[llvm] [X86] Optimize (a & 0xff) | ((b & 0xff) << 8) into high-byte register insertions (PR #214645)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 7 08:01:21 PDT 2026
https://github.com/AZero13 updated https://github.com/llvm/llvm-project/pull/214645
>From eb4a27eed1b0a16043273371fa28374ad86dfa9d Mon Sep 17 00:00:00 2001
From: AZero13 <gfunni234 at gmail.com>
Date: Fri, 7 Aug 2026 01:24:04 -0400
Subject: [PATCH 1/2] Pre-commit test (NFC)
---
llvm/test/CodeGen/X86/h-register-insert.ll | 73 ++++++++++++++++++++++
1 file changed, 73 insertions(+)
create mode 100644 llvm/test/CodeGen/X86/h-register-insert.ll
diff --git a/llvm/test/CodeGen/X86/h-register-insert.ll b/llvm/test/CodeGen/X86/h-register-insert.ll
new file mode 100644
index 0000000000000..77c1d8d6ffa24
--- /dev/null
+++ b/llvm/test/CodeGen/X86/h-register-insert.ll
@@ -0,0 +1,73 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-unknown-unknown | FileCheck %s --check-prefix=X86
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefix=X64
+
+define i32 @test_insert_32(i8 inreg %A, i8 inreg %B) {
+; X86-LABEL: test_insert_32:
+; X86: # %bb.0:
+; X86-NEXT: movzbl %al, %ecx
+; X86-NEXT: movzbl %dl, %eax
+; X86-NEXT: shll $8, %eax
+; X86-NEXT: orl %ecx, %eax
+; X86-NEXT: retl
+;
+; X64-LABEL: test_insert_32:
+; X64: # %bb.0:
+; X64-NEXT: movzbl %dil, %ecx
+; X64-NEXT: movzbl %sil, %eax
+; X64-NEXT: shll $8, %eax
+; X64-NEXT: orl %ecx, %eax
+; X64-NEXT: retq
+ %conv = zext i8 %A to i32
+ %conv1 = zext i8 %B to i32
+ %shl = shl nuw nsw i32 %conv1, 8
+ %or = or disjoint i32 %shl, %conv
+ ret i32 %or
+}
+
+define i16 @test_insert_16(i8 inreg %A, i8 inreg %B) {
+; X86-LABEL: test_insert_16:
+; X86: # %bb.0:
+; X86-NEXT: movzbl %al, %eax
+; X86-NEXT: shll $8, %edx
+; X86-NEXT: orl %edx, %eax
+; X86-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-NEXT: retl
+;
+; X64-LABEL: test_insert_16:
+; X64: # %bb.0:
+; X64-NEXT: movzbl %dil, %eax
+; X64-NEXT: shll $8, %esi
+; X64-NEXT: orl %esi, %eax
+; X64-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-NEXT: retq
+ %conv = zext i8 %A to i16
+ %conv1 = zext i8 %B to i16
+ %shl = shl nuw nsw i16 %conv1, 8
+ %or = or disjoint i16 %shl, %conv
+ ret i16 %or
+}
+
+define i64 @test_insert_64(i8 inreg %A, i8 inreg %B) {
+; X86-LABEL: test_insert_64:
+; X86: # %bb.0:
+; X86-NEXT: movzbl %al, %ecx
+; X86-NEXT: movzbl %dl, %eax
+; X86-NEXT: shll $8, %eax
+; X86-NEXT: orl %ecx, %eax
+; X86-NEXT: xorl %edx, %edx
+; X86-NEXT: retl
+;
+; X64-LABEL: test_insert_64:
+; X64: # %bb.0:
+; X64-NEXT: movzbl %dil, %ecx
+; X64-NEXT: movzbl %sil, %eax
+; X64-NEXT: shll $8, %eax
+; X64-NEXT: orl %ecx, %eax
+; X64-NEXT: retq
+ %conv = zext i8 %A to i64
+ %conv1 = zext i8 %B to i64
+ %shl = shl nuw nsw i64 %conv1, 8
+ %or = or disjoint i64 %shl, %conv
+ ret i64 %or
+}
>From 01ed66ecae742df86851f8b1fcc3e083426130c2 Mon Sep 17 00:00:00 2001
From: AZero13 <gfunni234 at gmail.com>
Date: Fri, 7 Aug 2026 02:05:48 -0400
Subject: [PATCH 2/2] [X86] Optimize (a & 0xff) | ((b & 0xff) << 8) into
high-byte register insertions
This had to be done in dagtodag because tablegen would not work.
