[llvm] [GlobalISel] Add G_EXTRACT_SUBVECTOR to computeKnownBits (PR #214533)
Deepak Shirke via llvm-commits
llvm-commits at lists.llvm.org
Sun Aug 9 21:25:46 PDT 2026
https://github.com/deepakshirkem updated https://github.com/llvm/llvm-project/pull/214533
>From 5e36f53f43a761c66afc21701212b1cbc704c0d6 Mon Sep 17 00:00:00 2001
From: deepakshirkem <deepakshirke509 at gmail.com>
Date: Thu, 6 Aug 2026 22:40:46 +0530
Subject: [PATCH 1/3] [GlobalISel] Add G_EXTRACT_SUBVECTOR to computeKnownBits
Port the SDAG EXTRACT_SUBVECTOR computeKnownBits handling to GlobalISel.
Offsets the demanded elements by the subvector index to propagate known
bits from the source vector to the extracted subvector.
---
.../CodeGen/GlobalISel/GISelValueTracking.cpp | 12 +++
.../knownbits-extract-subvector.mir | 73 +++++++++++++++++++
2 files changed, 85 insertions(+)
create mode 100644 llvm/test/CodeGen/AArch64/GlobalISel/knownbits-extract-subvector.mir
diff --git a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
index 139d70f28a4d9..2e980a47398fc 100644
--- a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
@@ -1064,6 +1064,18 @@ void GISelValueTracking::computeKnownBitsImpl(Register R, KnownBits &Known,
}
break;
}
+ case TargetOpcode::G_EXTRACT_SUBVECTOR: {
+ // Offset the demanded elts by the subvector index.
+ Register SrcReg = MI.getOperand(1).getReg();
+ LLT SrcTy = MRI.getType(SrcReg);
+ if (SrcTy.isScalableVector())
+ break;
+ uint64_t Idx = MI.getOperand(2).getImm();
+ unsigned NumSrcElts = SrcTy.getNumElements();
+ APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx);
+ computeKnownBitsImpl(SrcReg, Known, DemandedSrcElts, Depth + 1);
+ break;
+ }
case TargetOpcode::G_SHUFFLE_VECTOR: {
APInt DemandedLHS, DemandedRHS;
// Collect the known bits that are shared by every vector element referenced
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-extract-subvector.mir b/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-extract-subvector.mir
new file mode 100644
index 0000000000000..dafcd3a3ace1d
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-extract-subvector.mir
@@ -0,0 +1,73 @@
+# NOTE: Assertions have been autogenerated by utils/update_givaluetracking_test_checks.py UTC_ARGS: --version 6
+# RUN: llc -mtriple aarch64 -passes="print<gisel-value-tracking>" %s -filetype=null 2>&1 | FileCheck %s
+---
+# Extract subvector from constant vector - known bits propagate
+name: ExtractSubvector_Constant
+body: |
+ bb.0:
+ ; CHECK-LABEL: name: @ExtractSubvector_Constant
+ ; CHECK-NEXT: %0:_ KnownBits:0000000000001111 SignBits:12
+ ; CHECK-NEXT: %1:_ KnownBits:0000000000000111 SignBits:13
+ ; CHECK-NEXT: %2:_ KnownBits:0000000000000011 SignBits:14
+ ; CHECK-NEXT: %3:_ KnownBits:0000000000000001 SignBits:15
+ ; CHECK-NEXT: %4:_ KnownBits:0000000011111111 SignBits:8
+ ; CHECK-NEXT: %5:_ KnownBits:0000000000111111 SignBits:10
+ ; CHECK-NEXT: %6:_ KnownBits:0000000000011111 SignBits:11
+ ; CHECK-NEXT: %7:_ KnownBits:0000000001111111 SignBits:9
+ ; CHECK-NEXT: %8:_ KnownBits:00000000???????1 SignBits:8
+ ; CHECK-NEXT: %9:_ KnownBits:000000000000???1 SignBits:12
+ %0:_(i16) = G_CONSTANT i16 15
+ %1:_(i16) = G_CONSTANT i16 7
+ %2:_(i16) = G_CONSTANT i16 3
+ %3:_(i16) = G_CONSTANT i16 1
+ %4:_(i16) = G_CONSTANT i16 255
+ %5:_(i16) = G_CONSTANT i16 63
+ %6:_(i16) = G_CONSTANT i16 31
+ %7:_(i16) = G_CONSTANT i16 127
+ %8:_(<8 x i16>) = G_BUILD_VECTOR %0, %1, %2, %3, %4, %5, %6, %7
+ %9:_(<4 x i16>) = G_EXTRACT_SUBVECTOR %8, 0
+...
