[llvm-branch-commits] [llvm] [GlobalISel] Lower vector `bitinsert` and `bitextract` (PR #240587)
Pedro Lobo via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Thu Nov 12 04:47:20 PST 2026
https://github.com/pedroclobo updated https://github.com/llvm/llvm-project/pull/240587
>From 7660aa328e923776f1f74b30c786f0eb0f7c328e Mon Sep 17 00:00:00 2001
From: Pedro Lobo <pedro.lobo at tecnico.ulisboa.pt>
Date: Fri, 6 Nov 2026 13:02:04 +0000
Subject: [PATCH 1/2] [GlobalISel] Lower vector `bitinsert` and `bitextract`
---
llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp | 25 +--
.../test/CodeGen/AArch64/bitextract-vector.ll | 28 ++-
llvm/test/CodeGen/AArch64/bitinsert-vector.ll | 161 ++++++++++++------
3 files changed, 143 insertions(+), 71 deletions(-)
diff --git a/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp b/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
index 38d198046af23..558f132dbf638 100644
--- a/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
@@ -4359,16 +4359,16 @@ bool IRTranslatorImpl::translateBitInsert(const User &U,
LLT BaseTy = MRI.getType(Base);
LLT ValTy = MRI.getType(Val);
- assert(BaseTy.getSizeInBits() >= ValTy.getSizeInBits() &&
+ assert(BaseTy.getScalarSizeInBits() >= ValTy.getScalarSizeInBits() &&
"bitinsert val wider than base should be rejected by verifier");
// If Val is a floating-point type, bitcast it to an integer of the same
// size so buildZExtOrTrunc can safely extend or truncate it.
- if (ValTy.isFloat()) {
- ValTy = LLT::scalar(ValTy.getSizeInBits());
+ if (ValTy.isFloatOrFloatVector()) {
+ ValTy = ValTy.changeToInteger();
Val = MIRBuilder.buildBitcast(ValTy, Val).getReg(0);
- } else if (ValTy.isPointer()) {
- ValTy = LLT::scalar(ValTy.getSizeInBits());
+ } else if (ValTy.isPointerOrPointerVector()) {
+ ValTy = ValTy.changeElementType(LLT::integer(ValTy.getScalarSizeInBits()));
Val = MIRBuilder.buildPtrToInt(ValTy, Val).getReg(0);
}
@@ -4380,8 +4380,8 @@ bool IRTranslatorImpl::translateBitInsert(const User &U,
// Truncate or extend Val to BaseTy so only the inserted bit range remains.
Register ExtVal = MIRBuilder.buildZExtOrTrunc(BaseTy, Val).getReg(0);
- unsigned BaseBitWidth = BaseTy.getSizeInBits();
- unsigned ValBitWidth = ValTy.getSizeInBits();
+ unsigned BaseBitWidth = BaseTy.getScalarSizeInBits();
+ unsigned ValBitWidth = ValTy.getScalarSizeInBits();
APInt InsertMask = APInt::getLowBitsSet(BaseBitWidth, ValBitWidth);
Register MaskConst = MIRBuilder.buildConstant(BaseTy, InsertMask).getReg(0);
Register ShiftedMask =
@@ -4403,7 +4403,7 @@ bool IRTranslatorImpl::translateBitExtract(const User &U,
LLT SrcTy = MRI.getType(Src);
LLT ResTy = MRI.getType(Res);
- assert(ResTy.getSizeInBits() <= SrcTy.getSizeInBits() &&
+ assert(ResTy.getScalarSizeInBits() <= SrcTy.getScalarSizeInBits() &&
"bitextract result wider than source should be rejected by verifier");
// Convert Offset to the target's preferred shift amount type.
@@ -4414,9 +4414,9 @@ bool IRTranslatorImpl::translateBitExtract(const User &U,
// Shift right by Offset to bring the target field down to bit 0.