---
llvm/lib/Target/X86/X86ISelDAGToDAG.cpp | 121 +++++++++++++++++++++
llvm/test/CodeGen/X86/h-register-insert.ll | 37 +++++--
2 files changed, 148 insertions(+), 10 deletions(-)
diff --git a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
index d078117061677..072275ad3f5aa 100644
--- a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -599,6 +599,7 @@ namespace {
uint8_t Imm);
bool tryVPTESTM(SDNode *Root, SDValue Setcc, SDValue Mask);
bool tryMatchBitSelect(SDNode *N);
+ bool tryMatchHRegisterInsert(SDNode *N);
MachineSDNode *emitPCMPISTR(unsigned ROpc, unsigned MOpc, bool MayFoldLoad,
const SDLoc &dl, MVT VT, SDNode *Node);
@@ -5403,6 +5404,124 @@ bool X86DAGToDAGISel::tryMatchBitSelect(SDNode *N) {
Ternlog.getNode(), A, B, C, 0xCA);
}
+bool X86DAGToDAGISel::tryMatchHRegisterInsert(SDNode *N) {
+ if (Subtarget->is64Bit())
+ return false;
+
+ if (N->getOpcode() != ISD::OR)
+ return false;
+ MVT VT = N->getSimpleValueType(0);
+ // On 32-bit x86, i64 is expanded to two i32 halves and i16 is promoted to i32
+ // during type legalization. So we only ever see i32 here.
+ if (VT != MVT::i32)
+ return false;
+
+ SDValue LHS = N->getOperand(0);
+ SDValue RHS = N->getOperand(1);
+
+ bool ZeroExtendLo = false;
+ SDValue LoBase;
+ auto isLoBase = [&](SDValue V) {
+ if (V.getOpcode() == ISD::AND) {
+ auto *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
+ if (!C)
+ return false;
+ uint64_t Mask = C->getZExtValue();
+ if (Mask == 0xFF) {
+ ZeroExtendLo = true;
+ LoBase = V.getOperand(0);
+ return true;
+ }
+ unsigned BitWidth = V.getValueSizeInBits();
+ uint64_t NotFF00 = (~0xFF00ULL) & maskTrailingOnes<uint64_t>(BitWidth);
+ if (Mask == NotFF00) {
+ ZeroExtendLo = false;
+ LoBase = V.getOperand(0);
+ return true;
+ }
+ } else if (V.getOpcode() == ISD::ZERO_EXTEND && V.getOperand(0).getValueType() == MVT::i8) {
+ ZeroExtendLo = false;
+ LoBase = V;
+ return true;
+ } else if (V.getOpcode() == ISD::ANY_EXTEND && V.getOperand(0).getValueType() == MVT::i8) {
+ ZeroExtendLo = true;
+ LoBase = V.getOperand(0);
+ return true;
+ }
+ return false;
+ };
+
+ SDValue HiBase;
+ auto isShl8 = [&](SDValue V) {
+ if (V.getOpcode() != ISD::SHL)
+ return false;
+ auto *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
+ if (!C || C->getZExtValue() != 8)
+ return false;
+ SDValue Src = V.getOperand(0);
+ if (Src.getOpcode() == ISD::AND) {
+ if (auto *Mask = dyn_cast<ConstantSDNode>(Src.getOperand(1))) {
+ if (Mask->getZExtValue() == 255)
+ Src = Src.getOperand(0);
+ }
+ } else if (Src.getOpcode() == ISD::ZERO_EXTEND || Src.getOpcode() == ISD::ANY_EXTEND) {
+ if (Src.getOperand(0).getValueType() == MVT::i8)
+ Src = Src.getOperand(0);
+ }
+ // Make sure the shifted value has only one use, so we don't duplicate work.
