[X86] shouldFoldMaskToVariableShiftPair should be true for scalars up 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.

Author: AZero13

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -3659,11 +3659,8 @@ 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

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

@@ -0,0 +1,212 @@
+; 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-NOBMI
+; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,+tbm,+bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X86-BMI2
+; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,-tbm,+bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X86-BMI2
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=-bmi,-tbm,-bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X64-NOBMI
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,+tbm,+bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X64-BMI2
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,-tbm,+bmi2,+fast-bextr < %s | FileCheck %s --check-prefixes=X64-BMI2
+
+define i32 @mask_pair(i32 %x, i32 %y) nounwind {
+; 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-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-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) nounwind {
+; 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-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-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) nounwind {
+; X86-NOBMI-LABEL: mask_pair_128:
+; X86-NOBMI:       # %bb.0:
+; X86-NOBMI-NEXT:    pushl %ebx
+; X86-NOBMI-NEXT:    pushl %edi
+; X86-NOBMI-NEXT:    pushl %esi
+; X86-NOBMI-NEXT:    subl $32, %esp
+; 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:    popl %esi
+; X86-NOBMI-NEXT:    popl %edi
+; X86-NOBMI-NEXT:    popl %ebx
+; X86-NOBMI-NEXT:    retl $4
+;
+; X86-BMI2-LABEL: mask_pair_128:
+; X86-BMI2:       # %bb.0:
+; X86-BMI2-NEXT:    pushl %ebx
+; X86-BMI2-NEXT:    pushl %edi
+; X86-BMI2-NEXT:    pushl %esi
+; X86-BMI2-NEXT:    subl $32, %esp
+; 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:    popl %esi
+; X86-BMI2-NEXT:    popl %edi
+; X86-BMI2-NEXT:    popl %ebx
+; 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-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
+}