Register Shifted = MIRBuilder.buildLShr(SrcTy, Src, LegalOffset).getReg(0);
- if (ResTy.isFloat()) {
+ if (ResTy.isFloatOrFloatVector()) {
// Drop into the integer domain to safely handle the size conversion
- LLT IntResTy = LLT::scalar(ResTy.getSizeInBits());
+ LLT IntResTy = ResTy.changeToInteger();
Register IntRes = MRI.createGenericVirtualRegister(IntResTy);
if (SrcTy == IntResTy)
@@ -4427,9 +4427,10 @@ bool IRTranslatorImpl::translateBitExtract(const User &U,
// Bitcast the raw integer bits back into the requested floating-point
// register
MIRBuilder.buildBitcast(Res, IntRes);
- } else if (ResTy.isPointer()) {
+ } else if (ResTy.isPointerOrPointerVector()) {
// Drop into the integer domain to safely handle the size conversion
- LLT IntResTy = LLT::scalar(ResTy.getSizeInBits());
+ LLT IntResTy =
+ ResTy.changeElementType(LLT::integer(ResTy.getScalarSizeInBits()));
Register IntRes = MRI.createGenericVirtualRegister(IntResTy);
if (SrcTy == IntResTy)
diff --git a/llvm/test/CodeGen/AArch64/bitextract-vector.ll b/llvm/test/CodeGen/AArch64/bitextract-vector.ll
index e43dd11e7bebd..1a073361c6f55 100644
--- a/llvm/test/CodeGen/AArch64/bitextract-vector.ll
+++ b/llvm/test/CodeGen/AArch64/bitextract-vector.ll
@@ -1,5 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
-; RUN: llc -mtriple=aarch64 < %s | FileCheck %s
+; RUN: llc -mtriple=aarch64 < %s | FileCheck %s --check-prefixes=CHECK,CHECK-SD
+; RUN: llc -mtriple=aarch64 -global-isel < %s | FileCheck %s --check-prefixes=CHECK,CHECK-GI
define <4 x i8> @test_bitextract_v4i8_const(<4 x b32> %src) {
; CHECK-LABEL: test_bitextract_v4i8_const:
@@ -42,14 +43,23 @@ define <2 x ptr> @test_bitextract_v2ptr(<2 x b64> %src) {
}
define <16 x i4> @test_bitextract_v16i4_var(<16 x b8> %src, <16 x i32> %off) {
-; CHECK-LABEL: test_bitextract_v16i4_var:
-; CHECK: // %bb.0:
-; CHECK-NEXT: uzp1 v3.8h, v3.8h, v4.8h
-; CHECK-NEXT: uzp1 v1.8h, v1.8h, v2.8h
-; CHECK-NEXT: uzp1 v1.16b, v1.16b, v3.16b
-; CHECK-NEXT: neg v1.16b, v1.16b
-; CHECK-NEXT: ushl v0.16b, v0.16b, v1.16b
-; CHECK-NEXT: ret
+; CHECK-SD-LABEL: test_bitextract_v16i4_var:
+; CHECK-SD: // %bb.0:
+; CHECK-SD-NEXT: uzp1 v3.8h, v3.8h, v4.8h
+; CHECK-SD-NEXT: uzp1 v1.8h, v1.8h, v2.8h
+; CHECK-SD-NEXT: uzp1 v1.16b, v1.16b, v3.16b
+; CHECK-SD-NEXT: neg v1.16b, v1.16b
+; CHECK-SD-NEXT: ushl v0.16b, v0.16b, v1.16b
+; CHECK-SD-NEXT: ret
+;
