[X86] shouldFoldMaskToVariableShiftPair should be true for scalars up to the biggest legal type #158068

AZero13 · 2025-09-11T13:17:43Z

For X86, we want to do this for scalars up to the biggest legal type. Having it be for types bigger results in bloated code.

llvmbot · 2025-09-11T13:18:20Z

@llvm/pr-subscribers-backend-x86

Author: AZero13 (AZero13)

Changes

For X86, we want to do this for scalars up to the biggest legal type. Having it be for types bigger results in bloated code.

Full diff: https://github.com/llvm/llvm-project/pull/158068.diff

2 Files Affected:

(modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+2-5)
(added) llvm/test/CodeGen/X86/and-mask-variable.ll (+356)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 08ae0d52d795e..a5dc7fae4a12a 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -3663,11 +3663,8 @@ bool X86TargetLowering::shouldFoldMaskToVariableShiftPair(SDValue Y) const {
   if (VT.isVector())
     return false;
 
-  // 64-bit shifts on 32-bit targets produce really bad bloated code.
-  if (VT == MVT::i64 && !Subtarget.is64Bit())
-    return false;
-
-  return true;
+  unsigned MaxWidth = Subtarget.is64Bit() ? 64 : 32;
+  return VT.getScalarSizeInBits() <= MaxWidth;
 }
 
