>
%ret = icmp slt <3 x i8> %tmp0, %x %ret = icmp slt <3 x i8> %tmp0, %x
ret <3 x i1> %ret ret <3 x i1> %ret
} }
define <3 x i1> @p3_vec_nonsplat_undef(<3 x i8> %x) { define <3 x i1> @p3_vec_nonsplat_undef(<3 x i8> %x) {
; CHECK-LABEL: @p3_vec_nonsplat_undef( ; CHECK-LABEL: @p3_vec_nonsplat_undef(
; CHECK-NEXT: [[TMP1:%.*]] = icmp sgt <3 x i8> [[X:%.*]], <i8 undef, i8 15, i8 3> ; CHECK-NEXT: [[TMP1:%.*]] = icmp sgt <3 x i8> [[X:%.*]], <i8 15, i8 15, i8 3>
; CHECK-NEXT: ret <3 x i1> [[TMP1]] ; CHECK-NEXT: ret <3 x i1> [[TMP1]]
; ;
%tmp0 = and <3 x i8> %x, <i8 undef, i8 15, i8 3> %tmp0 = and <3 x i8> %x, <i8 undef, i8 15, i8 3>
%ret = icmp slt <3 x i8> %tmp0, %x %ret = icmp slt <3 x i8> %tmp0, %x
ret <3 x i1> %ret ret <3 x i1> %ret
} }
; ============================================================================ ; ; ============================================================================ ;
▲ Show 20 Lines • Show All 161 Lines • Show Last 20 Lines

llvm/test/Transforms/InstCombine/canonicalize-constant-low-bit-mask-and-icmp-ugt-to-icmp-ugt.ll

Show First 20 Lines • Show All 93 Lines • ▼ Show 20 Lines;
%tmp0 = and <2 x i8> %x, <i8 3, i8 -1> %tmp0 = and <2 x i8> %x, <i8 3, i8 -1>
%ret = icmp ugt <2 x i8> %x, %tmp0 %ret = icmp ugt <2 x i8> %x, %tmp0
ret <2 x i1> %ret ret <2 x i1> %ret
} }
define <3 x i1> @p3_vec_splat_undef() { define <3 x i1> @p3_vec_splat_undef() {
; CHECK-LABEL: @p3_vec_splat_undef( ; CHECK-LABEL: @p3_vec_splat_undef(
; CHECK-NEXT: [[X:%.*]] = call <3 x i8> @gen3x8() ; CHECK-NEXT: [[X:%.*]] = call <3 x i8> @gen3x8()
; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X]], <i8 3, i8 undef, i8 3> ; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X]], <i8 3, i8 3, i8 3>
; CHECK-NEXT: ret <3 x i1> [[TMP1]] ; CHECK-NEXT: ret <3 x i1> [[TMP1]]
; ;
%x = call <3 x i8> @gen3x8() %x = call <3 x i8> @gen3x8()
%tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 3> %tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 3>
%ret = icmp ugt <3 x i8> %x, %tmp0 %ret = icmp ugt <3 x i8> %x, %tmp0
ret <3 x i1> %ret ret <3 x i1> %ret
} }
define <3 x i1> @p3_vec_nonsplat_undef() { define <3 x i1> @p3_vec_nonsplat_undef() {
; CHECK-LABEL: @p3_vec_nonsplat_undef( ; CHECK-LABEL: @p3_vec_nonsplat_undef(
; CHECK-NEXT: [[X:%.*]] = call <3 x i8> @gen3x8() ; CHECK-NEXT: [[X:%.*]] = call <3 x i8> @gen3x8()
; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X]], <i8 3, i8 undef, i8 15> ; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X]], <i8 3, i8 3, i8 15>
; CHECK-NEXT: ret <3 x i1> [[TMP1]] ; CHECK-NEXT: ret <3 x i1> [[TMP1]]
; ;
%x = call <3 x i8> @gen3x8() %x = call <3 x i8> @gen3x8()
%tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 15> %tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 15>
%ret = icmp ugt <3 x i8> %x, %tmp0 %ret = icmp ugt <3 x i8> %x, %tmp0
ret <3 x i1> %ret ret <3 x i1> %ret
} }
▲ Show 20 Lines • Show All 116 Lines • Show Last 20 Lines

llvm/test/Transforms/InstCombine/canonicalize-constant-low-bit-mask-and-icmp-ult-to-icmp-ugt.ll

Show First 20 Lines • Show All 77 Lines • ▼ Show 20 Lines
; ;
%tmp0 = and <2 x i8> %x, <i8 3, i8 -1> %tmp0 = and <2 x i8> %x, <i8 3, i8 -1>
%ret = icmp ult <2 x i8> %tmp0, %x %ret = icmp ult <2 x i8> %tmp0, %x
ret <2 x i1> %ret ret <2 x i1> %ret
} }
define <3 x i1> @p3_vec_splat_undef(<3 x i8> %x) { define <3 x i1> @p3_vec_splat_undef(<3 x i8> %x) {
; CHECK-LABEL: @p3_vec_splat_undef( ; CHECK-LABEL: @p3_vec_splat_undef(
; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X:%.*]], <i8 3, i8 undef, i8 3> ; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X:%.*]], <i8 3, i8 3, i8 3>
; CHECK-NEXT: ret <3 x i1> [[TMP1]] ; CHECK-NEXT: ret <3 x i1> [[TMP1]]
; ;
%tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 3> %tmp0 = and <3 x i8> %x, <i8 3, i8 undef, i8 3>
%ret = icmp ult <3 x i8> %tmp0, %x %ret = icmp ult <3 x i8> %tmp0, %x
ret <3 x i1> %ret ret <3 x i1> %ret
} }
define <3 x i1> @p3_vec_nonsplat_undef(<3 x i8> %x) { define <3 x i1> @p3_vec_nonsplat_undef(<3 x i8> %x) {
; CHECK-LABEL: @p3_vec_nonsplat_undef( ; CHECK-LABEL: @p3_vec_nonsplat_undef(
; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X:%.*]], <i8 7, i8 31, i8 undef> ; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt <3 x i8> [[X:%.*]], <i8 7, i8 31, i8 7>
; CHECK-NEXT: ret <3 x i1> [[TMP1]] ; CHECK-NEXT: ret <3 x i1> [[TMP1]]
; ;
%tmp0 = and <3 x i8> %x, <i8 7, i8 31, i8 undef> %tmp0 = and <3 x i8> %x, <i8 7, i8 31, i8 undef>
%ret = icmp ult <3 x i8> %tmp0, %x %ret = icmp ult <3 x i8> %tmp0, %x
ret <3 x i1> %ret ret <3 x i1> %ret
} }
; ============================================================================ ; ; ============================================================================ ;
▲ Show 20 Lines • Show All 114 Lines • Show Last 20 Lines