+; CHECK-GI-LABEL: test_bitextract_v16i4_var:
+; CHECK-GI: // %bb.0:
+; CHECK-GI-NEXT: uzp1 v1.8h, v1.8h, v2.8h
+; CHECK-GI-NEXT: uzp1 v2.8h, v3.8h, v4.8h
+; CHECK-GI-NEXT: uzp1 v1.16b, v1.16b, v2.16b
+; CHECK-GI-NEXT: neg v1.16b, v1.16b
+; CHECK-GI-NEXT: ushl v0.16b, v0.16b, v1.16b
+; CHECK-GI-NEXT: ret
%result = bitextract <16 x i4>, <16 x b8> %src, <16 x i32> %off
ret <16 x i4> %result
}
diff --git a/llvm/test/CodeGen/AArch64/bitinsert-vector.ll b/llvm/test/CodeGen/AArch64/bitinsert-vector.ll
index 0d6f3a9fc18e9..33f56ecece49e 100644
--- a/llvm/test/CodeGen/AArch64/bitinsert-vector.ll
+++ b/llvm/test/CodeGen/AArch64/bitinsert-vector.ll
@@ -1,45 +1,80 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
-; RUN: llc -mtriple=aarch64 < %s | FileCheck %s
+; RUN: llc -mtriple=aarch64 < %s | FileCheck %s --check-prefixes=CHECK,CHECK-SD
+; RUN: llc -mtriple=aarch64 -global-isel < %s | FileCheck %s --check-prefixes=CHECK,CHECK-GI
define <4 x b32> @test_bitinsert_v4i8_const(<4 x b32> %base, <4 x i8> %val) {
-; CHECK-LABEL: test_bitinsert_v4i8_const:
-; CHECK: // %bb.0:
-; CHECK-NEXT: bic v1.4h, #255, lsl #8
-; CHECK-NEXT: bic v0.4s, #255, lsl #8
-; CHECK-NEXT: ushll v1.4s, v1.4h, #8
-; CHECK-NEXT: orr v0.16b, v0.16b, v1.16b
-; CHECK-NEXT: ret
+; CHECK-SD-LABEL: test_bitinsert_v4i8_const:
+; CHECK-SD: // %bb.0:
+; CHECK-SD-NEXT: bic v1.4h, #255, lsl #8
+; CHECK-SD-NEXT: bic v0.4s, #255, lsl #8
+; CHECK-SD-NEXT: ushll v1.4s, v1.4h, #8
+; CHECK-SD-NEXT: orr v0.16b, v0.16b, v1.16b
+; CHECK-SD-NEXT: ret
+;
+; CHECK-GI-LABEL: test_bitinsert_v4i8_const:
+; CHECK-GI: // %bb.0:
+; CHECK-GI-NEXT: movi v2.2d, #0x0000ff000000ff
+; CHECK-GI-NEXT: ushll v1.4s, v1.4h, #0
+; CHECK-GI-NEXT: movi v3.2d, #0xffff00ffffff00ff
+; CHECK-GI-NEXT: and v1.16b, v1.16b, v2.16b
+; CHECK-GI-NEXT: and v0.16b, v0.16b, v3.16b
+; CHECK-GI-NEXT: shl v1.4s, v1.4s, #8
+; CHECK-GI-NEXT: orr v0.16b, v0.16b, v1.16b
+; CHECK-GI-NEXT: ret
%result = bitinsert <4 x b32> %base, <4 x i8> %val, <4 x i32> splat (i32 8)
ret <4 x b32> %result
}
define <4 x b32> @test_bitinsert_v4i8_var(<4 x b32> %base, <4 x i8> %val, <4 x i32> %off) {
-; CHECK-LABEL: test_bitinsert_v4i8_var:
-; CHECK: // %bb.0:
-; CHECK-NEXT: movi v3.2d, #0x0000ff000000ff
-; CHECK-NEXT: bic v1.4h, #255, lsl #8
-; CHECK-NEXT: ushll v1.4s, v1.4h, #0
-; CHECK-NEXT: ushl v3.4s, v3.4s, v2.4s
-; CHECK-NEXT: ushl v1.4s, v1.4s, v2.4s