 TargetLowering::ShiftLegalizationStrategy
diff --git a/llvm/test/CodeGen/X86/and-mask-variable.ll b/llvm/test/CodeGen/X86/and-mask-variable.ll
new file mode 100644
index 0000000000000..844a413391d75
--- /dev/null
+++ b/llvm/test/CodeGen/X86/and-mask-variable.ll
@@ -0,0 +1,356 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=-bmi,-tbm,-bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X86,X86-NOBMI
+; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,-tbm,-bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X86,X86-BMINOTBM,X86-BMI1
+; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,+tbm,-bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X86,X86-BMITBM,X86-BMI1
+; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,+tbm,+bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X86,X86-BMITBM,X86-BMI2
+; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,-tbm,+bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X86,X86-BMINOTBM,X86-BMI2
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=-bmi,-tbm,-bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X64,X64-NOBMI
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,-tbm,-bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X64,X64-BMINOTBM,X64-BMI1
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,+tbm,-bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X64,X64-BMITBM,X64-BMI1
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,+tbm,+bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X64,X64-BMITBM,X64-BMI2
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,-tbm,+bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X64,X64-BMINOTBM,X64-BMI2
+
+define i32 @mask_pair(i32 %x, i32 %y) {
+; X86-NOBMI-LABEL: mask_pair:
+; X86-NOBMI:       # %bb.0:
+; X86-NOBMI-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
+; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NOBMI-NEXT:    shrl %cl, %eax
+; X86-NOBMI-NEXT:    shll %cl, %eax
+; X86-NOBMI-NEXT:    retl
+;
+; X86-BMI1-LABEL: mask_pair:
+; X86-BMI1:       # %bb.0:
+; X86-BMI1-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
+; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-BMI1-NEXT:    shrl %cl, %eax
+; X86-BMI1-NEXT:    shll %cl, %eax
+; X86-BMI1-NEXT:    retl
+;
+; X86-BMI2-LABEL: mask_pair:
+; X86-BMI2:       # %bb.0:
+; X86-BMI2-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
+; X86-BMI2-NEXT:    shrxl %eax, {{[0-9]+}}(%esp), %ecx
+; X86-BMI2-NEXT:    shlxl %eax, %ecx, %eax
+; X86-BMI2-NEXT:    retl
+;
+; X64-NOBMI-LABEL: mask_pair:
+; X64-NOBMI:       # %bb.0:
+; X64-NOBMI-NEXT:    movl %esi, %ecx
+; X64-NOBMI-NEXT:    movl %edi, %eax
+; X64-NOBMI-NEXT:    shrl %cl, %eax
+; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
+; X64-NOBMI-NEXT:    shll %cl, %eax
+; X64-NOBMI-NEXT:    retq
+;
+; X64-BMI1-LABEL: mask_pair:
+; X64-BMI1:       # %bb.0:
+; X64-BMI1-NEXT:    movl %esi, %ecx
+; X64-BMI1-NEXT:    movl %edi, %eax
+; X64-BMI1-NEXT:    shrl %cl, %eax
+; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
+; X64-BMI1-NEXT:    shll %cl, %eax
+; X64-BMI1-NEXT:    retq
+;
+; X64-BMI2-LABEL: mask_pair:
+; X64-BMI2:       # %bb.0:
+; X64-BMI2-NEXT:    shrxl %esi, %edi, %eax
+; X64-BMI2-NEXT:    shlxl %esi, %eax, %eax
+; X64-BMI2-NEXT:    retq
+  %shl = shl nsw i32 -1, %y
+  %and = and i32 %shl, %x
+  ret i32 %and
+}
+
+define i64 @mask_pair_64(i64 %x, i64 %y) {
+; X86-NOBMI-LABEL: mask_pair_64:
+; X86-NOBMI:       # %bb.0:
+; X86-NOBMI-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
+; X86-NOBMI-NEXT:    movl $-1, %edx
+; X86-NOBMI-NEXT:    movl $-1, %eax