+---
+# Extract subvector with known zero high bits
+name: ExtractSubvector_KnownZero
+body: |
+ bb.0:
+ ; CHECK-LABEL: name: @ExtractSubvector_KnownZero
+ ; CHECK-NEXT: %0:_ KnownBits:???????????????? SignBits:1
+ ; CHECK-NEXT: %1:_ KnownBits:0000000011111111 SignBits:8
+ ; CHECK-NEXT: %2:_ KnownBits:0000000011111111 SignBits:8
+ ; CHECK-NEXT: %3:_ KnownBits:00000000???????? SignBits:8
+ ; CHECK-NEXT: %4:_ KnownBits:00000000???????? SignBits:8
+ %0:_(<8 x i16>) = COPY $q0
+ %1:_(i16) = G_CONSTANT i16 255
+ %2:_(<8 x i16>) = G_BUILD_VECTOR %1, %1, %1, %1, %1, %1, %1, %1
+ %3:_(<8 x i16>) = G_AND %0, %2
+ %4:_(<4 x i16>) = G_EXTRACT_SUBVECTOR %3, 0
+...
+---
+# Extract from non-zero index
+name: ExtractSubvector_HighIndex
+body: |
+ bb.0:
+ ; CHECK-LABEL: name: @ExtractSubvector_HighIndex
+ ; CHECK-NEXT: %0:_ KnownBits:0000000000001111 SignBits:12
+ ; CHECK-NEXT: %1:_ KnownBits:0000000000000111 SignBits:13
+ ; CHECK-NEXT: %2:_ KnownBits:0000000000000011 SignBits:14
+ ; CHECK-NEXT: %3:_ KnownBits:0000000000000001 SignBits:15
+ ; CHECK-NEXT: %4:_ KnownBits:0000000011111111 SignBits:8
+ ; CHECK-NEXT: %5:_ KnownBits:0000000000111111 SignBits:10
+ ; CHECK-NEXT: %6:_ KnownBits:0000000000011111 SignBits:11
+ ; CHECK-NEXT: %7:_ KnownBits:0000000001111111 SignBits:9
+ ; CHECK-NEXT: %8:_ KnownBits:00000000???????1 SignBits:8
+ ; CHECK-NEXT: %9:_ KnownBits:00000000???11111 SignBits:8
+ %0:_(i16) = G_CONSTANT i16 15
+ %1:_(i16) = G_CONSTANT i16 7
+ %2:_(i16) = G_CONSTANT i16 3
+ %3:_(i16) = G_CONSTANT i16 1
+ %4:_(i16) = G_CONSTANT i16 255
+ %5:_(i16) = G_CONSTANT i16 63
+ %6:_(i16) = G_CONSTANT i16 31
+ %7:_(i16) = G_CONSTANT i16 127
+ %8:_(<8 x i16>) = G_BUILD_VECTOR %0, %1, %2, %3, %4, %5, %6, %7
+ %9:_(<4 x i16>) = G_EXTRACT_SUBVECTOR %8, 4
+...
>From ad8c66cfe771bf26b73305bae44ac6f0d8809744 Mon Sep 17 00:00:00 2001
From: deepakshirkem <deepakshirke509 at gmail.com>
Date: Fri, 7 Aug 2026 09:46:46 +0530
Subject: [PATCH 2/3] [AArch64] Update extract-subvec-combine tests for
improved GlobalISel known-bits
GlobalISel now correctly propagates known bits through G_EXTRACT_SUBVECTOR,
allowing it to eliminate redundant AND operations after zero-extension.
GlobalISel output now matches SelectionDAG for these patterns.
[GlobalISel] Add G_EXTRACT_SUBVECTOR to computeKnownBits
Port the SDAG EXTRACT_SUBVECTOR computeKnownBits handling to GlobalISel.
Offsets the demanded elements by the subvector index to propagate known
bits from the source vector to the extracted subvector.