-; CHECK-NEXT: bic v0.16b, v0.16b, v3.16b
-; CHECK-NEXT: orr v0.16b, v0.16b, v1.16b
-; CHECK-NEXT: ret
+; CHECK-SD-LABEL: test_bitinsert_v4i8_var:
+; CHECK-SD: // %bb.0:
+; CHECK-SD-NEXT: movi v3.2d, #0x0000ff000000ff
+; CHECK-SD-NEXT: bic v1.4h, #255, lsl #8
+; CHECK-SD-NEXT: ushll v1.4s, v1.4h, #0
+; CHECK-SD-NEXT: ushl v3.4s, v3.4s, v2.4s
+; CHECK-SD-NEXT: ushl v1.4s, v1.4s, v2.4s
+; CHECK-SD-NEXT: bic v0.16b, v0.16b, v3.16b
+; CHECK-SD-NEXT: orr v0.16b, v0.16b, v1.16b
+; CHECK-SD-NEXT: ret
+;
+; CHECK-GI-LABEL: test_bitinsert_v4i8_var:
+; CHECK-GI: // %bb.0:
+; CHECK-GI-NEXT: movi v3.2d, #0x0000ff000000ff
+; CHECK-GI-NEXT: ushll v1.4s, v1.4h, #0
+; CHECK-GI-NEXT: and v1.16b, v1.16b, v3.16b
+; CHECK-GI-NEXT: ushl v3.4s, v3.4s, v2.4s
+; CHECK-GI-NEXT: bic v0.16b, v0.16b, v3.16b
+; CHECK-GI-NEXT: ushl v1.4s, v1.4s, v2.4s
+; CHECK-GI-NEXT: orr v0.16b, v0.16b, v1.16b
+; CHECK-GI-NEXT: ret
%result = bitinsert <4 x b32> %base, <4 x i8> %val, <4 x i32> %off
ret <4 x b32> %result
}
define <2 x b64> @test_bitinsert_v2f32(<2 x b64> %base, <2 x float> %val) {
-; CHECK-LABEL: test_bitinsert_v2f32:
-; CHECK: // %bb.0:
-; CHECK-NEXT: adrp x8, .LCPI2_0
-; CHECK-NEXT: ushll v1.2d, v1.2s, #0
-; CHECK-NEXT: adrp x9, .LCPI2_1
-; CHECK-NEXT: ldr q2, [x8, :lo12:.LCPI2_0]
-; CHECK-NEXT: ldr q3, [x9, :lo12:.LCPI2_1]
-; CHECK-NEXT: ushl v1.2d, v1.2d, v2.2d
-; CHECK-NEXT: and v0.16b, v0.16b, v3.16b
-; CHECK-NEXT: orr v0.16b, v0.16b, v1.16b
-; CHECK-NEXT: ret
+; CHECK-SD-LABEL: test_bitinsert_v2f32:
+; CHECK-SD: // %bb.0:
+; CHECK-SD-NEXT: adrp x8, .LCPI2_0
+; CHECK-SD-NEXT: ushll v1.2d, v1.2s, #0
+; CHECK-SD-NEXT: adrp x9, .LCPI2_1
+; CHECK-SD-NEXT: ldr q2, [x8, :lo12:.LCPI2_0]
+; CHECK-SD-NEXT: ldr q3, [x9, :lo12:.LCPI2_1]
+; CHECK-SD-NEXT: ushl v1.2d, v1.2d, v2.2d
+; CHECK-SD-NEXT: and v0.16b, v0.16b, v3.16b
+; CHECK-SD-NEXT: orr v0.16b, v0.16b, v1.16b
+; CHECK-SD-NEXT: ret
+;
+; CHECK-GI-LABEL: test_bitinsert_v2f32:
+; CHECK-GI: // %bb.0:
+; CHECK-GI-NEXT: adrp x8, .LCPI2_1
+; CHECK-GI-NEXT: ushll v1.2d, v1.2s, #0
+; CHECK-GI-NEXT: adrp x9, .LCPI2_0
+; CHECK-GI-NEXT: ldr q2, [x8, :lo12:.LCPI2_1]
+; CHECK-GI-NEXT: ldr q3, [x9, :lo12:.LCPI2_0]
+; CHECK-GI-NEXT: and v0.16b, v0.16b, v3.16b
+; CHECK-GI-NEXT: ushl v1.2d, v1.2d, v2.2d
+; CHECK-GI-NEXT: orr v0.16b, v0.16b, v1.16b
+; CHECK-GI-NEXT: ret
%result = bitinsert <2 x b64> %base, <2 x float> %val, <2 x i32> <i32 0, i32 32>