+; X86-NOBMI-NEXT:    shll %cl, %eax
+; X86-NOBMI-NEXT:    testb $32, %cl
+; X86-NOBMI-NEXT:    je .LBB1_2
+; X86-NOBMI-NEXT:  # %bb.1:
+; X86-NOBMI-NEXT:    movl %eax, %edx
+; X86-NOBMI-NEXT:    xorl %eax, %eax
+; X86-NOBMI-NEXT:  .LBB1_2:
+; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
+; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edx
+; X86-NOBMI-NEXT:    retl
+;
+; X86-BMI1-LABEL: mask_pair_64:
+; X86-BMI1:       # %bb.0:
+; X86-BMI1-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
+; X86-BMI1-NEXT:    movl $-1, %edx
+; X86-BMI1-NEXT:    movl $-1, %eax
+; X86-BMI1-NEXT:    shll %cl, %eax
+; X86-BMI1-NEXT:    testb $32, %cl
+; X86-BMI1-NEXT:    je .LBB1_2
+; X86-BMI1-NEXT:  # %bb.1:
+; X86-BMI1-NEXT:    movl %eax, %edx
+; X86-BMI1-NEXT:    xorl %eax, %eax
+; X86-BMI1-NEXT:  .LBB1_2:
+; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %eax
+; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %edx
+; X86-BMI1-NEXT:    retl
+;
+; X86-BMI2-LABEL: mask_pair_64:
+; X86-BMI2:       # %bb.0:
+; X86-BMI2-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
+; X86-BMI2-NEXT:    movl $-1, %edx
+; X86-BMI2-NEXT:    shlxl %ecx, %edx, %eax
+; X86-BMI2-NEXT:    testb $32, %cl
+; X86-BMI2-NEXT:    je .LBB1_2
+; X86-BMI2-NEXT:  # %bb.1:
+; X86-BMI2-NEXT:    movl %eax, %edx
+; X86-BMI2-NEXT:    xorl %eax, %eax
+; X86-BMI2-NEXT:  .LBB1_2:
+; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %eax
+; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %edx
+; X86-BMI2-NEXT:    retl
+;
+; X64-NOBMI-LABEL: mask_pair_64:
+; X64-NOBMI:       # %bb.0:
+; X64-NOBMI-NEXT:    movq %rsi, %rcx
+; X64-NOBMI-NEXT:    movq %rdi, %rax
+; X64-NOBMI-NEXT:    shrq %cl, %rax
+; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
+; X64-NOBMI-NEXT:    shlq %cl, %rax
+; X64-NOBMI-NEXT:    retq
+;
+; X64-BMI1-LABEL: mask_pair_64:
+; X64-BMI1:       # %bb.0:
+; X64-BMI1-NEXT:    movq %rsi, %rcx
+; X64-BMI1-NEXT:    movq %rdi, %rax
+; X64-BMI1-NEXT:    shrq %cl, %rax
+; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $rcx
+; X64-BMI1-NEXT:    shlq %cl, %rax
+; X64-BMI1-NEXT:    retq
+;
+; X64-BMI2-LABEL: mask_pair_64:
+; X64-BMI2:       # %bb.0:
+; X64-BMI2-NEXT:    shrxq %rsi, %rdi, %rax
+; X64-BMI2-NEXT:    shlxq %rsi, %rax, %rax
+; X64-BMI2-NEXT:    retq
+  %shl = shl nsw i64 -1, %y
+  %and = and i64 %shl, %x
+  ret i64 %and
+}
+
+define i128 @mask_pair_128(i128 %x, i128 %y) {
+; X86-NOBMI-LABEL: mask_pair_128:
+; X86-NOBMI:       # %bb.0:
+; X86-NOBMI-NEXT:    pushl %ebx
+; X86-NOBMI-NEXT:    .cfi_def_cfa_offset 8
+; X86-NOBMI-NEXT:    pushl %edi
+; X86-NOBMI-NEXT:    .cfi_def_cfa_offset 12
+; X86-NOBMI-NEXT:    pushl %esi
+; X86-NOBMI-NEXT:    .cfi_def_cfa_offset 16
+; X86-NOBMI-NEXT:    subl $32, %esp
+; X86-NOBMI-NEXT:    .cfi_def_cfa_offset 48
+; X86-NOBMI-NEXT:    .cfi_offset %esi, -16
+; X86-NOBMI-NEXT:    .cfi_offset %edi, -12
+; X86-NOBMI-NEXT:    .cfi_offset %ebx, -8
+; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NOBMI-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-NOBMI-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-NOBMI-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-NOBMI-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-NOBMI-NEXT:    movl $0, {{[0-9]+}}(%esp)
+; X86-NOBMI-NEXT:    movl $0, {{[0-9]+}}(%esp)
+; X86-NOBMI-NEXT:    movl $0, {{[0-9]+}}(%esp)
+; X86-NOBMI-NEXT:    movl $0, (%esp)
+; X86-NOBMI-NEXT:    movl %ecx, %edx
+; X86-NOBMI-NEXT:    shrb $3, %dl
+; X86-NOBMI-NEXT:    andb $12, %dl
+; X86-NOBMI-NEXT:    negb %dl
+; X86-NOBMI-NEXT:    movsbl %dl, %ebx
+; X86-NOBMI-NEXT:    movl 24(%esp,%ebx), %edx