+ if (!V.hasOneUse())
+ return false;
+ HiBase = Src;
+ return true;
+ };
+
+ if (isLoBase(LHS) && isShl8(RHS)) {
+ } else if (isLoBase(RHS) && isShl8(LHS)) {
+ } else {
+ return false;
+ }
+
+ SDLoc dl(N);
+ SDValue Sub8 = CurDAG->getTargetConstant(X86::sub_8bit, dl, MVT::i32);
+ SDValue RC_ABCD = CurDAG->getTargetConstant(X86::GR32_ABCDRegClassID, dl, MVT::i32);
+
+ SDValue DestBase;
+ if (ZeroExtendLo) {
+ SDValue LoExtr;
+ if (LoBase.getValueType() == MVT::i8) {
+ LoExtr = LoBase;
+ } else {
+ LoExtr = SDValue(CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl,
+ MVT::i8, LoBase, Sub8), 0);
+ }
+ DestBase = SDValue(
+ CurDAG->getMachineNode(X86::MOVZX32rr8, dl, MVT::i32, LoExtr), 0);
+ DestBase =
+ SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl,
+ MVT::i32, DestBase, RC_ABCD),
+ 0);
+ } else {
+ DestBase = SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
+ dl, MVT::i32, LoBase, RC_ABCD),
+ 0);
+ }
+
+ SDValue HiExtr1;
+ if (HiBase.getValueType() == MVT::i8) {
+ HiExtr1 = HiBase;
+ } else {
+ HiExtr1 = SDValue(CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
+ dl, MVT::i8, HiBase, Sub8), 0);
+ }
+
+ SDValue Sub8Hi = CurDAG->getTargetConstant(X86::sub_8bit_hi, dl, MVT::i32);
+ SDNode *Insert = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT,
+ DestBase, HiExtr1, Sub8Hi);
+
+ ReplaceNode(N, Insert);
+ return true;
+}
+
void X86DAGToDAGISel::Select(SDNode *Node) {
MVT NVT = Node->getSimpleValueType(0);
unsigned Opcode = Node->getOpcode();
@@ -5715,6 +5834,8 @@ void X86DAGToDAGISel::Select(SDNode *Node) {
return;
if (Opcode == ISD::OR && tryMatchBitSelect(Node))
return;
+ if (Opcode == ISD::OR && tryMatchHRegisterInsert(Node))
+ return;
if (tryVPTERNLOG(Node))
return;
diff --git a/llvm/test/CodeGen/X86/h-register-insert.ll b/llvm/test/CodeGen/X86/h-register-insert.ll
index 77c1d8d6ffa24..a5d8d825d1560 100644
--- a/llvm/test/CodeGen/X86/h-register-insert.ll
+++ b/llvm/test/CodeGen/X86/h-register-insert.ll
@@ -5,10 +5,8 @@
define i32 @test_insert_32(i8 inreg %A, i8 inreg %B) {
; X86-LABEL: test_insert_32:
; X86: # %bb.0:
-; X86-NEXT: movzbl %al, %ecx
-; X86-NEXT: movzbl %dl, %eax
-; X86-NEXT: shll $8, %eax
-; X86-NEXT: orl %ecx, %eax
+; X86-NEXT: movzbl %al, %eax
+; X86-NEXT: movb %dl, %ah
; X86-NEXT: retl
;
; X64-LABEL: test_insert_32:
@@ -29,8 +27,7 @@ define i16 @test_insert_16(i8 inreg %A, i8 inreg %B) {
; X86-LABEL: test_insert_16:
; X86: # %bb.0:
; X86-NEXT: movzbl %al, %eax
-; X86-NEXT: shll $8, %edx
-; X86-NEXT: orl %edx, %eax
+; X86-NEXT: movb %dl, %ah
; X86-NEXT: # kill: def $ax killed $ax killed $eax
; X86-NEXT: retl
;
@@ -51,10 +48,8 @@ define i16 @test_insert_16(i8 inreg %A, i8 inreg %B) {
define i64 @test_insert_64(i8 inreg %A, i8 inreg %B) {
; X86-LABEL: test_insert_64:
; X86: # %bb.0:
-; X86-NEXT: movzbl %al, %ecx
-; X86-NEXT: movzbl %dl, %eax
-; X86-NEXT: shll $8, %eax
-; X86-NEXT: orl %ecx, %eax
+; X86-NEXT: movzbl %al, %eax
+; X86-NEXT: movb %dl, %ah
; X86-NEXT: xorl %edx, %edx
; X86-NEXT: retl
;
@@ -71,3 +66,25 @@ define i64 @test_insert_64(i8 inreg %A, i8 inreg %B) {
%or = or disjoint i64 %shl, %conv
ret i64 %or
}
+
+define i32 @test_insert_mask_32(i32 %a, i8 inreg %b) {
+; X86-LABEL: test_insert_mask_32:
+; X86: # %bb.0:
+; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT: movb %al, %ch
+; X86-NEXT: movl %ecx, %eax
+; X86-NEXT: retl
+;
+; X64-LABEL: test_insert_mask_32:
+; X64: # %bb.0:
+; X64-NEXT: andl $-65281, %edi # imm = 0xFFFF00FF
+; X64-NEXT: movzbl %sil, %eax
+; X64-NEXT: shll $8, %eax
+; X64-NEXT: orl %edi, %eax
+; X64-NEXT: retq
+ %mask = and i32 %a, 4294902015 ; 0xFFFF00FF
+ %conv = zext i8 %b to i32
+ %shl = shl nuw nsw i32 %conv, 8
+ %or = or disjoint i32 %mask, %shl
+ ret i32 %or
+}
More information about the llvm-commits
mailing list