[llvm] [ValueTracking] Improve `Bitcast` handling to match SDAG (PR #125935)
Abhishek Kaushik via llvm-commits
llvm-commits at lists.llvm.org
Sat Feb 22 10:18:38 PST 2025
https://github.com/abhishek-kaushik22 updated https://github.com/llvm/llvm-project/pull/125935
>From 2c804af30631ee0880179620b0a77b4f4e75c833 Mon Sep 17 00:00:00 2001
From: abhishek-kaushik22 <abhishek.kaushik at intel.com>
Date: Thu, 6 Feb 2025 02:55:03 +0530
Subject: [PATCH 1/4] [ValueTracking] Improve `Bitcast` handling to match SDAG
Closes #125228
---
llvm/lib/Analysis/ValueTracking.cpp | 27 ++++++++++++++++++++++++---
1 file changed, 24 insertions(+), 3 deletions(-)
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index e3e026f7979da..dfe93bea9bee8 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -1325,6 +1325,9 @@ static void computeKnownBitsFromOperator(const Operator *I,
// Look through a cast from narrow vector elements to wider type.
// Examples: v4i32 -> v2i64, v3i8 -> v24
unsigned SubBitWidth = SrcVecTy->getScalarSizeInBits();
+ unsigned NumElts = DemandedElts.getBitWidth();
+ unsigned SubScale = BitWidth / SubBitWidth;
+ bool isLE = Q.DL.isLittleEndian();
if (BitWidth % SubBitWidth == 0) {
// Known bits are automatically intersected across demanded elements of a
// vector. So for example, if a bit is computed as known zero, it must be
@@ -1340,8 +1343,6 @@ static void computeKnownBitsFromOperator(const Operator *I,
//
// The known bits of each sub-element are then inserted into place
// (dependent on endian) to form the full result of known bits.
- unsigned NumElts = DemandedElts.getBitWidth();
- unsigned SubScale = BitWidth / SubBitWidth;
APInt SubDemandedElts = APInt::getZero(NumElts * SubScale);
for (unsigned i = 0; i != NumElts; ++i) {
if (DemandedElts[i])
@@ -1352,10 +1353,30 @@ static void computeKnownBitsFromOperator(const Operator *I,
for (unsigned i = 0; i != SubScale; ++i) {
computeKnownBits(I->getOperand(0), SubDemandedElts.shl(i), KnownSrc,
Depth + 1, Q);
- unsigned ShiftElt = Q.DL.isLittleEndian() ? i : SubScale - 1 - i;
+ unsigned ShiftElt = isLE ? i : SubScale - 1 - i;
Known.insertBits(KnownSrc, ShiftElt * SubBitWidth);
}
}
+
+ if (SubBitWidth % BitWidth == 0) {
+ KnownBits KnownSrc(SubBitWidth);
+ APInt SubDemandedElts =
+ APIntOps::ScaleBitMask(DemandedElts, NumElts / SubScale);
+ computeKnownBits(I->getOperand(0), SubDemandedElts, KnownSrc, Depth + 1,
+ Q);
+
+ Known.Zero.setAllBits();
+ Known.One.setAllBits();
+ for (unsigned i = 0; i != SubScale; ++i) {
+ if (DemandedElts[i]) {
+ unsigned Shifts = isLE ? i : NumElts - 1 - i;
+ unsigned Offset = (Shifts % SubScale) * BitWidth;
+ Known = Known.intersectWith(KnownSrc.extractBits(BitWidth, Offset));
+ if (Known.isUnknown())
+ break;
+ }
+ }
+ }
break;
}
case Instruction::SExt: {
>From 92ebaf5b66dd32f30948ff90eb36ab4ab25180c9 Mon Sep 17 00:00:00 2001
From: abhishek-kaushik22 <abhishek.kaushik at intel.com>
Date: Thu, 6 Feb 2025 21:09:20 +0530
Subject: [PATCH 2/4] Minor change and tests added
---
llvm/lib/Analysis/ValueTracking.cpp | 11 ++++++-----
.../InstCombine/X86/x86-vector-shifts.ll | 14 ++++++++------
2 files changed, 14 insertions(+), 11 deletions(-)
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index dfe93bea9bee8..8a5499a58c772 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -1322,12 +1322,11 @@ static void computeKnownBitsFromOperator(const Operator *I,
isa<ScalableVectorType>(I->getType()))
break;
+ unsigned NumElts = DemandedElts.getBitWidth();
+ bool IsLE = Q.DL.isLittleEndian();
// Look through a cast from narrow vector elements to wider type.
// Examples: v4i32 -> v2i64, v3i8 -> v24
unsigned SubBitWidth = SrcVecTy->getScalarSizeInBits();
- unsigned NumElts = DemandedElts.getBitWidth();
- unsigned SubScale = BitWidth / SubBitWidth;
- bool isLE = Q.DL.isLittleEndian();
if (BitWidth % SubBitWidth == 0) {
// Known bits are automatically intersected across demanded elements of a
// vector. So for example, if a bit is computed as known zero, it must be
@@ -1343,6 +1342,7 @@ static void computeKnownBitsFromOperator(const Operator *I,
//
// The known bits of each sub-element are then inserted into place
// (dependent on endian) to form the full result of known bits.
+ unsigned SubScale = BitWidth / SubBitWidth;
APInt SubDemandedElts = APInt::getZero(NumElts * SubScale);
for (unsigned i = 0; i != NumElts; ++i) {
if (DemandedElts[i])
@@ -1353,12 +1353,13 @@ static void computeKnownBitsFromOperator(const Operator *I,
for (unsigned i = 0; i != SubScale; ++i) {
computeKnownBits(I->getOperand(0), SubDemandedElts.shl(i), KnownSrc,
Depth + 1, Q);
- unsigned ShiftElt = isLE ? i : SubScale - 1 - i;
+ unsigned ShiftElt = IsLE ? i : SubScale - 1 - i;
Known.insertBits(KnownSrc, ShiftElt * SubBitWidth);
}
}
if (SubBitWidth % BitWidth == 0) {
+ unsigned SubScale = SubBitWidth / BitWidth;
KnownBits KnownSrc(SubBitWidth);
APInt SubDemandedElts =
APIntOps::ScaleBitMask(DemandedElts, NumElts / SubScale);
@@ -1369,7 +1370,7 @@ static void computeKnownBitsFromOperator(const Operator *I,
Known.One.setAllBits();
for (unsigned i = 0; i != SubScale; ++i) {
if (DemandedElts[i]) {
- unsigned Shifts = isLE ? i : NumElts - 1 - i;
+ unsigned Shifts = IsLE ? i : NumElts - 1 - i;
unsigned Offset = (Shifts % SubScale) * BitWidth;
Known = Known.intersectWith(KnownSrc.extractBits(BitWidth, Offset));
if (Known.isUnknown())
diff --git a/llvm/test/Transforms/InstCombine/X86/x86-vector-shifts.ll b/llvm/test/Transforms/InstCombine/X86/x86-vector-shifts.ll
index db56080a3ea2b..9fb4bbc557f3c 100644
--- a/llvm/test/Transforms/InstCombine/X86/x86-vector-shifts.ll
+++ b/llvm/test/Transforms/InstCombine/X86/x86-vector-shifts.ll
@@ -3732,19 +3732,21 @@ define <4 x i64> @test_avx2_psrl_0() {
ret <4 x i64> %16
}
-; FIXME: Failure to peek through bitcasts to ensure psllq shift amount is within bounds.
-define <2 x i64> @PR125228(<2 x i64> %v, <2 x i64> %s) {
-; CHECK-LABEL: @PR125228(
+define <2 x i64> @pr125228(<2 x i64> %v, <2 x i64> %s) {
+; CHECK-LABEL: @pr125228(
+; CHECK-NEXT: entry:
; CHECK-NEXT: [[MASK:%.*]] = and <2 x i64> [[S:%.*]], splat (i64 63)
-; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <2 x i64> [[MASK]], <2 x i64> poison, <2 x i32> zeroinitializer
-; CHECK-NEXT: [[SLL0:%.*]] = shl <2 x i64> [[V:%.*]], [[TMP1]]
+; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <2 x i64> [[MASK]], <2 x i64> poison, <2 x i32> zeroinitializer
+; CHECK-NEXT: [[SLL0:%.*]] = shl <2 x i64> [[V:%.*]], [[TMP0]]
; CHECK-NEXT: [[CAST:%.*]] = bitcast <2 x i64> [[MASK]] to <16 x i8>
; CHECK-NEXT: [[PSRLDQ:%.*]] = shufflevector <16 x i8> [[CAST]], <16 x i8> poison, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[CAST3:%.*]] = bitcast <16 x i8> [[PSRLDQ]] to <2 x i64>
-; CHECK-NEXT: [[SLL1:%.*]] = call <2 x i64> @llvm.x86.sse2.psll.q(<2 x i64> [[V]], <2 x i64> [[CAST3]])
+; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <2 x i64> [[CAST3]], <2 x i64> poison, <2 x i32> zeroinitializer
+; CHECK-NEXT: [[SLL1:%.*]] = shl <2 x i64> [[V]], [[TMP1]]
; CHECK-NEXT: [[SHUFP_UNCASTED:%.*]] = shufflevector <2 x i64> [[SLL0]], <2 x i64> [[SLL1]], <2 x i32> <i32 0, i32 3>
; CHECK-NEXT: ret <2 x i64> [[SHUFP_UNCASTED]]
;
+entry:
%mask = and <2 x i64> %s, splat (i64 63)
%sll0 = call <2 x i64> @llvm.x86.sse2.psll.q(<2 x i64> %v, <2 x i64> %mask)
%cast = bitcast <2 x i64> %mask to <16 x i8>
>From 345c8fa8a9cea10595fbaf549f3709ccb2d8bef1 Mon Sep 17 00:00:00 2001
From: abhishek-kaushik22 <abhishek.kaushik at intel.com>
Date: Thu, 6 Feb 2025 21:22:09 +0530
Subject: [PATCH 3/4] Add comment
---
llvm/lib/Analysis/ValueTracking.cpp | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 8a5499a58c772..741353710452e 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -1357,7 +1357,8 @@ static void computeKnownBitsFromOperator(const Operator *I,
Known.insertBits(KnownSrc, ShiftElt * SubBitWidth);
}
}
-
+ // Look through a cast from wider vector elements to narrow type.
+ // Examples: v2i64 -> v4i32
if (SubBitWidth % BitWidth == 0) {
unsigned SubScale = SubBitWidth / BitWidth;
KnownBits KnownSrc(SubBitWidth);
>From 544b87bcc52f6c68d80202912365c0573b4d8a49 Mon Sep 17 00:00:00 2001
From: abhishek-kaushik22 <abhishek.kaushik at intel.com>
Date: Sat, 22 Feb 2025 23:48:18 +0530
Subject: [PATCH 4/4] Update x86-vector-shifts.ll
---
.../InstCombine/X86/x86-vector-shifts.ll | 14 ++++++--------
1 file changed, 6 insertions(+), 8 deletions(-)
diff --git a/llvm/test/Transforms/InstCombine/X86/x86-vector-shifts.ll b/llvm/test/Transforms/InstCombine/X86/x86-vector-shifts.ll
index 9fb4bbc557f3c..cc252ae53803b 100644
--- a/llvm/test/Transforms/InstCombine/X86/x86-vector-shifts.ll
+++ b/llvm/test/Transforms/InstCombine/X86/x86-vector-shifts.ll
@@ -3732,21 +3732,19 @@ define <4 x i64> @test_avx2_psrl_0() {
ret <4 x i64> %16
}
-define <2 x i64> @pr125228(<2 x i64> %v, <2 x i64> %s) {
-; CHECK-LABEL: @pr125228(
-; CHECK-NEXT: entry:
+define <2 x i64> @PR125228(<2 x i64> %v, <2 x i64> %s) {
+; CHECK-LABEL: @PR125228(
; CHECK-NEXT: [[MASK:%.*]] = and <2 x i64> [[S:%.*]], splat (i64 63)
-; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <2 x i64> [[MASK]], <2 x i64> poison, <2 x i32> zeroinitializer
-; CHECK-NEXT: [[SLL0:%.*]] = shl <2 x i64> [[V:%.*]], [[TMP0]]
+; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <2 x i64> [[MASK]], <2 x i64> poison, <2 x i32> zeroinitializer
+; CHECK-NEXT: [[SLL0:%.*]] = shl <2 x i64> [[V:%.*]], [[TMP1]]
; CHECK-NEXT: [[CAST:%.*]] = bitcast <2 x i64> [[MASK]] to <16 x i8>
; CHECK-NEXT: [[PSRLDQ:%.*]] = shufflevector <16 x i8> [[CAST]], <16 x i8> poison, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[CAST3:%.*]] = bitcast <16 x i8> [[PSRLDQ]] to <2 x i64>
-; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <2 x i64> [[CAST3]], <2 x i64> poison, <2 x i32> zeroinitializer
-; CHECK-NEXT: [[SLL1:%.*]] = shl <2 x i64> [[V]], [[TMP1]]
+; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <2 x i64> [[CAST3]], <2 x i64> poison, <2 x i32> zeroinitializer
+; CHECK-NEXT: [[SLL1:%.*]] = shl <2 x i64> [[V]], [[TMP2]]
; CHECK-NEXT: [[SHUFP_UNCASTED:%.*]] = shufflevector <2 x i64> [[SLL0]], <2 x i64> [[SLL1]], <2 x i32> <i32 0, i32 3>
; CHECK-NEXT: ret <2 x i64> [[SHUFP_UNCASTED]]
;
-entry:
%mask = and <2 x i64> %s, splat (i64 63)
%sll0 = call <2 x i64> @llvm.x86.sse2.psll.q(<2 x i64> %v, <2 x i64> %mask)
%cast = bitcast <2 x i64> %mask to <16 x i8>
More information about the llvm-commits
mailing list