+; X86-NOBMI-NEXT:    movl 28(%esp,%ebx), %esi
+; X86-NOBMI-NEXT:    shldl %cl, %edx, %esi
+; X86-NOBMI-NEXT:    movl 16(%esp,%ebx), %edi
+; X86-NOBMI-NEXT:    movl 20(%esp,%ebx), %ebx
+; X86-NOBMI-NEXT:    shldl %cl, %ebx, %edx
+; X86-NOBMI-NEXT:    shldl %cl, %edi, %ebx
+; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
+; X86-NOBMI-NEXT:    shll %cl, %edi
+; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edx
+; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %esi
+; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edi
+; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %ebx
+; X86-NOBMI-NEXT:    movl %esi, 12(%eax)
+; X86-NOBMI-NEXT:    movl %edx, 8(%eax)
+; X86-NOBMI-NEXT:    movl %ebx, 4(%eax)
+; X86-NOBMI-NEXT:    movl %edi, (%eax)
+; X86-NOBMI-NEXT:    addl $32, %esp
+; X86-NOBMI-NEXT:    .cfi_def_cfa_offset 16
+; X86-NOBMI-NEXT:    popl %esi
+; X86-NOBMI-NEXT:    .cfi_def_cfa_offset 12
+; X86-NOBMI-NEXT:    popl %edi
+; X86-NOBMI-NEXT:    .cfi_def_cfa_offset 8
+; X86-NOBMI-NEXT:    popl %ebx
+; X86-NOBMI-NEXT:    .cfi_def_cfa_offset 4
+; X86-NOBMI-NEXT:    retl $4
+;
+; X86-BMI1-LABEL: mask_pair_128:
+; X86-BMI1:       # %bb.0:
+; X86-BMI1-NEXT:    pushl %ebx
+; X86-BMI1-NEXT:    .cfi_def_cfa_offset 8
+; X86-BMI1-NEXT:    pushl %edi
+; X86-BMI1-NEXT:    .cfi_def_cfa_offset 12
+; X86-BMI1-NEXT:    pushl %esi
+; X86-BMI1-NEXT:    .cfi_def_cfa_offset 16
+; X86-BMI1-NEXT:    subl $32, %esp
+; X86-BMI1-NEXT:    .cfi_def_cfa_offset 48
+; X86-BMI1-NEXT:    .cfi_offset %esi, -16
+; X86-BMI1-NEXT:    .cfi_offset %edi, -12
+; X86-BMI1-NEXT:    .cfi_offset %ebx, -8
+; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-BMI1-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-BMI1-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-BMI1-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-BMI1-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-BMI1-NEXT:    movl $0, {{[0-9]+}}(%esp)
+; X86-BMI1-NEXT:    movl $0, {{[0-9]+}}(%esp)
+; X86-BMI1-NEXT:    movl $0, {{[0-9]+}}(%esp)
+; X86-BMI1-NEXT:    movl $0, (%esp)
+; X86-BMI1-NEXT:    movl %ecx, %edx
+; X86-BMI1-NEXT:    shrb $3, %dl
+; X86-BMI1-NEXT:    andb $12, %dl
+; X86-BMI1-NEXT:    negb %dl
+; X86-BMI1-NEXT:    movsbl %dl, %ebx
+; X86-BMI1-NEXT:    movl 24(%esp,%ebx), %edx
+; X86-BMI1-NEXT:    movl 28(%esp,%ebx), %esi
+; X86-BMI1-NEXT:    shldl %cl, %edx, %esi
+; X86-BMI1-NEXT:    movl 16(%esp,%ebx), %edi
+; X86-BMI1-NEXT:    movl 20(%esp,%ebx), %ebx
+; X86-BMI1-NEXT:    shldl %cl, %ebx, %edx
+; X86-BMI1-NEXT:    shldl %cl, %edi, %ebx
+; X86-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
+; X86-BMI1-NEXT:    shll %cl, %edi
+; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %edx
+; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %esi
+; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %edi
+; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %ebx
+; X86-BMI1-NEXT:    movl %esi, 12(%eax)
+; X86-BMI1-NEXT:    movl %edx, 8(%eax)
+; X86-BMI1-NEXT:    movl %ebx, 4(%eax)
+; X86-BMI1-NEXT:    movl %edi, (%eax)
+; X86-BMI1-NEXT:    addl $32, %esp
+; X86-BMI1-NEXT:    .cfi_def_cfa_offset 16
+; X86-BMI1-NEXT:    popl %esi
+; X86-BMI1-NEXT:    .cfi_def_cfa_offset 12
+; X86-BMI1-NEXT:    popl %edi
+; X86-BMI1-NEXT:    .cfi_def_cfa_offset 8
+; X86-BMI1-NEXT:    popl %ebx
+; X86-BMI1-NEXT:    .cfi_def_cfa_offset 4
+; X86-BMI1-NEXT:    retl $4
+;
+; X86-BMI2-LABEL: mask_pair_128:
+; X86-BMI2:       # %bb.0:
+; X86-BMI2-NEXT:    pushl %ebx
+; X86-BMI2-NEXT:    .cfi_def_cfa_offset 8
+; X86-BMI2-NEXT:    pushl %edi
+; X86-BMI2-NEXT:    .cfi_def_cfa_offset 12