---
.../CodeGen/AArch64/extract-subvec-combine.ll | 35 ++++++-------------
1 file changed, 10 insertions(+), 25 deletions(-)
diff --git a/llvm/test/CodeGen/AArch64/extract-subvec-combine.ll b/llvm/test/CodeGen/AArch64/extract-subvec-combine.ll
index 2779f535adc4b..ce6e3c0a32b38 100644
--- a/llvm/test/CodeGen/AArch64/extract-subvec-combine.ll
+++ b/llvm/test/CodeGen/AArch64/extract-subvec-combine.ll
@@ -3,18 +3,11 @@
; RUN: llc -mtriple=aarch64 -global-isel -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-GI
define <2 x i32> @and_extract_zext_idx0(<4 x i16> %vec) nounwind {
-; CHECK-SD-LABEL: and_extract_zext_idx0:
-; CHECK-SD: // %bb.0:
-; CHECK-SD-NEXT: ushll v0.4s, v0.4h, #0
-; CHECK-SD-NEXT: // kill: def $d0 killed $d0 killed $q0
-; CHECK-SD-NEXT: ret
-;
-; CHECK-GI-LABEL: and_extract_zext_idx0:
-; CHECK-GI: // %bb.0:
-; CHECK-GI-NEXT: movi d1, #0x00ffff0000ffff
-; CHECK-GI-NEXT: ushll v0.4s, v0.4h, #0
-; CHECK-GI-NEXT: and v0.8b, v0.8b, v1.8b
-; CHECK-GI-NEXT: ret
+; CHECK-LABEL: and_extract_zext_idx0:
+; CHECK: // %bb.0:
+; CHECK-NEXT: ushll v0.4s, v0.4h, #0
+; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
+; CHECK-NEXT: ret
%zext = zext <4 x i16> %vec to <4 x i32>
%extract = call <2 x i32> @llvm.vector.extract.v2i32.v4i32(<4 x i32> %zext, i64 0)
%and = and <2 x i32> %extract, <i32 65535, i32 65535>
@@ -41,19 +34,11 @@ define <4 x i16> @and_extract_sext_idx0(<8 x i8> %vec) nounwind {
}
define <2 x i32> @and_extract_zext_idx2(<4 x i16> %vec) nounwind {
-; CHECK-SD-LABEL: and_extract_zext_idx2:
-; CHECK-SD: // %bb.0:
-; CHECK-SD-NEXT: ushll v0.4s, v0.4h, #0
-; CHECK-SD-NEXT: mov d0, v0.d[1]
-; CHECK-SD-NEXT: ret
-;
-; CHECK-GI-LABEL: and_extract_zext_idx2:
-; CHECK-GI: // %bb.0:
-; CHECK-GI-NEXT: ushll v0.4s, v0.4h, #0
-; CHECK-GI-NEXT: movi d1, #0x00ffff0000ffff
-; CHECK-GI-NEXT: mov d0, v0.d[1]
-; CHECK-GI-NEXT: and v0.8b, v0.8b, v1.8b
-; CHECK-GI-NEXT: ret
+; CHECK-LABEL: and_extract_zext_idx2:
+; CHECK: // %bb.0:
+; CHECK-NEXT: ushll v0.4s, v0.4h, #0
+; CHECK-NEXT: mov d0, v0.d[1]
+; CHECK-NEXT: ret
%zext = zext <4 x i16> %vec to <4 x i32>
%extract = call <2 x i32> @llvm.vector.extract.v2i32.v4i32(<4 x i32> %zext, i64 2)
%and = and <2 x i32> %extract, <i32 65535, i32 65535>
>From 9ea751ba8f7d8af14decb78fba1516778cd69a80 Mon Sep 17 00:00:00 2001
From: deepakshirkem <deepakshirke509 at gmail.com>
Date: Mon, 10 Aug 2026 09:55:23 +0530
Subject: [PATCH 3/3] Address review: handle scalable vectors conservatively in
G_EXTRACT_SUBVECTOR
Instead of breaking early for scalable vectors, use APInt(1, 1) to
conservatively demand all elements, matching the SDAG approach.
---
llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp | 14 ++++++++------
1 file changed, 8 insertions(+), 6 deletions(-)
diff --git a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
index 2e980a47398fc..7c98e025599ee 100644
--- a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
@@ -1065,14 +1065,16 @@ void GISelValueTracking::computeKnownBitsImpl(Register R, KnownBits &Known,
break;
}
case TargetOpcode::G_EXTRACT_SUBVECTOR: {
- // Offset the demanded elts by the subvector index.
Register SrcReg = MI.getOperand(1).getReg();
LLT SrcTy = MRI.getType(SrcReg);
- if (SrcTy.isScalableVector())
- break;
- uint64_t Idx = MI.getOperand(2).getImm();
- unsigned NumSrcElts = SrcTy.getNumElements();
- APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx);
+ APInt DemandedSrcElts;
+ if (SrcTy.isScalableVector()) {
+ DemandedSrcElts = APInt(1, 1);
+ } else {
+ uint64_t Idx = MI.getOperand(2).getImm();
+ unsigned NumSrcElts = SrcTy.getNumElements();
+ DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx);
+ }
computeKnownBitsImpl(SrcReg, Known, DemandedSrcElts, Depth + 1);
break;
}
More information about the llvm-commits
mailing list