ret <2 x b64> %result
}
@@ -54,32 +89,58 @@ define <2 x b64> @test_bitinsert_v2ptr(<2 x b64> %base, <2 x ptr> %val) {
}
define <2 x b128> @test_bitinsert_v2ptr_b128(<2 x b128> %base, <2 x ptr> %val) {
-; CHECK-LABEL: test_bitinsert_v2ptr_b128:
-; CHECK: // %bb.0:
-; CHECK-NEXT: mov x8, v0.d[1]
-; CHECK-NEXT: fmov x9, d0
-; CHECK-NEXT: bfi x0, x9, #32, #32
-; CHECK-NEXT: bfxil x1, x9, #32, #32
-; CHECK-NEXT: bfi x2, x8, #32, #32
-; CHECK-NEXT: bfxil x3, x8, #32, #32
-; CHECK-NEXT: ret
+; CHECK-SD-LABEL: test_bitinsert_v2ptr_b128:
+; CHECK-SD: // %bb.0:
+; CHECK-SD-NEXT: mov x8, v0.d[1]
+; CHECK-SD-NEXT: fmov x9, d0
+; CHECK-SD-NEXT: bfi x0, x9, #32, #32
+; CHECK-SD-NEXT: bfxil x1, x9, #32, #32
+; CHECK-SD-NEXT: bfi x2, x8, #32, #32
+; CHECK-SD-NEXT: bfxil x3, x8, #32, #32
+; CHECK-SD-NEXT: ret
+;
+; CHECK-GI-LABEL: test_bitinsert_v2ptr_b128:
+; CHECK-GI: // %bb.0:
+; CHECK-GI-NEXT: mov d1, v0.d[1]
+; CHECK-GI-NEXT: fmov x9, d0
+; CHECK-GI-NEXT: and x8, x1, #0xffffffff00000000
+; CHECK-GI-NEXT: and x10, x3, #0xffffffff00000000
+; CHECK-GI-NEXT: bfi x0, x9, #32, #32
+; CHECK-GI-NEXT: orr x1, x8, x9, lsr #32
+; CHECK-GI-NEXT: fmov x11, d1
+; CHECK-GI-NEXT: bfi x2, x11, #32, #32
+; CHECK-GI-NEXT: orr x3, x10, x11, lsr #32
+; CHECK-GI-NEXT: ret
%result = bitinsert <2 x b128> %base, <2 x ptr> %val, <2 x i32> splat (i32 32)
ret <2 x b128> %result
}
define <16 x b8> @test_bitinsert_v16i4_var(<16 x b8> %base, <16 x i4> %val, <16 x i32> %off) {
-; CHECK-LABEL: test_bitinsert_v16i4_var:
-; CHECK: // %bb.0:
-; CHECK-NEXT: uzp1 v4.8h, v4.8h, v5.8h
-; CHECK-NEXT: uzp1 v2.8h, v2.8h, v3.8h
-; CHECK-NEXT: movi v3.16b, #15
-; CHECK-NEXT: uzp1 v2.16b, v2.16b, v4.16b
-; CHECK-NEXT: and v1.16b, v1.16b, v3.16b
-; CHECK-NEXT: ushl v3.16b, v3.16b, v2.16b
-; CHECK-NEXT: ushl v1.16b, v1.16b, v2.16b
-; CHECK-NEXT: bic v0.16b, v0.16b, v3.16b
-; CHECK-NEXT: orr v0.16b, v0.16b, v1.16b
-; CHECK-NEXT: ret
+; CHECK-SD-LABEL: test_bitinsert_v16i4_var:
+; CHECK-SD: // %bb.0:
+; CHECK-SD-NEXT: uzp1 v4.8h, v4.8h, v5.8h
+; CHECK-SD-NEXT: uzp1 v2.8h, v2.8h, v3.8h
+; CHECK-SD-NEXT: movi v3.16b, #15
+; CHECK-SD-NEXT: uzp1 v2.16b, v2.16b, v4.16b
+; CHECK-SD-NEXT: and v1.16b, v1.16b, v3.16b
+; CHECK-SD-NEXT: ushl v3.16b, v3.16b, v2.16b
+; CHECK-SD-NEXT: ushl v1.16b, v1.16b, v2.16b
+; CHECK-SD-NEXT: bic v0.16b, v0.16b, v3.16b
+; CHECK-SD-NEXT: orr v0.16b, v0.16b, v1.16b
+; CHECK-SD-NEXT: ret
+;
+; CHECK-GI-LABEL: test_bitinsert_v16i4_var:
+; CHECK-GI: // %bb.0:
+; CHECK-GI-NEXT: uzp1 v2.8h, v2.8h, v3.8h
+; CHECK-GI-NEXT: uzp1 v3.8h, v4.8h, v5.8h
+; CHECK-GI-NEXT: uzp1 v2.16b, v2.16b, v3.16b
+; CHECK-GI-NEXT: movi v3.16b, #15
+; CHECK-GI-NEXT: and v1.16b, v1.16b, v3.16b
+; CHECK-GI-NEXT: ushl v3.16b, v3.16b, v2.16b
+; CHECK-GI-NEXT: bic v0.16b, v0.16b, v3.16b
+; CHECK-GI-NEXT: ushl v1.16b, v1.16b, v2.16b
+; CHECK-GI-NEXT: orr v0.16b, v0.16b, v1.16b
+; CHECK-GI-NEXT: ret
%result = bitinsert <16 x b8> %base, <16 x i4> %val, <16 x i32> %off
ret <16 x b8> %result
}
>From 047c70db14e7736526bdcfba32e8cbf5c80c297b Mon Sep 17 00:00:00 2001
From: Pedro Lobo <pedro.lobo at tecnico.ulisboa.pt>
Date: Tue, 10 Nov 2026 22:55:07 +0000
Subject: [PATCH 2/2] [AArch64] Use `-global-isel=0` in vector
`bitinsert`/`bitextract` test
---
llvm/test/CodeGen/AArch64/bitextract-vector.ll | 2 +-
llvm/test/CodeGen/AArch64/bitinsert-vector.ll | 2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/llvm/test/CodeGen/AArch64/bitextract-vector.ll b/llvm/test/CodeGen/AArch64/bitextract-vector.ll
index 1a073361c6f55..4ab2aac7f3556 100644
--- a/llvm/test/CodeGen/AArch64/bitextract-vector.ll
+++ b/llvm/test/CodeGen/AArch64/bitextract-vector.ll
@@ -1,5 +1,5 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
-; RUN: llc -mtriple=aarch64 < %s | FileCheck %s --check-prefixes=CHECK,CHECK-SD
+; RUN: llc -mtriple=aarch64 -global-isel=0 < %s | FileCheck %s --check-prefixes=CHECK,CHECK-SD
; RUN: llc -mtriple=aarch64 -global-isel < %s | FileCheck %s --check-prefixes=CHECK,CHECK-GI
define <4 x i8> @test_bitextract_v4i8_const(<4 x b32> %src) {
diff --git a/llvm/test/CodeGen/AArch64/bitinsert-vector.ll b/llvm/test/CodeGen/AArch64/bitinsert-vector.ll
index 33f56ecece49e..05530b3dd6edf 100644
--- a/llvm/test/CodeGen/AArch64/bitinsert-vector.ll
+++ b/llvm/test/CodeGen/AArch64/bitinsert-vector.ll
@@ -1,5 +1,5 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
-; RUN: llc -mtriple=aarch64 < %s | FileCheck %s --check-prefixes=CHECK,CHECK-SD
+; RUN: llc -mtriple=aarch64 -global-isel=0 < %s | FileCheck %s --check-prefixes=CHECK,CHECK-SD
; RUN: llc -mtriple=aarch64 -global-isel < %s | FileCheck %s --check-prefixes=CHECK,CHECK-GI
define <4 x b32> @test_bitinsert_v4i8_const(<4 x b32> %base, <4 x i8> %val) {
More information about the llvm-branch-commits
mailing list