+; X86-BMI2-NEXT:    pushl %esi
+; X86-BMI2-NEXT:    .cfi_def_cfa_offset 16
+; X86-BMI2-NEXT:    subl $32, %esp
+; X86-BMI2-NEXT:    .cfi_def_cfa_offset 48
+; X86-BMI2-NEXT:    .cfi_offset %esi, -16
+; X86-BMI2-NEXT:    .cfi_offset %edi, -12
+; X86-BMI2-NEXT:    .cfi_offset %ebx, -8
+; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-BMI2-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-BMI2-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-BMI2-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-BMI2-NEXT:    movl $-1, {{[0-9]+}}(%esp)
+; X86-BMI2-NEXT:    movl $0, {{[0-9]+}}(%esp)
+; X86-BMI2-NEXT:    movl $0, {{[0-9]+}}(%esp)
+; X86-BMI2-NEXT:    movl $0, {{[0-9]+}}(%esp)
+; X86-BMI2-NEXT:    movl $0, (%esp)
+; X86-BMI2-NEXT:    movl %ecx, %edx
+; X86-BMI2-NEXT:    shrb $3, %dl
+; X86-BMI2-NEXT:    andb $12, %dl
+; X86-BMI2-NEXT:    negb %dl
+; X86-BMI2-NEXT:    movsbl %dl, %edi
+; X86-BMI2-NEXT:    movl 24(%esp,%edi), %edx
+; X86-BMI2-NEXT:    movl 28(%esp,%edi), %esi
+; X86-BMI2-NEXT:    shldl %cl, %edx, %esi
+; X86-BMI2-NEXT:    movl 16(%esp,%edi), %ebx
+; X86-BMI2-NEXT:    movl 20(%esp,%edi), %edi
+; X86-BMI2-NEXT:    shldl %cl, %edi, %edx
+; X86-BMI2-NEXT:    shldl %cl, %ebx, %edi
+; X86-BMI2-NEXT:    shlxl %ecx, %ebx, %ecx
+; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %edx
+; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %esi
+; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %ecx
+; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %edi
+; X86-BMI2-NEXT:    movl %esi, 12(%eax)
+; X86-BMI2-NEXT:    movl %edx, 8(%eax)
+; X86-BMI2-NEXT:    movl %edi, 4(%eax)
+; X86-BMI2-NEXT:    movl %ecx, (%eax)
+; X86-BMI2-NEXT:    addl $32, %esp
+; X86-BMI2-NEXT:    .cfi_def_cfa_offset 16
+; X86-BMI2-NEXT:    popl %esi
+; X86-BMI2-NEXT:    .cfi_def_cfa_offset 12
+; X86-BMI2-NEXT:    popl %edi
+; X86-BMI2-NEXT:    .cfi_def_cfa_offset 8
+; X86-BMI2-NEXT:    popl %ebx
+; X86-BMI2-NEXT:    .cfi_def_cfa_offset 4
+; X86-BMI2-NEXT:    retl $4
+;
+; X64-NOBMI-LABEL: mask_pair_128:
+; X64-NOBMI:       # %bb.0:
+; X64-NOBMI-NEXT:    movq %rdx, %rcx
+; X64-NOBMI-NEXT:    movq $-1, %rdx
+; X64-NOBMI-NEXT:    movq $-1, %r8
+; X64-NOBMI-NEXT:    shlq %cl, %r8
+; X64-NOBMI-NEXT:    xorl %eax, %eax
+; X64-NOBMI-NEXT:    testb $64, %cl
+; X64-NOBMI-NEXT:    cmovneq %r8, %rdx
+; X64-NOBMI-NEXT:    cmoveq %r8, %rax
+; X64-NOBMI-NEXT:    andq %rdi, %rax
+; X64-NOBMI-NEXT:    andq %rsi, %rdx
+; X64-NOBMI-NEXT:    retq
+;
+; X64-BMI1-LABEL: mask_pair_128:
+; X64-BMI1:       # %bb.0:
+; X64-BMI1-NEXT:    movq %rdx, %rcx
+; X64-BMI1-NEXT:    movq $-1, %rdx
+; X64-BMI1-NEXT:    movq $-1, %r8
+; X64-BMI1-NEXT:    shlq %cl, %r8
+; X64-BMI1-NEXT:    xorl %eax, %eax
+; X64-BMI1-NEXT:    testb $64, %cl
+; X64-BMI1-NEXT:    cmovneq %r8, %rdx
+; X64-BMI1-NEXT:    cmoveq %r8, %rax
+; X64-BMI1-NEXT:    andq %rdi, %rax
+; X64-BMI1-NEXT:    andq %rsi, %rdx
+; X64-BMI1-NEXT:    retq
+;
+; X64-BMI2-LABEL: mask_pair_128:
+; X64-BMI2:       # %bb.0:
+; X64-BMI2-NEXT:    movq $-1, %rcx
+; X64-BMI2-NEXT:    shlxq %rdx, %rcx, %r8
+; X64-BMI2-NEXT:    xorl %eax, %eax
+; X64-BMI2-NEXT:    testb $64, %dl
+; X64-BMI2-NEXT:    cmovneq %r8, %rcx
+; X64-BMI2-NEXT:    cmoveq %r8, %rax
+; X64-BMI2-NEXT:    andq %rdi, %rax
+; X64-BMI2-NEXT:    andq %rsi, %rcx
+; X64-BMI2-NEXT:    movq %rcx, %rdx
+; X64-BMI2-NEXT:    retq
+  %shl = shl nsw i128 -1, %y
+  %and = and i128 %shl, %x
+  ret i128 %and
+}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; X64: {{.*}}
+; X64-BMINOTBM: {{.*}}
+; X64-BMITBM: {{.*}}
+; X86: {{.*}}
+; X86-BMINOTBM: {{.*}}
+; X86-BMITBM: {{.*}}

AZero13 · 2025-09-11T13:27:06Z

@topperc

AZero13 · 2025-10-03T14:24:39Z

@phoebewang Thoughts?

llvm/test/CodeGen/X86/and-mask-variable.ll

llvm/lib/Target/X86/X86ISelLowering.cpp

AZero13 · 2025-10-07T14:26:38Z

@phoebewang Fixed.

llvm/test/CodeGen/X86/and-mask-variable.ll

… to the biggest legal type For X86, we want to do this for scalars up to the biggest legal type.

… to the biggest legal type (#158068) For X86, we want to do this for scalars up to the biggest legal type. Having it be for types bigger results in bloated code.

… to the biggest legal type (llvm#158068) For X86, we want to do this for scalars up to the biggest legal type. Having it be for types bigger results in bloated code.

llvmbot added the backend:X86 label Sep 11, 2025

AZero13 mentioned this pull request Sep 11, 2025

[X86][ARM][AArch64] shouldFoldMaskToVariableShiftPair should be true for scalars up to the biggest legal type #156886

Closed

phoebewang requested review from RKSimon and phoebewang October 4, 2025 01:33

phoebewang reviewed Oct 4, 2025

View reviewed changes

llvm/test/CodeGen/X86/and-mask-variable.ll Outdated Show resolved Hide resolved

phoebewang reviewed Oct 4, 2025

View reviewed changes

llvm/test/CodeGen/X86/and-mask-variable.ll Outdated Show resolved Hide resolved

phoebewang reviewed Oct 4, 2025

View reviewed changes

llvm/lib/Target/X86/X86ISelLowering.cpp Show resolved Hide resolved

AZero13 force-pushed the x86-shifty branch 2 times, most recently from 47a6e2d to f16e4cf Compare October 7, 2025 14:25

AZero13 requested a review from phoebewang October 7, 2025 14:26

phoebewang reviewed Oct 8, 2025

View reviewed changes

llvm/test/CodeGen/X86/and-mask-variable.ll Show resolved Hide resolved

phoebewang reviewed Oct 8, 2025

View reviewed changes

llvm/test/CodeGen/X86/and-mask-variable.ll Outdated Show resolved Hide resolved

AZero13 requested a review from phoebewang October 8, 2025 20:30

AZero13 force-pushed the x86-shifty branch from 3a297ff to cf8d8ba Compare October 8, 2025 20:33

AZero13 added 2 commits October 8, 2025 17:03

Pre-commit test (NFC)

23c1ed5

[X86] shouldFoldMaskToVariableShiftPair should be true for scalars up…

7bd2b3d

… to the biggest legal type For X86, we want to do this for scalars up to the biggest legal type.

AZero13 force-pushed the x86-shifty branch from cf8d8ba to 7bd2b3d Compare October 8, 2025 21:03

phoebewang approved these changes Oct 9, 2025

View reviewed changes

phoebewang merged commit a4a9803 into llvm:main Oct 9, 2025
7 of 8 checks passed

AZero13 deleted the x86-shifty branch October 9, 2025 02:04

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[X86] shouldFoldMaskToVariableShiftPair should be true for scalars up to the biggest legal type #158068

[X86] shouldFoldMaskToVariableShiftPair should be true for scalars up to the biggest legal type #158068

Uh oh!

AZero13 commented Sep 11, 2025

Uh oh!

llvmbot commented Sep 11, 2025

Uh oh!

AZero13 commented Sep 11, 2025

Uh oh!

AZero13 commented Oct 3, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

AZero13 commented Oct 7, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

[X86] shouldFoldMaskToVariableShiftPair should be true for scalars up to the biggest legal type #158068

[X86] shouldFoldMaskToVariableShiftPair should be true for scalars up to the biggest legal type #158068

Uh oh!

Conversation

AZero13 commented Sep 11, 2025

Uh oh!

llvmbot commented Sep 11, 2025

Uh oh!

AZero13 commented Sep 11, 2025

Uh oh!

AZero13 commented Oct 3, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

AZero13 commented Oct 7, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants