llvm-for-llvmta/test/CodeGen/X86/extract-lowbits.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=-bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X86,X86-NOBMI
; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X86,X86-BMI1
; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,+tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X86,X86-BMI1
; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,+tbm,+bmi2 < %s | FileCheck %s --check-prefixes=X86,X86-BMI2
; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,-tbm,+bmi2 < %s | FileCheck %s --check-prefixes=X86,X86-BMI2
; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=-bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X64,X64-NOBMI
; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X64,X64-BMI1,X64-BMI1NOTBM
; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,+tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X64,X64-BMI1,X64-BMI1TBM
; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,+tbm,+bmi2 < %s | FileCheck %s --check-prefixes=X64,X64-BMI2,X64-BMI2TBM
; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,-tbm,+bmi2 < %s | FileCheck %s --check-prefixes=X64,X64-BMI2,X64-BMI2NOTBM

; *Please* keep in sync with test/CodeGen/AArch64/extract-lowbits.ll

; https://bugs.llvm.org/show_bug.cgi?id=36419
; https://bugs.llvm.org/show_bug.cgi?id=37603
; https://bugs.llvm.org/show_bug.cgi?id=37610

; Patterns:
;   a) x &  (1 << nbits) - 1
;   b) x & ~(-1 << nbits)
;   c) x &  (-1 >> (32 - y))
;   d) x << (32 - y) >> (32 - y)
; are equivalent.

; ---------------------------------------------------------------------------- ;
; Pattern a. 32-bit
; ---------------------------------------------------------------------------- ;

define i32 @bzhi32_a0(i32 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_a0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    decl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_a0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_a0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_a0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    decl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_a0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_a0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %onebit = shl i32 1, %numlowbits
  %mask = add nsw i32 %onebit, -1
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_a1_indexzext(i32 %val, i8 zeroext %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_a1_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    decl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_a1_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_a1_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_a1_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    decl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_a1_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_a1_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %conv = zext i8 %numlowbits to i32
  %onebit = shl i32 1, %conv
  %mask = add nsw i32 %onebit, -1
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_a2_load(i32* %w, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_a2_load:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    decl %eax
; X86-NOBMI-NEXT:    andl (%edx), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_a2_load:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    shll $8, %ecx
; X86-BMI1-NEXT:    bextrl %ecx, (%eax), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_a2_load:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    bzhil %ecx, (%eax), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_a2_load:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    decl %eax
; X64-NOBMI-NEXT:    andl (%rdi), %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_a2_load:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, (%rdi), %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_a2_load:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, (%rdi), %eax
; X64-BMI2-NEXT:    retq
  %val = load i32, i32* %w
  %onebit = shl i32 1, %numlowbits
  %mask = add nsw i32 %onebit, -1
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_a3_load_indexzext(i32* %w, i8 zeroext %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_a3_load_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    decl %eax
; X86-NOBMI-NEXT:    andl (%edx), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_a3_load_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    shll $8, %ecx
; X86-BMI1-NEXT:    bextrl %ecx, (%eax), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_a3_load_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    bzhil %ecx, (%eax), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_a3_load_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    decl %eax
; X64-NOBMI-NEXT:    andl (%rdi), %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_a3_load_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, (%rdi), %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_a3_load_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, (%rdi), %eax
; X64-BMI2-NEXT:    retq
  %val = load i32, i32* %w
  %conv = zext i8 %numlowbits to i32
  %onebit = shl i32 1, %conv
  %mask = add nsw i32 %onebit, -1
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_a4_commutative(i32 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_a4_commutative:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    decl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_a4_commutative:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_a4_commutative:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_a4_commutative:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    decl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_a4_commutative:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_a4_commutative:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %onebit = shl i32 1, %numlowbits
  %mask = add nsw i32 %onebit, -1
  %masked = and i32 %val, %mask ; swapped order
  ret i32 %masked
}

; 64-bit

define i64 @bzhi64_a0(i64 %val, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_a0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    xorl %edx, %edx
; X86-NOBMI-NEXT:    shldl %cl, %eax, %edx
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    je .LBB5_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %eax, %edx
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:  .LBB5_2:
; X86-NOBMI-NEXT:    addl $-1, %eax
; X86-NOBMI-NEXT:    adcl $-1, %edx
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_a0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $1, %eax
; X86-BMI1-NEXT:    xorl %edx, %edx
; X86-BMI1-NEXT:    shldl %cl, %eax, %edx
; X86-BMI1-NEXT:    shll %cl, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB5_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:  .LBB5_2:
; X86-BMI1-NEXT:    addl $-1, %eax
; X86-BMI1-NEXT:    adcl $-1, %edx
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_a0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    movl $1, %eax
; X86-BMI2-NEXT:    xorl %edx, %edx
; X86-BMI2-NEXT:    shldl %cl, %eax, %edx
; X86-BMI2-NEXT:    shlxl %ecx, %eax, %eax
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB5_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:  .LBB5_2:
; X86-BMI2-NEXT:    addl $-1, %eax
; X86-BMI2-NEXT:    adcl $-1, %edx
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_a0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    decq %rax
; X64-NOBMI-NEXT:    andq %rdi, %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_a0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, %rdi, %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_a0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhiq %rsi, %rdi, %rax
; X64-BMI2-NEXT:    retq
  %onebit = shl i64 1, %numlowbits
  %mask = add nsw i64 %onebit, -1
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_a1_indexzext(i64 %val, i8 zeroext %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_a1_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    xorl %edx, %edx
; X86-NOBMI-NEXT:    shldl %cl, %eax, %edx
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    je .LBB6_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %eax, %edx
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:  .LBB6_2:
; X86-NOBMI-NEXT:    addl $-1, %eax
; X86-NOBMI-NEXT:    adcl $-1, %edx
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_a1_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $1, %eax
; X86-BMI1-NEXT:    xorl %edx, %edx
; X86-BMI1-NEXT:    shldl %cl, %eax, %edx
; X86-BMI1-NEXT:    shll %cl, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB6_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:  .LBB6_2:
; X86-BMI1-NEXT:    addl $-1, %eax
; X86-BMI1-NEXT:    adcl $-1, %edx
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_a1_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    movl $1, %eax
; X86-BMI2-NEXT:    xorl %edx, %edx
; X86-BMI2-NEXT:    shldl %cl, %eax, %edx
; X86-BMI2-NEXT:    shlxl %ecx, %eax, %eax
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB6_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:  .LBB6_2:
; X86-BMI2-NEXT:    addl $-1, %eax
; X86-BMI2-NEXT:    adcl $-1, %edx
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_a1_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    decq %rax
; X64-NOBMI-NEXT:    andq %rdi, %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_a1_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, %rdi, %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_a1_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI2-NEXT:    bzhiq %rsi, %rdi, %rax
; X64-BMI2-NEXT:    retq
  %conv = zext i8 %numlowbits to i64
  %onebit = shl i64 1, %conv
  %mask = add nsw i64 %onebit, -1
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_a2_load(i64* %w, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_a2_load:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    xorl %edx, %edx
; X86-NOBMI-NEXT:    shldl %cl, %eax, %edx
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    je .LBB7_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %eax, %edx
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:  .LBB7_2:
; X86-NOBMI-NEXT:    addl $-1, %eax
; X86-NOBMI-NEXT:    adcl $-1, %edx
; X86-NOBMI-NEXT:    andl 4(%esi), %edx
; X86-NOBMI-NEXT:    andl (%esi), %eax
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_a2_load:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $1, %eax
; X86-BMI1-NEXT:    xorl %edx, %edx
; X86-BMI1-NEXT:    shldl %cl, %eax, %edx
; X86-BMI1-NEXT:    shll %cl, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB7_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:  .LBB7_2:
; X86-BMI1-NEXT:    addl $-1, %eax
; X86-BMI1-NEXT:    adcl $-1, %edx
; X86-BMI1-NEXT:    andl 4(%esi), %edx
; X86-BMI1-NEXT:    andl (%esi), %eax
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_a2_load:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    movl $1, %eax
; X86-BMI2-NEXT:    xorl %edx, %edx
; X86-BMI2-NEXT:    shldl %cl, %eax, %edx
; X86-BMI2-NEXT:    shlxl %ecx, %eax, %eax
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB7_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:  .LBB7_2:
; X86-BMI2-NEXT:    addl $-1, %eax
; X86-BMI2-NEXT:    adcl $-1, %edx
; X86-BMI2-NEXT:    andl 4(%esi), %edx
; X86-BMI2-NEXT:    andl (%esi), %eax
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_a2_load:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    decq %rax
; X64-NOBMI-NEXT:    andq (%rdi), %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_a2_load:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, (%rdi), %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_a2_load:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhiq %rsi, (%rdi), %rax
; X64-BMI2-NEXT:    retq
  %val = load i64, i64* %w
  %onebit = shl i64 1, %numlowbits
  %mask = add nsw i64 %onebit, -1
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_a3_load_indexzext(i64* %w, i8 zeroext %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_a3_load_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    xorl %edx, %edx
; X86-NOBMI-NEXT:    shldl %cl, %eax, %edx
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    je .LBB8_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %eax, %edx
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:  .LBB8_2:
; X86-NOBMI-NEXT:    addl $-1, %eax
; X86-NOBMI-NEXT:    adcl $-1, %edx
; X86-NOBMI-NEXT:    andl 4(%esi), %edx
; X86-NOBMI-NEXT:    andl (%esi), %eax
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_a3_load_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $1, %eax
; X86-BMI1-NEXT:    xorl %edx, %edx
; X86-BMI1-NEXT:    shldl %cl, %eax, %edx
; X86-BMI1-NEXT:    shll %cl, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB8_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:  .LBB8_2:
; X86-BMI1-NEXT:    addl $-1, %eax
; X86-BMI1-NEXT:    adcl $-1, %edx
; X86-BMI1-NEXT:    andl 4(%esi), %edx
; X86-BMI1-NEXT:    andl (%esi), %eax
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_a3_load_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    movl $1, %eax
; X86-BMI2-NEXT:    xorl %edx, %edx
; X86-BMI2-NEXT:    shldl %cl, %eax, %edx
; X86-BMI2-NEXT:    shlxl %ecx, %eax, %eax
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB8_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:  .LBB8_2:
; X86-BMI2-NEXT:    addl $-1, %eax
; X86-BMI2-NEXT:    adcl $-1, %edx
; X86-BMI2-NEXT:    andl 4(%esi), %edx
; X86-BMI2-NEXT:    andl (%esi), %eax
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_a3_load_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    decq %rax
; X64-NOBMI-NEXT:    andq (%rdi), %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_a3_load_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, (%rdi), %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_a3_load_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI2-NEXT:    bzhiq %rsi, (%rdi), %rax
; X64-BMI2-NEXT:    retq
  %val = load i64, i64* %w
  %conv = zext i8 %numlowbits to i64
  %onebit = shl i64 1, %conv
  %mask = add nsw i64 %onebit, -1
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_a4_commutative(i64 %val, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_a4_commutative:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    xorl %edx, %edx
; X86-NOBMI-NEXT:    shldl %cl, %eax, %edx
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    je .LBB9_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %eax, %edx
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:  .LBB9_2:
; X86-NOBMI-NEXT:    addl $-1, %eax
; X86-NOBMI-NEXT:    adcl $-1, %edx
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_a4_commutative:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $1, %eax
; X86-BMI1-NEXT:    xorl %edx, %edx
; X86-BMI1-NEXT:    shldl %cl, %eax, %edx
; X86-BMI1-NEXT:    shll %cl, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB9_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:  .LBB9_2:
; X86-BMI1-NEXT:    addl $-1, %eax
; X86-BMI1-NEXT:    adcl $-1, %edx
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_a4_commutative:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    movl $1, %eax
; X86-BMI2-NEXT:    xorl %edx, %edx
; X86-BMI2-NEXT:    shldl %cl, %eax, %edx
; X86-BMI2-NEXT:    shlxl %ecx, %eax, %eax
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB9_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:  .LBB9_2:
; X86-BMI2-NEXT:    addl $-1, %eax
; X86-BMI2-NEXT:    adcl $-1, %edx
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_a4_commutative:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    decq %rax
; X64-NOBMI-NEXT:    andq %rdi, %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_a4_commutative:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, %rdi, %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_a4_commutative:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhiq %rsi, %rdi, %rax
; X64-BMI2-NEXT:    retq
  %onebit = shl i64 1, %numlowbits
  %mask = add nsw i64 %onebit, -1
  %masked = and i64 %val, %mask ; swapped order
  ret i64 %masked
}

; 64-bit, but with 32-bit output

; Everything done in 64-bit, truncation happens last.
define i32 @bzhi64_32_a0(i64 %val, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_a0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %edx
; X86-NOBMI-NEXT:    shll %cl, %edx
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB10_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %edx, %eax
; X86-NOBMI-NEXT:  .LBB10_2:
; X86-NOBMI-NEXT:    decl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_a0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $1, %edx
; X86-BMI1-NEXT:    shll %cl, %edx
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    jne .LBB10_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %edx, %eax
; X86-BMI1-NEXT:  .LBB10_2:
; X86-BMI1-NEXT:    decl %eax
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_a0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    jne .LBB10_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl $1, %eax
; X86-BMI2-NEXT:    shlxl %ecx, %eax, %eax
; X86-BMI2-NEXT:  .LBB10_2:
; X86-BMI2-NEXT:    decl %eax
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_a0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    decl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_a0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_a0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %onebit = shl i64 1, %numlowbits
  %mask = add nsw i64 %onebit, -1
  %masked = and i64 %mask, %val
  %res = trunc i64 %masked to i32
  ret i32 %res
}

; Shifting happens in 64-bit, then truncation. Masking is 32-bit.
define i32 @bzhi64_32_a1(i64 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_a1:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    decl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_a1:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_a1:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_a1:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    decl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_a1:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_a1:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %truncval = trunc i64 %val to i32
  %onebit = shl i32 1, %numlowbits
  %mask = add nsw i32 %onebit, -1
  %masked = and i32 %mask, %truncval
  ret i32 %masked
}

; Shifting happens in 64-bit, then truncation (with extra use).
; Masking is 32-bit.
define i32 @bzhi64_32_a1_trunc_extrause(i64 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_a1_trunc_extrause:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %ebx
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    pushl %eax
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %bl
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    movl %esi, (%esp)
; X86-NOBMI-NEXT:    calll use32@PLT
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    movl %ebx, %ecx
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    decl %eax
; X86-NOBMI-NEXT:    andl %esi, %eax
; X86-NOBMI-NEXT:    addl $4, %esp
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %ebx
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_a1_trunc_extrause:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %ebx
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    pushl %eax
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %bl
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-BMI1-NEXT:    movl %esi, (%esp)
; X86-BMI1-NEXT:    calll use32@PLT
; X86-BMI1-NEXT:    shll $8, %ebx
; X86-BMI1-NEXT:    bextrl %ebx, %esi, %eax
; X86-BMI1-NEXT:    addl $4, %esp
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    popl %ebx
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_a1_trunc_extrause:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %ebx
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    pushl %eax
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %bl
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-BMI2-NEXT:    movl %esi, (%esp)
; X86-BMI2-NEXT:    calll use32@PLT
; X86-BMI2-NEXT:    bzhil %ebx, %esi, %eax
; X86-BMI2-NEXT:    addl $4, %esp
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    popl %ebx
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_a1_trunc_extrause:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    pushq %rbp
; X64-NOBMI-NEXT:    pushq %rbx
; X64-NOBMI-NEXT:    pushq %rax
; X64-NOBMI-NEXT:    movl %esi, %ebp
; X64-NOBMI-NEXT:    movq %rdi, %rbx
; X64-NOBMI-NEXT:    callq use32@PLT
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    movl %ebp, %ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    decl %eax
; X64-NOBMI-NEXT:    andl %ebx, %eax
; X64-NOBMI-NEXT:    addq $8, %rsp
; X64-NOBMI-NEXT:    popq %rbx
; X64-NOBMI-NEXT:    popq %rbp
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_a1_trunc_extrause:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    pushq %r14
; X64-BMI1-NEXT:    pushq %rbx
; X64-BMI1-NEXT:    pushq %rax
; X64-BMI1-NEXT:    movl %esi, %ebx
; X64-BMI1-NEXT:    movq %rdi, %r14
; X64-BMI1-NEXT:    callq use32@PLT
; X64-BMI1-NEXT:    shll $8, %ebx
; X64-BMI1-NEXT:    bextrl %ebx, %r14d, %eax
; X64-BMI1-NEXT:    addq $8, %rsp
; X64-BMI1-NEXT:    popq %rbx
; X64-BMI1-NEXT:    popq %r14
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_a1_trunc_extrause:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    pushq %rbp
; X64-BMI2-NEXT:    pushq %rbx
; X64-BMI2-NEXT:    pushq %rax
; X64-BMI2-NEXT:    movl %esi, %ebp
; X64-BMI2-NEXT:    movq %rdi, %rbx
; X64-BMI2-NEXT:    callq use32@PLT
; X64-BMI2-NEXT:    bzhil %ebp, %ebx, %eax
; X64-BMI2-NEXT:    addq $8, %rsp
; X64-BMI2-NEXT:    popq %rbx
; X64-BMI2-NEXT:    popq %rbp
; X64-BMI2-NEXT:    retq
  %truncval = trunc i64 %val to i32
  call void @use32(i32 %truncval)
  %onebit = shl i32 1, %numlowbits
  %mask = add nsw i32 %onebit, -1
  %masked = and i32 %mask, %truncval
  ret i32 %masked
}

; Shifting happens in 64-bit. Mask is 32-bit, but extended to 64-bit.
; Masking is 64-bit. Then truncation.
define i32 @bzhi64_32_a2(i64 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_a2:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    decl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_a2:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_a2:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_a2:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    decl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_a2:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_a2:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %onebit = shl i32 1, %numlowbits
  %mask = add nsw i32 %onebit, -1
  %zextmask = zext i32 %mask to i64
  %masked = and i64 %zextmask, %val
  %truncmasked = trunc i64 %masked to i32
  ret i32 %truncmasked
}

; Shifting happens in 64-bit. Mask is 32-bit, but calculated in 64-bit.
; Masking is 64-bit. Then truncation.
define i32 @bzhi64_32_a3(i64 %val, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_a3:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $1, %edx
; X86-NOBMI-NEXT:    shll %cl, %edx
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB14_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %edx, %eax
; X86-NOBMI-NEXT:  .LBB14_2:
; X86-NOBMI-NEXT:    decl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_a3:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $1, %edx
; X86-BMI1-NEXT:    shll %cl, %edx
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    jne .LBB14_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %edx, %eax
; X86-BMI1-NEXT:  .LBB14_2:
; X86-BMI1-NEXT:    decl %eax
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_a3:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    jne .LBB14_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl $1, %eax
; X86-BMI2-NEXT:    shlxl %ecx, %eax, %eax
; X86-BMI2-NEXT:  .LBB14_2:
; X86-BMI2-NEXT:    decl %eax
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_a3:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movl $1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    decl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_a3:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_a3:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %onebit = shl i64 1, %numlowbits
  %mask = add nsw i64 %onebit, 4294967295
  %masked = and i64 %mask, %val
  %truncmasked = trunc i64 %masked to i32
  ret i32 %truncmasked
}

; ---------------------------------------------------------------------------- ;
; Pattern b. 32-bit
; ---------------------------------------------------------------------------- ;

define i32 @bzhi32_b0(i32 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_b0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_b0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_b0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_b0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $-1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    notl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_b0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_b0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %notmask = shl i32 -1, %numlowbits
  %mask = xor i32 %notmask, -1
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_b1_indexzext(i32 %val, i8 zeroext %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_b1_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_b1_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_b1_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_b1_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $-1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    notl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_b1_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_b1_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %conv = zext i8 %numlowbits to i32
  %notmask = shl i32 -1, %conv
  %mask = xor i32 %notmask, -1
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_b2_load(i32* %w, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_b2_load:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl (%edx), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_b2_load:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    shll $8, %ecx
; X86-BMI1-NEXT:    bextrl %ecx, (%eax), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_b2_load:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    bzhil %ecx, (%eax), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_b2_load:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $-1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    notl %eax
; X64-NOBMI-NEXT:    andl (%rdi), %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_b2_load:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, (%rdi), %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_b2_load:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, (%rdi), %eax
; X64-BMI2-NEXT:    retq
  %val = load i32, i32* %w
  %notmask = shl i32 -1, %numlowbits
  %mask = xor i32 %notmask, -1
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_b3_load_indexzext(i32* %w, i8 zeroext %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_b3_load_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl (%edx), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_b3_load_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    shll $8, %ecx
; X86-BMI1-NEXT:    bextrl %ecx, (%eax), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_b3_load_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    bzhil %ecx, (%eax), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_b3_load_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $-1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    notl %eax
; X64-NOBMI-NEXT:    andl (%rdi), %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_b3_load_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, (%rdi), %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_b3_load_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, (%rdi), %eax
; X64-BMI2-NEXT:    retq
  %val = load i32, i32* %w
  %conv = zext i8 %numlowbits to i32
  %notmask = shl i32 -1, %conv
  %mask = xor i32 %notmask, -1
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_b4_commutative(i32 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_b4_commutative:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_b4_commutative:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_b4_commutative:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_b4_commutative:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $-1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    notl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_b4_commutative:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_b4_commutative:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %notmask = shl i32 -1, %numlowbits
  %mask = xor i32 %notmask, -1
  %masked = and i32 %val, %mask ; swapped order
  ret i32 %masked
}

; 64-bit

define i64 @bzhi64_b0(i64 %val, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_b0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %edx
; X86-NOBMI-NEXT:    movl $-1, %esi
; X86-NOBMI-NEXT:    shll %cl, %esi
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB20_1
; X86-NOBMI-NEXT:  # %bb.2:
; X86-NOBMI-NEXT:    movl %esi, %eax
; X86-NOBMI-NEXT:    jmp .LBB20_3
; X86-NOBMI-NEXT:  .LBB20_1:
; X86-NOBMI-NEXT:    movl %esi, %edx
; X86-NOBMI-NEXT:  .LBB20_3:
; X86-NOBMI-NEXT:    notl %edx
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_b0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; 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 .LBB20_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:  .LBB20_2:
; X86-BMI1-NEXT:    andnl {{[0-9]+}}(%esp), %eax, %eax
; X86-BMI1-NEXT:    andnl {{[0-9]+}}(%esp), %edx, %edx
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_b0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %dl
; X86-BMI2-NEXT:    movl $-1, %ecx
; X86-BMI2-NEXT:    shlxl %edx, %ecx, %eax
; X86-BMI2-NEXT:    testb $32, %dl
; X86-BMI2-NEXT:    je .LBB20_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %eax, %ecx
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:  .LBB20_2:
; X86-BMI2-NEXT:    andnl {{[0-9]+}}(%esp), %eax, %eax
; X86-BMI2-NEXT:    andnl {{[0-9]+}}(%esp), %ecx, %edx
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_b0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movq $-1, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    notq %rax
; X64-NOBMI-NEXT:    andq %rdi, %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_b0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, %rdi, %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_b0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhiq %rsi, %rdi, %rax
; X64-BMI2-NEXT:    retq
  %notmask = shl i64 -1, %numlowbits
  %mask = xor i64 %notmask, -1
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_b1_indexzext(i64 %val, i8 zeroext %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_b1_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %edx
; X86-NOBMI-NEXT:    movl $-1, %esi
; X86-NOBMI-NEXT:    shll %cl, %esi
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB21_1
; X86-NOBMI-NEXT:  # %bb.2:
; X86-NOBMI-NEXT:    movl %esi, %eax
; X86-NOBMI-NEXT:    jmp .LBB21_3
; X86-NOBMI-NEXT:  .LBB21_1:
; X86-NOBMI-NEXT:    movl %esi, %edx
; X86-NOBMI-NEXT:  .LBB21_3:
; X86-NOBMI-NEXT:    notl %edx
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_b1_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; 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 .LBB21_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:  .LBB21_2:
; X86-BMI1-NEXT:    andnl {{[0-9]+}}(%esp), %eax, %eax
; X86-BMI1-NEXT:    andnl {{[0-9]+}}(%esp), %edx, %edx
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_b1_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %dl
; X86-BMI2-NEXT:    movl $-1, %ecx
; X86-BMI2-NEXT:    shlxl %edx, %ecx, %eax
; X86-BMI2-NEXT:    testb $32, %dl
; X86-BMI2-NEXT:    je .LBB21_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %eax, %ecx
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:  .LBB21_2:
; X86-BMI2-NEXT:    andnl {{[0-9]+}}(%esp), %eax, %eax
; X86-BMI2-NEXT:    andnl {{[0-9]+}}(%esp), %ecx, %edx
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_b1_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movq $-1, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    notq %rax
; X64-NOBMI-NEXT:    andq %rdi, %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_b1_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, %rdi, %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_b1_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI2-NEXT:    bzhiq %rsi, %rdi, %rax
; X64-BMI2-NEXT:    retq
  %conv = zext i8 %numlowbits to i64
  %notmask = shl i64 -1, %conv
  %mask = xor i64 %notmask, -1
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_b2_load(i64* %w, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_b2_load:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %edx
; X86-NOBMI-NEXT:    movl $-1, %edi
; X86-NOBMI-NEXT:    shll %cl, %edi
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB22_1
; X86-NOBMI-NEXT:  # %bb.2:
; X86-NOBMI-NEXT:    movl %edi, %eax
; X86-NOBMI-NEXT:    jmp .LBB22_3
; X86-NOBMI-NEXT:  .LBB22_1:
; X86-NOBMI-NEXT:    movl %edi, %edx
; X86-NOBMI-NEXT:  .LBB22_3:
; X86-NOBMI-NEXT:    notl %edx
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl (%esi), %eax
; X86-NOBMI-NEXT:    andl 4(%esi), %edx
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %edi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_b2_load:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %esi
; X86-BMI1-NEXT:    movl $-1, %eax
; X86-BMI1-NEXT:    shll %cl, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB22_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %esi
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:  .LBB22_2:
; X86-BMI1-NEXT:    andnl (%edx), %eax, %eax
; X86-BMI1-NEXT:    andnl 4(%edx), %esi, %edx
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_b2_load:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %ebx
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %bl
; X86-BMI2-NEXT:    movl $-1, %edx
; X86-BMI2-NEXT:    shlxl %ebx, %edx, %eax
; X86-BMI2-NEXT:    testb $32, %bl
; X86-BMI2-NEXT:    je .LBB22_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:  .LBB22_2:
; X86-BMI2-NEXT:    andnl (%ecx), %eax, %eax
; X86-BMI2-NEXT:    andnl 4(%ecx), %edx, %edx
; X86-BMI2-NEXT:    popl %ebx
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_b2_load:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movq $-1, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    notq %rax
; X64-NOBMI-NEXT:    andq (%rdi), %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_b2_load:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, (%rdi), %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_b2_load:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhiq %rsi, (%rdi), %rax
; X64-BMI2-NEXT:    retq
  %val = load i64, i64* %w
  %notmask = shl i64 -1, %numlowbits
  %mask = xor i64 %notmask, -1
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_b3_load_indexzext(i64* %w, i8 zeroext %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_b3_load_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %edx
; X86-NOBMI-NEXT:    movl $-1, %edi
; X86-NOBMI-NEXT:    shll %cl, %edi
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB23_1
; X86-NOBMI-NEXT:  # %bb.2:
; X86-NOBMI-NEXT:    movl %edi, %eax
; X86-NOBMI-NEXT:    jmp .LBB23_3
; X86-NOBMI-NEXT:  .LBB23_1:
; X86-NOBMI-NEXT:    movl %edi, %edx
; X86-NOBMI-NEXT:  .LBB23_3:
; X86-NOBMI-NEXT:    notl %edx
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl (%esi), %eax
; X86-NOBMI-NEXT:    andl 4(%esi), %edx
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %edi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_b3_load_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %esi
; X86-BMI1-NEXT:    movl $-1, %eax
; X86-BMI1-NEXT:    shll %cl, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB23_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %esi
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:  .LBB23_2:
; X86-BMI1-NEXT:    andnl (%edx), %eax, %eax
; X86-BMI1-NEXT:    andnl 4(%edx), %esi, %edx
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_b3_load_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %ebx
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %bl
; X86-BMI2-NEXT:    movl $-1, %edx
; X86-BMI2-NEXT:    shlxl %ebx, %edx, %eax
; X86-BMI2-NEXT:    testb $32, %bl
; X86-BMI2-NEXT:    je .LBB23_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:  .LBB23_2:
; X86-BMI2-NEXT:    andnl (%ecx), %eax, %eax
; X86-BMI2-NEXT:    andnl 4(%ecx), %edx, %edx
; X86-BMI2-NEXT:    popl %ebx
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_b3_load_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movq $-1, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    notq %rax
; X64-NOBMI-NEXT:    andq (%rdi), %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_b3_load_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, (%rdi), %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_b3_load_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI2-NEXT:    bzhiq %rsi, (%rdi), %rax
; X64-BMI2-NEXT:    retq
  %val = load i64, i64* %w
  %conv = zext i8 %numlowbits to i64
  %notmask = shl i64 -1, %conv
  %mask = xor i64 %notmask, -1
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_b4_commutative(i64 %val, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_b4_commutative:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %edx
; X86-NOBMI-NEXT:    movl $-1, %esi
; X86-NOBMI-NEXT:    shll %cl, %esi
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB24_1
; X86-NOBMI-NEXT:  # %bb.2:
; X86-NOBMI-NEXT:    movl %esi, %eax
; X86-NOBMI-NEXT:    jmp .LBB24_3
; X86-NOBMI-NEXT:  .LBB24_1:
; X86-NOBMI-NEXT:    movl %esi, %edx
; X86-NOBMI-NEXT:  .LBB24_3:
; X86-NOBMI-NEXT:    notl %edx
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_b4_commutative:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; 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 .LBB24_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:    xorl %eax, %eax
; X86-BMI1-NEXT:  .LBB24_2:
; X86-BMI1-NEXT:    andnl {{[0-9]+}}(%esp), %eax, %eax
; X86-BMI1-NEXT:    andnl {{[0-9]+}}(%esp), %edx, %edx
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_b4_commutative:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %dl
; X86-BMI2-NEXT:    movl $-1, %ecx
; X86-BMI2-NEXT:    shlxl %edx, %ecx, %eax
; X86-BMI2-NEXT:    testb $32, %dl
; X86-BMI2-NEXT:    je .LBB24_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %eax, %ecx
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:  .LBB24_2:
; X86-BMI2-NEXT:    andnl {{[0-9]+}}(%esp), %eax, %eax
; X86-BMI2-NEXT:    andnl {{[0-9]+}}(%esp), %ecx, %edx
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_b4_commutative:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movq $-1, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    notq %rax
; X64-NOBMI-NEXT:    andq %rdi, %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_b4_commutative:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, %rdi, %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_b4_commutative:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhiq %rsi, %rdi, %rax
; X64-BMI2-NEXT:    retq
  %notmask = shl i64 -1, %numlowbits
  %mask = xor i64 %notmask, -1
  %masked = and i64 %val, %mask ; swapped order
  ret i64 %masked
}

; 64-bit, but with 32-bit output

; Everything done in 64-bit, truncation happens last.
define i32 @bzhi64_32_b0(i64 %val, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_b0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %edx
; X86-NOBMI-NEXT:    shll %cl, %edx
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB25_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %edx, %eax
; X86-NOBMI-NEXT:  .LBB25_2:
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_b0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %eax
; X86-BMI1-NEXT:    shll %cl, %eax
; X86-BMI1-NEXT:    xorl %edx, %edx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    jne .LBB25_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:  .LBB25_2:
; X86-BMI1-NEXT:    andnl {{[0-9]+}}(%esp), %edx, %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_b0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    xorl %ecx, %ecx
; X86-BMI2-NEXT:    testb $32, %al
; X86-BMI2-NEXT:    jne .LBB25_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl $-1, %ecx
; X86-BMI2-NEXT:    shlxl %eax, %ecx, %ecx
; X86-BMI2-NEXT:  .LBB25_2:
; X86-BMI2-NEXT:    andnl {{[0-9]+}}(%esp), %ecx, %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_b0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movq $-1, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    notl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_b0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_b0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %widenumlowbits = zext i8 %numlowbits to i64
  %notmask = shl nsw i64 -1, %widenumlowbits
  %mask = xor i64 %notmask, -1
  %wideres = and i64 %val, %mask
  %res = trunc i64 %wideres to i32
  ret i32 %res
}

; Shifting happens in 64-bit, then truncation. Masking is 32-bit.
define i32 @bzhi64_32_b1(i64 %val, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_b1:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_b1:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_b1:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_b1:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $-1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    notl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_b1:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_b1:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %truncval = trunc i64 %val to i32
  %widenumlowbits = zext i8 %numlowbits to i32
  %notmask = shl nsw i32 -1, %widenumlowbits
  %mask = xor i32 %notmask, -1
  %res = and i32 %truncval, %mask
  ret i32 %res
}

; Shifting happens in 64-bit. Mask is 32-bit, but extended to 64-bit.
; Masking is 64-bit. Then truncation.
define i32 @bzhi64_32_b2(i64 %val, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_b2:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %eax
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_b2:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_b2:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_b2:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $-1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    notl %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_b2:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_b2:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %widenumlowbits = zext i8 %numlowbits to i32
  %notmask = shl nsw i32 -1, %widenumlowbits
  %mask = xor i32 %notmask, -1
  %zextmask = zext i32 %mask to i64
  %wideres = and i64 %val, %zextmask
  %res = trunc i64 %wideres to i32
  ret i32 %res
}

; Shifting happens in 64-bit. Mask is 32-bit, but calculated in 64-bit.
; Masking is 64-bit. Then truncation.
define i32 @bzhi64_32_b3(i64 %val, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_b3:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %edx
; X86-NOBMI-NEXT:    shll %cl, %edx
; X86-NOBMI-NEXT:    xorl %eax, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB28_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %edx, %eax
; X86-NOBMI-NEXT:  .LBB28_2:
; X86-NOBMI-NEXT:    notl %eax
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_b3:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %eax
; X86-BMI1-NEXT:    shll %cl, %eax
; X86-BMI1-NEXT:    xorl %edx, %edx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    jne .LBB28_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:  .LBB28_2:
; X86-BMI1-NEXT:    andnl {{[0-9]+}}(%esp), %edx, %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_b3:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    xorl %ecx, %ecx
; X86-BMI2-NEXT:    testb $32, %al
; X86-BMI2-NEXT:    jne .LBB28_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl $-1, %ecx
; X86-BMI2-NEXT:    shlxl %eax, %ecx, %ecx
; X86-BMI2-NEXT:  .LBB28_2:
; X86-BMI2-NEXT:    andnl {{[0-9]+}}(%esp), %ecx, %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_b3:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl $4294967295, %eax # imm = 0xFFFFFFFF
; X64-NOBMI-NEXT:    movl $4294967295, %edx # imm = 0xFFFFFFFF
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shlq %cl, %rdx
; X64-NOBMI-NEXT:    xorl %edx, %eax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_b3:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_b3:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %widenumlowbits = zext i8 %numlowbits to i64
  %notmask = shl nsw i64 4294967295, %widenumlowbits
  %mask = xor i64 %notmask, 4294967295
  %wideres = and i64 %val, %mask
  %res = trunc i64 %wideres to i32
  ret i32 %res
}

; ---------------------------------------------------------------------------- ;
; Pattern c. 32-bit
; ---------------------------------------------------------------------------- ;

declare void @use32(i32)

define i32 @bzhi32_c0(i32 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_c0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    subl $8, %esp
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %esi
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %esi
; X86-NOBMI-NEXT:    movl %esi, (%esp)
; X86-NOBMI-NEXT:    calll use32@PLT
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    movl %esi, %eax
; X86-NOBMI-NEXT:    addl $8, %esp
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_c0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    subl $8, %esp
; X86-BMI1-NEXT:    xorl %ecx, %ecx
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %esi
; X86-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-BMI1-NEXT:    shrl %cl, %esi
; X86-BMI1-NEXT:    movl %esi, (%esp)
; X86-BMI1-NEXT:    calll use32@PLT
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-BMI1-NEXT:    movl %esi, %eax
; X86-BMI1-NEXT:    addl $8, %esp
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_c0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %ebx
; X86-BMI2-NEXT:    subl $8, %esp
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %bl
; X86-BMI2-NEXT:    movl %ebx, %eax
; X86-BMI2-NEXT:    negb %al
; X86-BMI2-NEXT:    movl $-1, %ecx
; X86-BMI2-NEXT:    shrxl %eax, %ecx, %eax
; X86-BMI2-NEXT:    movl %eax, (%esp)
; X86-BMI2-NEXT:    calll use32@PLT
; X86-BMI2-NEXT:    bzhil %ebx, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    addl $8, %esp
; X86-BMI2-NEXT:    popl %ebx
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_c0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    pushq %rbp
; X64-NOBMI-NEXT:    pushq %rbx
; X64-NOBMI-NEXT:    pushq %rax
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl %edi, %ebx
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movl $-1, %ebp
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %ebp
; X64-NOBMI-NEXT:    movl %ebp, %edi
; X64-NOBMI-NEXT:    callq use32@PLT
; X64-NOBMI-NEXT:    andl %ebx, %ebp
; X64-NOBMI-NEXT:    movl %ebp, %eax
; X64-NOBMI-NEXT:    addq $8, %rsp
; X64-NOBMI-NEXT:    popq %rbx
; X64-NOBMI-NEXT:    popq %rbp
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_c0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    pushq %rbp
; X64-BMI1-NEXT:    pushq %rbx
; X64-BMI1-NEXT:    pushq %rax
; X64-BMI1-NEXT:    movl %esi, %ecx
; X64-BMI1-NEXT:    movl %edi, %ebx
; X64-BMI1-NEXT:    negb %cl
; X64-BMI1-NEXT:    movl $-1, %ebp
; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-BMI1-NEXT:    shrl %cl, %ebp
; X64-BMI1-NEXT:    movl %ebp, %edi
; X64-BMI1-NEXT:    callq use32@PLT
; X64-BMI1-NEXT:    andl %ebx, %ebp
; X64-BMI1-NEXT:    movl %ebp, %eax
; X64-BMI1-NEXT:    addq $8, %rsp
; X64-BMI1-NEXT:    popq %rbx
; X64-BMI1-NEXT:    popq %rbp
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_c0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    pushq %rbp
; X64-BMI2-NEXT:    pushq %rbx
; X64-BMI2-NEXT:    pushq %rax
; X64-BMI2-NEXT:    movl %esi, %ebx
; X64-BMI2-NEXT:    movl %edi, %ebp
; X64-BMI2-NEXT:    movl %ebx, %eax
; X64-BMI2-NEXT:    negb %al
; X64-BMI2-NEXT:    movl $-1, %ecx
; X64-BMI2-NEXT:    shrxl %eax, %ecx, %edi
; X64-BMI2-NEXT:    callq use32@PLT
; X64-BMI2-NEXT:    bzhil %ebx, %ebp, %eax
; X64-BMI2-NEXT:    addq $8, %rsp
; X64-BMI2-NEXT:    popq %rbx
; X64-BMI2-NEXT:    popq %rbp
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i32 32, %numlowbits
  %mask = lshr i32 -1, %numhighbits
  call void @use32(i32 %mask)
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_c1_indexzext(i32 %val, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_c1_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    subl $8, %esp
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %esi
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %esi
; X86-NOBMI-NEXT:    movl %esi, (%esp)
; X86-NOBMI-NEXT:    calll use32@PLT
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    movl %esi, %eax
; X86-NOBMI-NEXT:    addl $8, %esp
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_c1_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    subl $8, %esp
; X86-BMI1-NEXT:    xorl %ecx, %ecx
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %esi
; X86-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-BMI1-NEXT:    shrl %cl, %esi
; X86-BMI1-NEXT:    movl %esi, (%esp)
; X86-BMI1-NEXT:    calll use32@PLT
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-BMI1-NEXT:    movl %esi, %eax
; X86-BMI1-NEXT:    addl $8, %esp
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_c1_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %ebx
; X86-BMI2-NEXT:    subl $8, %esp
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %bl
; X86-BMI2-NEXT:    movl %ebx, %eax
; X86-BMI2-NEXT:    negb %al
; X86-BMI2-NEXT:    movl $-1, %ecx
; X86-BMI2-NEXT:    shrxl %eax, %ecx, %eax
; X86-BMI2-NEXT:    movl %eax, (%esp)
; X86-BMI2-NEXT:    calll use32@PLT
; X86-BMI2-NEXT:    bzhil %ebx, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    addl $8, %esp
; X86-BMI2-NEXT:    popl %ebx
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_c1_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    pushq %rbp
; X64-NOBMI-NEXT:    pushq %rbx
; X64-NOBMI-NEXT:    pushq %rax
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl %edi, %ebx
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movl $-1, %ebp
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %ebp
; X64-NOBMI-NEXT:    movl %ebp, %edi
; X64-NOBMI-NEXT:    callq use32@PLT
; X64-NOBMI-NEXT:    andl %ebx, %ebp
; X64-NOBMI-NEXT:    movl %ebp, %eax
; X64-NOBMI-NEXT:    addq $8, %rsp
; X64-NOBMI-NEXT:    popq %rbx
; X64-NOBMI-NEXT:    popq %rbp
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_c1_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    pushq %rbp
; X64-BMI1-NEXT:    pushq %rbx
; X64-BMI1-NEXT:    pushq %rax
; X64-BMI1-NEXT:    movl %esi, %ecx
; X64-BMI1-NEXT:    movl %edi, %ebx
; X64-BMI1-NEXT:    negb %cl
; X64-BMI1-NEXT:    movl $-1, %ebp
; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-BMI1-NEXT:    shrl %cl, %ebp
; X64-BMI1-NEXT:    movl %ebp, %edi
; X64-BMI1-NEXT:    callq use32@PLT
; X64-BMI1-NEXT:    andl %ebx, %ebp
; X64-BMI1-NEXT:    movl %ebp, %eax
; X64-BMI1-NEXT:    addq $8, %rsp
; X64-BMI1-NEXT:    popq %rbx
; X64-BMI1-NEXT:    popq %rbp
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_c1_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    pushq %rbp
; X64-BMI2-NEXT:    pushq %rbx
; X64-BMI2-NEXT:    pushq %rax
; X64-BMI2-NEXT:    movl %esi, %ebx
; X64-BMI2-NEXT:    movl %edi, %ebp
; X64-BMI2-NEXT:    movl %ebx, %eax
; X64-BMI2-NEXT:    negb %al
; X64-BMI2-NEXT:    movl $-1, %ecx
; X64-BMI2-NEXT:    shrxl %eax, %ecx, %edi
; X64-BMI2-NEXT:    callq use32@PLT
; X64-BMI2-NEXT:    bzhil %ebx, %ebp, %eax
; X64-BMI2-NEXT:    addq $8, %rsp
; X64-BMI2-NEXT:    popq %rbx
; X64-BMI2-NEXT:    popq %rbp
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i8 32, %numlowbits
  %sh_prom = zext i8 %numhighbits to i32
  %mask = lshr i32 -1, %sh_prom
  call void @use32(i32 %mask)
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_c2_load(i32* %w, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_c2_load:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    subl $8, %esp
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %edx
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %edx
; X86-NOBMI-NEXT:    movl (%eax), %esi
; X86-NOBMI-NEXT:    andl %edx, %esi
; X86-NOBMI-NEXT:    movl %edx, (%esp)
; X86-NOBMI-NEXT:    calll use32@PLT
; X86-NOBMI-NEXT:    movl %esi, %eax
; X86-NOBMI-NEXT:    addl $8, %esp
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_c2_load:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    subl $8, %esp
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    xorl %ecx, %ecx
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %edx
; X86-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-BMI1-NEXT:    shrl %cl, %edx
; X86-BMI1-NEXT:    movl (%eax), %esi
; X86-BMI1-NEXT:    andl %edx, %esi
; X86-BMI1-NEXT:    movl %edx, (%esp)
; X86-BMI1-NEXT:    calll use32@PLT
; X86-BMI1-NEXT:    movl %esi, %eax
; X86-BMI1-NEXT:    addl $8, %esp
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_c2_load:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    subl $8, %esp
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    bzhil %ecx, (%eax), %esi
; X86-BMI2-NEXT:    # kill: def $cl killed $cl killed $ecx def $ecx
; X86-BMI2-NEXT:    negb %cl
; X86-BMI2-NEXT:    movl $-1, %eax
; X86-BMI2-NEXT:    shrxl %ecx, %eax, %eax
; X86-BMI2-NEXT:    movl %eax, (%esp)
; X86-BMI2-NEXT:    calll use32@PLT
; X86-BMI2-NEXT:    movl %esi, %eax
; X86-BMI2-NEXT:    addl $8, %esp
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_c2_load:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    pushq %rbx
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movl $-1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %eax
; X64-NOBMI-NEXT:    movl (%rdi), %ebx
; X64-NOBMI-NEXT:    andl %eax, %ebx
; X64-NOBMI-NEXT:    movl %eax, %edi
; X64-NOBMI-NEXT:    callq use32@PLT
; X64-NOBMI-NEXT:    movl %ebx, %eax
; X64-NOBMI-NEXT:    popq %rbx
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_c2_load:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    pushq %rbx
; X64-BMI1-NEXT:    movl %esi, %ecx
; X64-BMI1-NEXT:    negb %cl
; X64-BMI1-NEXT:    movl $-1, %eax
; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-BMI1-NEXT:    shrl %cl, %eax
; X64-BMI1-NEXT:    movl (%rdi), %ebx
; X64-BMI1-NEXT:    andl %eax, %ebx
; X64-BMI1-NEXT:    movl %eax, %edi
; X64-BMI1-NEXT:    callq use32@PLT
; X64-BMI1-NEXT:    movl %ebx, %eax
; X64-BMI1-NEXT:    popq %rbx
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_c2_load:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    pushq %rbx
; X64-BMI2-NEXT:    bzhil %esi, (%rdi), %ebx
; X64-BMI2-NEXT:    # kill: def $sil killed $sil killed $esi def $esi
; X64-BMI2-NEXT:    negb %sil
; X64-BMI2-NEXT:    movl $-1, %eax
; X64-BMI2-NEXT:    shrxl %esi, %eax, %edi
; X64-BMI2-NEXT:    callq use32@PLT
; X64-BMI2-NEXT:    movl %ebx, %eax
; X64-BMI2-NEXT:    popq %rbx
; X64-BMI2-NEXT:    retq
  %val = load i32, i32* %w
  %numhighbits = sub i32 32, %numlowbits
  %mask = lshr i32 -1, %numhighbits
  call void @use32(i32 %mask)
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_c3_load_indexzext(i32* %w, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_c3_load_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    subl $8, %esp
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %edx
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %edx
; X86-NOBMI-NEXT:    movl (%eax), %esi
; X86-NOBMI-NEXT:    andl %edx, %esi
; X86-NOBMI-NEXT:    movl %edx, (%esp)
; X86-NOBMI-NEXT:    calll use32@PLT
; X86-NOBMI-NEXT:    movl %esi, %eax
; X86-NOBMI-NEXT:    addl $8, %esp
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_c3_load_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    subl $8, %esp
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    xorl %ecx, %ecx
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %edx
; X86-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-BMI1-NEXT:    shrl %cl, %edx
; X86-BMI1-NEXT:    movl (%eax), %esi
; X86-BMI1-NEXT:    andl %edx, %esi
; X86-BMI1-NEXT:    movl %edx, (%esp)
; X86-BMI1-NEXT:    calll use32@PLT
; X86-BMI1-NEXT:    movl %esi, %eax
; X86-BMI1-NEXT:    addl $8, %esp
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_c3_load_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    subl $8, %esp
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    bzhil %ecx, (%eax), %esi
; X86-BMI2-NEXT:    # kill: def $cl killed $cl killed $ecx def $ecx
; X86-BMI2-NEXT:    negb %cl
; X86-BMI2-NEXT:    movl $-1, %eax
; X86-BMI2-NEXT:    shrxl %ecx, %eax, %eax
; X86-BMI2-NEXT:    movl %eax, (%esp)
; X86-BMI2-NEXT:    calll use32@PLT
; X86-BMI2-NEXT:    movl %esi, %eax
; X86-BMI2-NEXT:    addl $8, %esp
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_c3_load_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    pushq %rbx
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movl $-1, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %eax
; X64-NOBMI-NEXT:    movl (%rdi), %ebx
; X64-NOBMI-NEXT:    andl %eax, %ebx
; X64-NOBMI-NEXT:    movl %eax, %edi
; X64-NOBMI-NEXT:    callq use32@PLT
; X64-NOBMI-NEXT:    movl %ebx, %eax
; X64-NOBMI-NEXT:    popq %rbx
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_c3_load_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    pushq %rbx
; X64-BMI1-NEXT:    movl %esi, %ecx
; X64-BMI1-NEXT:    negb %cl
; X64-BMI1-NEXT:    movl $-1, %eax
; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-BMI1-NEXT:    shrl %cl, %eax
; X64-BMI1-NEXT:    movl (%rdi), %ebx
; X64-BMI1-NEXT:    andl %eax, %ebx
; X64-BMI1-NEXT:    movl %eax, %edi
; X64-BMI1-NEXT:    callq use32@PLT
; X64-BMI1-NEXT:    movl %ebx, %eax
; X64-BMI1-NEXT:    popq %rbx
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_c3_load_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    pushq %rbx
; X64-BMI2-NEXT:    bzhil %esi, (%rdi), %ebx
; X64-BMI2-NEXT:    # kill: def $sil killed $sil killed $esi def $esi
; X64-BMI2-NEXT:    negb %sil
; X64-BMI2-NEXT:    movl $-1, %eax
; X64-BMI2-NEXT:    shrxl %esi, %eax, %edi
; X64-BMI2-NEXT:    callq use32@PLT
; X64-BMI2-NEXT:    movl %ebx, %eax
; X64-BMI2-NEXT:    popq %rbx
; X64-BMI2-NEXT:    retq
  %val = load i32, i32* %w
  %numhighbits = sub i8 32, %numlowbits
  %sh_prom = zext i8 %numhighbits to i32
  %mask = lshr i32 -1, %sh_prom
  call void @use32(i32 %mask)
  %masked = and i32 %mask, %val
  ret i32 %masked
}

define i32 @bzhi32_c4_commutative(i32 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_c4_commutative:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    subl $8, %esp
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %esi
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %esi
; X86-NOBMI-NEXT:    movl %esi, (%esp)
; X86-NOBMI-NEXT:    calll use32@PLT
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    movl %esi, %eax
; X86-NOBMI-NEXT:    addl $8, %esp
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_c4_commutative:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    subl $8, %esp
; X86-BMI1-NEXT:    xorl %ecx, %ecx
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %esi
; X86-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-BMI1-NEXT:    shrl %cl, %esi
; X86-BMI1-NEXT:    movl %esi, (%esp)
; X86-BMI1-NEXT:    calll use32@PLT
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-BMI1-NEXT:    movl %esi, %eax
; X86-BMI1-NEXT:    addl $8, %esp
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_c4_commutative:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %ebx
; X86-BMI2-NEXT:    subl $8, %esp
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %bl
; X86-BMI2-NEXT:    movl %ebx, %eax
; X86-BMI2-NEXT:    negb %al
; X86-BMI2-NEXT:    movl $-1, %ecx
; X86-BMI2-NEXT:    shrxl %eax, %ecx, %eax
; X86-BMI2-NEXT:    movl %eax, (%esp)
; X86-BMI2-NEXT:    calll use32@PLT
; X86-BMI2-NEXT:    bzhil %ebx, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    addl $8, %esp
; X86-BMI2-NEXT:    popl %ebx
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_c4_commutative:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    pushq %rbp
; X64-NOBMI-NEXT:    pushq %rbx
; X64-NOBMI-NEXT:    pushq %rax
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl %edi, %ebx
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movl $-1, %ebp
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %ebp
; X64-NOBMI-NEXT:    movl %ebp, %edi
; X64-NOBMI-NEXT:    callq use32@PLT
; X64-NOBMI-NEXT:    andl %ebx, %ebp
; X64-NOBMI-NEXT:    movl %ebp, %eax
; X64-NOBMI-NEXT:    addq $8, %rsp
; X64-NOBMI-NEXT:    popq %rbx
; X64-NOBMI-NEXT:    popq %rbp
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_c4_commutative:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    pushq %rbp
; X64-BMI1-NEXT:    pushq %rbx
; X64-BMI1-NEXT:    pushq %rax
; X64-BMI1-NEXT:    movl %esi, %ecx
; X64-BMI1-NEXT:    movl %edi, %ebx
; X64-BMI1-NEXT:    negb %cl
; X64-BMI1-NEXT:    movl $-1, %ebp
; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-BMI1-NEXT:    shrl %cl, %ebp
; X64-BMI1-NEXT:    movl %ebp, %edi
; X64-BMI1-NEXT:    callq use32@PLT
; X64-BMI1-NEXT:    andl %ebx, %ebp
; X64-BMI1-NEXT:    movl %ebp, %eax
; X64-BMI1-NEXT:    addq $8, %rsp
; X64-BMI1-NEXT:    popq %rbx
; X64-BMI1-NEXT:    popq %rbp
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_c4_commutative:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    pushq %rbp
; X64-BMI2-NEXT:    pushq %rbx
; X64-BMI2-NEXT:    pushq %rax
; X64-BMI2-NEXT:    movl %esi, %ebx
; X64-BMI2-NEXT:    movl %edi, %ebp
; X64-BMI2-NEXT:    movl %ebx, %eax
; X64-BMI2-NEXT:    negb %al
; X64-BMI2-NEXT:    movl $-1, %ecx
; X64-BMI2-NEXT:    shrxl %eax, %ecx, %edi
; X64-BMI2-NEXT:    callq use32@PLT
; X64-BMI2-NEXT:    bzhil %ebx, %ebp, %eax
; X64-BMI2-NEXT:    addq $8, %rsp
; X64-BMI2-NEXT:    popq %rbx
; X64-BMI2-NEXT:    popq %rbp
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i32 32, %numlowbits
  %mask = lshr i32 -1, %numhighbits
  call void @use32(i32 %mask)
  %masked = and i32 %val, %mask ; swapped order
  ret i32 %masked
}

; 64-bit

declare void @use64(i64)

define i64 @bzhi64_c0(i64 %val, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_c0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    pushl %eax
; X86-NOBMI-NEXT:    movb $64, %cl
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %esi
; X86-NOBMI-NEXT:    movl $-1, %edi
; X86-NOBMI-NEXT:    shrl %cl, %edi
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    je .LBB34_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %edi, %esi
; X86-NOBMI-NEXT:    xorl %edi, %edi
; X86-NOBMI-NEXT:  .LBB34_2:
; X86-NOBMI-NEXT:    subl $8, %esp
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    calll use64@PLT
; X86-NOBMI-NEXT:    addl $16, %esp
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edi
; X86-NOBMI-NEXT:    movl %esi, %eax
; X86-NOBMI-NEXT:    movl %edi, %edx
; X86-NOBMI-NEXT:    addl $4, %esp
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %edi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_c0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    pushl %eax
; X86-BMI1-NEXT:    movb $64, %cl
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %esi
; X86-BMI1-NEXT:    movl $-1, %edi
; X86-BMI1-NEXT:    shrl %cl, %edi
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB34_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %edi, %esi
; X86-BMI1-NEXT:    xorl %edi, %edi
; X86-BMI1-NEXT:  .LBB34_2:
; X86-BMI1-NEXT:    subl $8, %esp
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    calll use64@PLT
; X86-BMI1-NEXT:    addl $16, %esp
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %edi
; X86-BMI1-NEXT:    movl %esi, %eax
; X86-BMI1-NEXT:    movl %edi, %edx
; X86-BMI1-NEXT:    addl $4, %esp
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    popl %edi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_c0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    pushl %eax
; X86-BMI2-NEXT:    movb $64, %al
; X86-BMI2-NEXT:    subb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    movl $-1, %edi
; X86-BMI2-NEXT:    shrxl %eax, %edi, %esi
; X86-BMI2-NEXT:    testb $32, %al
; X86-BMI2-NEXT:    je .LBB34_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %esi, %edi
; X86-BMI2-NEXT:    xorl %esi, %esi
; X86-BMI2-NEXT:  .LBB34_2:
; X86-BMI2-NEXT:    subl $8, %esp
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    calll use64@PLT
; X86-BMI2-NEXT:    addl $16, %esp
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %edi
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-BMI2-NEXT:    movl %edi, %eax
; X86-BMI2-NEXT:    movl %esi, %edx
; X86-BMI2-NEXT:    addl $4, %esp
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    popl %edi
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_c0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    pushq %r14
; X64-NOBMI-NEXT:    pushq %rbx
; X64-NOBMI-NEXT:    pushq %rax
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movq %rdi, %r14
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movq $-1, %rbx
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shrq %cl, %rbx
; X64-NOBMI-NEXT:    movq %rbx, %rdi
; X64-NOBMI-NEXT:    callq use64@PLT
; X64-NOBMI-NEXT:    andq %r14, %rbx
; X64-NOBMI-NEXT:    movq %rbx, %rax
; X64-NOBMI-NEXT:    addq $8, %rsp
; X64-NOBMI-NEXT:    popq %rbx
; X64-NOBMI-NEXT:    popq %r14
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_c0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    pushq %r14
; X64-BMI1-NEXT:    pushq %rbx
; X64-BMI1-NEXT:    pushq %rax
; X64-BMI1-NEXT:    movq %rsi, %rcx
; X64-BMI1-NEXT:    movq %rdi, %r14
; X64-BMI1-NEXT:    negb %cl
; X64-BMI1-NEXT:    movq $-1, %rbx
; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-BMI1-NEXT:    shrq %cl, %rbx
; X64-BMI1-NEXT:    movq %rbx, %rdi
; X64-BMI1-NEXT:    callq use64@PLT
; X64-BMI1-NEXT:    andq %r14, %rbx
; X64-BMI1-NEXT:    movq %rbx, %rax
; X64-BMI1-NEXT:    addq $8, %rsp
; X64-BMI1-NEXT:    popq %rbx
; X64-BMI1-NEXT:    popq %r14
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_c0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    pushq %r14
; X64-BMI2-NEXT:    pushq %rbx
; X64-BMI2-NEXT:    pushq %rax
; X64-BMI2-NEXT:    movq %rsi, %rbx
; X64-BMI2-NEXT:    movq %rdi, %r14
; X64-BMI2-NEXT:    movl %ebx, %eax
; X64-BMI2-NEXT:    negb %al
; X64-BMI2-NEXT:    movq $-1, %rcx
; X64-BMI2-NEXT:    shrxq %rax, %rcx, %rdi
; X64-BMI2-NEXT:    callq use64@PLT
; X64-BMI2-NEXT:    bzhiq %rbx, %r14, %rax
; X64-BMI2-NEXT:    addq $8, %rsp
; X64-BMI2-NEXT:    popq %rbx
; X64-BMI2-NEXT:    popq %r14
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i64 64, %numlowbits
  %mask = lshr i64 -1, %numhighbits
  call void @use64(i64 %mask)
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_c1_indexzext(i64 %val, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_c1_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    pushl %eax
; X86-NOBMI-NEXT:    movb $64, %cl
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %esi
; X86-NOBMI-NEXT:    movl $-1, %edi
; X86-NOBMI-NEXT:    shrl %cl, %edi
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    je .LBB35_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %edi, %esi
; X86-NOBMI-NEXT:    xorl %edi, %edi
; X86-NOBMI-NEXT:  .LBB35_2:
; X86-NOBMI-NEXT:    subl $8, %esp
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    calll use64@PLT
; X86-NOBMI-NEXT:    addl $16, %esp
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edi
; X86-NOBMI-NEXT:    movl %esi, %eax
; X86-NOBMI-NEXT:    movl %edi, %edx
; X86-NOBMI-NEXT:    addl $4, %esp
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %edi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_c1_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    pushl %eax
; X86-BMI1-NEXT:    movb $64, %cl
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %esi
; X86-BMI1-NEXT:    movl $-1, %edi
; X86-BMI1-NEXT:    shrl %cl, %edi
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB35_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %edi, %esi
; X86-BMI1-NEXT:    xorl %edi, %edi
; X86-BMI1-NEXT:  .LBB35_2:
; X86-BMI1-NEXT:    subl $8, %esp
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    calll use64@PLT
; X86-BMI1-NEXT:    addl $16, %esp
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %edi
; X86-BMI1-NEXT:    movl %esi, %eax
; X86-BMI1-NEXT:    movl %edi, %edx
; X86-BMI1-NEXT:    addl $4, %esp
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    popl %edi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_c1_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    pushl %eax
; X86-BMI2-NEXT:    movb $64, %al
; X86-BMI2-NEXT:    subb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    movl $-1, %edi
; X86-BMI2-NEXT:    shrxl %eax, %edi, %esi
; X86-BMI2-NEXT:    testb $32, %al
; X86-BMI2-NEXT:    je .LBB35_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %esi, %edi
; X86-BMI2-NEXT:    xorl %esi, %esi
; X86-BMI2-NEXT:  .LBB35_2:
; X86-BMI2-NEXT:    subl $8, %esp
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    calll use64@PLT
; X86-BMI2-NEXT:    addl $16, %esp
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %edi
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-BMI2-NEXT:    movl %edi, %eax
; X86-BMI2-NEXT:    movl %esi, %edx
; X86-BMI2-NEXT:    addl $4, %esp
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    popl %edi
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_c1_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    pushq %r14
; X64-NOBMI-NEXT:    pushq %rbx
; X64-NOBMI-NEXT:    pushq %rax
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movq %rdi, %r14
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movq $-1, %rbx
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrq %cl, %rbx
; X64-NOBMI-NEXT:    movq %rbx, %rdi
; X64-NOBMI-NEXT:    callq use64@PLT
; X64-NOBMI-NEXT:    andq %r14, %rbx
; X64-NOBMI-NEXT:    movq %rbx, %rax
; X64-NOBMI-NEXT:    addq $8, %rsp
; X64-NOBMI-NEXT:    popq %rbx
; X64-NOBMI-NEXT:    popq %r14
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_c1_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    pushq %r14
; X64-BMI1-NEXT:    pushq %rbx
; X64-BMI1-NEXT:    pushq %rax
; X64-BMI1-NEXT:    movl %esi, %ecx
; X64-BMI1-NEXT:    movq %rdi, %r14
; X64-BMI1-NEXT:    negb %cl
; X64-BMI1-NEXT:    movq $-1, %rbx
; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-BMI1-NEXT:    shrq %cl, %rbx
; X64-BMI1-NEXT:    movq %rbx, %rdi
; X64-BMI1-NEXT:    callq use64@PLT
; X64-BMI1-NEXT:    andq %r14, %rbx
; X64-BMI1-NEXT:    movq %rbx, %rax
; X64-BMI1-NEXT:    addq $8, %rsp
; X64-BMI1-NEXT:    popq %rbx
; X64-BMI1-NEXT:    popq %r14
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_c1_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    pushq %r14
; X64-BMI2-NEXT:    pushq %rbx
; X64-BMI2-NEXT:    pushq %rax
; X64-BMI2-NEXT:    movl %esi, %ebx
; X64-BMI2-NEXT:    movq %rdi, %r14
; X64-BMI2-NEXT:    movl %ebx, %eax
; X64-BMI2-NEXT:    negb %al
; X64-BMI2-NEXT:    movq $-1, %rcx
; X64-BMI2-NEXT:    shrxq %rax, %rcx, %rdi
; X64-BMI2-NEXT:    callq use64@PLT
; X64-BMI2-NEXT:    bzhiq %rbx, %r14, %rax
; X64-BMI2-NEXT:    addq $8, %rsp
; X64-BMI2-NEXT:    popq %rbx
; X64-BMI2-NEXT:    popq %r14
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i8 64, %numlowbits
  %sh_prom = zext i8 %numhighbits to i64
  %mask = lshr i64 -1, %sh_prom
  call void @use64(i64 %mask)
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_c2_load(i64* %w, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_c2_load:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %ebx
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    movb $64, %cl
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %eax
; X86-NOBMI-NEXT:    movl $-1, %ebx
; X86-NOBMI-NEXT:    shrl %cl, %ebx
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    je .LBB36_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %ebx, %eax
; X86-NOBMI-NEXT:    xorl %ebx, %ebx
; X86-NOBMI-NEXT:  .LBB36_2:
; X86-NOBMI-NEXT:    movl 4(%edx), %esi
; X86-NOBMI-NEXT:    andl %ebx, %esi
; X86-NOBMI-NEXT:    movl (%edx), %edi
; X86-NOBMI-NEXT:    andl %eax, %edi
; X86-NOBMI-NEXT:    subl $8, %esp
; X86-NOBMI-NEXT:    pushl %ebx
; X86-NOBMI-NEXT:    pushl %eax
; X86-NOBMI-NEXT:    calll use64@PLT
; X86-NOBMI-NEXT:    addl $16, %esp
; X86-NOBMI-NEXT:    movl %edi, %eax
; X86-NOBMI-NEXT:    movl %esi, %edx
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %edi
; X86-NOBMI-NEXT:    popl %ebx
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_c2_load:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %ebx
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-BMI1-NEXT:    movb $64, %cl
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %eax
; X86-BMI1-NEXT:    movl $-1, %ebx
; X86-BMI1-NEXT:    shrl %cl, %ebx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB36_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %ebx, %eax
; X86-BMI1-NEXT:    xorl %ebx, %ebx
; X86-BMI1-NEXT:  .LBB36_2:
; X86-BMI1-NEXT:    movl 4(%edx), %esi
; X86-BMI1-NEXT:    andl %ebx, %esi
; X86-BMI1-NEXT:    movl (%edx), %edi
; X86-BMI1-NEXT:    andl %eax, %edi
; X86-BMI1-NEXT:    subl $8, %esp
; X86-BMI1-NEXT:    pushl %ebx
; X86-BMI1-NEXT:    pushl %eax
; X86-BMI1-NEXT:    calll use64@PLT
; X86-BMI1-NEXT:    addl $16, %esp
; X86-BMI1-NEXT:    movl %edi, %eax
; X86-BMI1-NEXT:    movl %esi, %edx
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    popl %edi
; X86-BMI1-NEXT:    popl %ebx
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_c2_load:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %ebx
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movb $64, %bl
; X86-BMI2-NEXT:    subb {{[0-9]+}}(%esp), %bl
; X86-BMI2-NEXT:    movl $-1, %ecx
; X86-BMI2-NEXT:    shrxl %ebx, %ecx, %edx
; X86-BMI2-NEXT:    testb $32, %bl
; X86-BMI2-NEXT:    je .LBB36_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %edx, %ecx
; X86-BMI2-NEXT:    xorl %edx, %edx
; X86-BMI2-NEXT:  .LBB36_2:
; X86-BMI2-NEXT:    movl 4(%eax), %esi
; X86-BMI2-NEXT:    andl %edx, %esi
; X86-BMI2-NEXT:    movl (%eax), %edi
; X86-BMI2-NEXT:    andl %ecx, %edi
; X86-BMI2-NEXT:    subl $8, %esp
; X86-BMI2-NEXT:    pushl %edx
; X86-BMI2-NEXT:    pushl %ecx
; X86-BMI2-NEXT:    calll use64@PLT
; X86-BMI2-NEXT:    addl $16, %esp
; X86-BMI2-NEXT:    movl %edi, %eax
; X86-BMI2-NEXT:    movl %esi, %edx
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    popl %edi
; X86-BMI2-NEXT:    popl %ebx
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_c2_load:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    pushq %rbx
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movq $-1, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shrq %cl, %rax
; X64-NOBMI-NEXT:    movq (%rdi), %rbx
; X64-NOBMI-NEXT:    andq %rax, %rbx
; X64-NOBMI-NEXT:    movq %rax, %rdi
; X64-NOBMI-NEXT:    callq use64@PLT
; X64-NOBMI-NEXT:    movq %rbx, %rax
; X64-NOBMI-NEXT:    popq %rbx
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_c2_load:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    pushq %rbx
; X64-BMI1-NEXT:    movq %rsi, %rcx
; X64-BMI1-NEXT:    negb %cl
; X64-BMI1-NEXT:    movq $-1, %rax
; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-BMI1-NEXT:    shrq %cl, %rax
; X64-BMI1-NEXT:    movq (%rdi), %rbx
; X64-BMI1-NEXT:    andq %rax, %rbx
; X64-BMI1-NEXT:    movq %rax, %rdi
; X64-BMI1-NEXT:    callq use64@PLT
; X64-BMI1-NEXT:    movq %rbx, %rax
; X64-BMI1-NEXT:    popq %rbx
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_c2_load:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    pushq %rbx
; X64-BMI2-NEXT:    bzhiq %rsi, (%rdi), %rbx
; X64-BMI2-NEXT:    # kill: def $sil killed $sil killed $rsi def $rsi
; X64-BMI2-NEXT:    negb %sil
; X64-BMI2-NEXT:    movq $-1, %rax
; X64-BMI2-NEXT:    shrxq %rsi, %rax, %rdi
; X64-BMI2-NEXT:    callq use64@PLT
; X64-BMI2-NEXT:    movq %rbx, %rax
; X64-BMI2-NEXT:    popq %rbx
; X64-BMI2-NEXT:    retq
  %val = load i64, i64* %w
  %numhighbits = sub i64 64, %numlowbits
  %mask = lshr i64 -1, %numhighbits
  call void @use64(i64 %mask)
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_c3_load_indexzext(i64* %w, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_c3_load_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %ebx
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    movb $64, %cl
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %eax
; X86-NOBMI-NEXT:    movl $-1, %ebx
; X86-NOBMI-NEXT:    shrl %cl, %ebx
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    je .LBB37_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %ebx, %eax
; X86-NOBMI-NEXT:    xorl %ebx, %ebx
; X86-NOBMI-NEXT:  .LBB37_2:
; X86-NOBMI-NEXT:    movl 4(%edx), %esi
; X86-NOBMI-NEXT:    andl %ebx, %esi
; X86-NOBMI-NEXT:    movl (%edx), %edi
; X86-NOBMI-NEXT:    andl %eax, %edi
; X86-NOBMI-NEXT:    subl $8, %esp
; X86-NOBMI-NEXT:    pushl %ebx
; X86-NOBMI-NEXT:    pushl %eax
; X86-NOBMI-NEXT:    calll use64@PLT
; X86-NOBMI-NEXT:    addl $16, %esp
; X86-NOBMI-NEXT:    movl %edi, %eax
; X86-NOBMI-NEXT:    movl %esi, %edx
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %edi
; X86-NOBMI-NEXT:    popl %ebx
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_c3_load_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %ebx
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-BMI1-NEXT:    movb $64, %cl
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %eax
; X86-BMI1-NEXT:    movl $-1, %ebx
; X86-BMI1-NEXT:    shrl %cl, %ebx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB37_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %ebx, %eax
; X86-BMI1-NEXT:    xorl %ebx, %ebx
; X86-BMI1-NEXT:  .LBB37_2:
; X86-BMI1-NEXT:    movl 4(%edx), %esi
; X86-BMI1-NEXT:    andl %ebx, %esi
; X86-BMI1-NEXT:    movl (%edx), %edi
; X86-BMI1-NEXT:    andl %eax, %edi
; X86-BMI1-NEXT:    subl $8, %esp
; X86-BMI1-NEXT:    pushl %ebx
; X86-BMI1-NEXT:    pushl %eax
; X86-BMI1-NEXT:    calll use64@PLT
; X86-BMI1-NEXT:    addl $16, %esp
; X86-BMI1-NEXT:    movl %edi, %eax
; X86-BMI1-NEXT:    movl %esi, %edx
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    popl %edi
; X86-BMI1-NEXT:    popl %ebx
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_c3_load_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %ebx
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movb $64, %bl
; X86-BMI2-NEXT:    subb {{[0-9]+}}(%esp), %bl
; X86-BMI2-NEXT:    movl $-1, %ecx
; X86-BMI2-NEXT:    shrxl %ebx, %ecx, %edx
; X86-BMI2-NEXT:    testb $32, %bl
; X86-BMI2-NEXT:    je .LBB37_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %edx, %ecx
; X86-BMI2-NEXT:    xorl %edx, %edx
; X86-BMI2-NEXT:  .LBB37_2:
; X86-BMI2-NEXT:    movl 4(%eax), %esi
; X86-BMI2-NEXT:    andl %edx, %esi
; X86-BMI2-NEXT:    movl (%eax), %edi
; X86-BMI2-NEXT:    andl %ecx, %edi
; X86-BMI2-NEXT:    subl $8, %esp
; X86-BMI2-NEXT:    pushl %edx
; X86-BMI2-NEXT:    pushl %ecx
; X86-BMI2-NEXT:    calll use64@PLT
; X86-BMI2-NEXT:    addl $16, %esp
; X86-BMI2-NEXT:    movl %edi, %eax
; X86-BMI2-NEXT:    movl %esi, %edx
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    popl %edi
; X86-BMI2-NEXT:    popl %ebx
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_c3_load_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    pushq %rbx
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movq $-1, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrq %cl, %rax
; X64-NOBMI-NEXT:    movq (%rdi), %rbx
; X64-NOBMI-NEXT:    andq %rax, %rbx
; X64-NOBMI-NEXT:    movq %rax, %rdi
; X64-NOBMI-NEXT:    callq use64@PLT
; X64-NOBMI-NEXT:    movq %rbx, %rax
; X64-NOBMI-NEXT:    popq %rbx
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_c3_load_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    pushq %rbx
; X64-BMI1-NEXT:    movl %esi, %ecx
; X64-BMI1-NEXT:    negb %cl
; X64-BMI1-NEXT:    movq $-1, %rax
; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-BMI1-NEXT:    shrq %cl, %rax
; X64-BMI1-NEXT:    movq (%rdi), %rbx
; X64-BMI1-NEXT:    andq %rax, %rbx
; X64-BMI1-NEXT:    movq %rax, %rdi
; X64-BMI1-NEXT:    callq use64@PLT
; X64-BMI1-NEXT:    movq %rbx, %rax
; X64-BMI1-NEXT:    popq %rbx
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_c3_load_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    pushq %rbx
; X64-BMI2-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI2-NEXT:    bzhiq %rsi, (%rdi), %rbx
; X64-BMI2-NEXT:    # kill: def $sil killed $sil killed $rsi def $rsi
; X64-BMI2-NEXT:    negb %sil
; X64-BMI2-NEXT:    movq $-1, %rax
; X64-BMI2-NEXT:    shrxq %rsi, %rax, %rdi
; X64-BMI2-NEXT:    callq use64@PLT
; X64-BMI2-NEXT:    movq %rbx, %rax
; X64-BMI2-NEXT:    popq %rbx
; X64-BMI2-NEXT:    retq
  %val = load i64, i64* %w
  %numhighbits = sub i8 64, %numlowbits
  %sh_prom = zext i8 %numhighbits to i64
  %mask = lshr i64 -1, %sh_prom
  call void @use64(i64 %mask)
  %masked = and i64 %mask, %val
  ret i64 %masked
}

define i64 @bzhi64_c4_commutative(i64 %val, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_c4_commutative:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    pushl %eax
; X86-NOBMI-NEXT:    movb $64, %cl
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %esi
; X86-NOBMI-NEXT:    movl $-1, %edi
; X86-NOBMI-NEXT:    shrl %cl, %edi
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    je .LBB38_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %edi, %esi
; X86-NOBMI-NEXT:    xorl %edi, %edi
; X86-NOBMI-NEXT:  .LBB38_2:
; X86-NOBMI-NEXT:    subl $8, %esp
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    calll use64@PLT
; X86-NOBMI-NEXT:    addl $16, %esp
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %edi
; X86-NOBMI-NEXT:    movl %esi, %eax
; X86-NOBMI-NEXT:    movl %edi, %edx
; X86-NOBMI-NEXT:    addl $4, %esp
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %edi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_c4_commutative:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    pushl %eax
; X86-BMI1-NEXT:    movb $64, %cl
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %esi
; X86-BMI1-NEXT:    movl $-1, %edi
; X86-BMI1-NEXT:    shrl %cl, %edi
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB38_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %edi, %esi
; X86-BMI1-NEXT:    xorl %edi, %edi
; X86-BMI1-NEXT:  .LBB38_2:
; X86-BMI1-NEXT:    subl $8, %esp
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    calll use64@PLT
; X86-BMI1-NEXT:    addl $16, %esp
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %edi
; X86-BMI1-NEXT:    movl %esi, %eax
; X86-BMI1-NEXT:    movl %edi, %edx
; X86-BMI1-NEXT:    addl $4, %esp
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    popl %edi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_c4_commutative:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    pushl %eax
; X86-BMI2-NEXT:    movb $64, %al
; X86-BMI2-NEXT:    subb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    movl $-1, %edi
; X86-BMI2-NEXT:    shrxl %eax, %edi, %esi
; X86-BMI2-NEXT:    testb $32, %al
; X86-BMI2-NEXT:    je .LBB38_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %esi, %edi
; X86-BMI2-NEXT:    xorl %esi, %esi
; X86-BMI2-NEXT:  .LBB38_2:
; X86-BMI2-NEXT:    subl $8, %esp
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    calll use64@PLT
; X86-BMI2-NEXT:    addl $16, %esp
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %edi
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %esi
; X86-BMI2-NEXT:    movl %edi, %eax
; X86-BMI2-NEXT:    movl %esi, %edx
; X86-BMI2-NEXT:    addl $4, %esp
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    popl %edi
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_c4_commutative:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    pushq %r14
; X64-NOBMI-NEXT:    pushq %rbx
; X64-NOBMI-NEXT:    pushq %rax
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movq %rdi, %r14
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movq $-1, %rbx
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shrq %cl, %rbx
; X64-NOBMI-NEXT:    movq %rbx, %rdi
; X64-NOBMI-NEXT:    callq use64@PLT
; X64-NOBMI-NEXT:    andq %r14, %rbx
; X64-NOBMI-NEXT:    movq %rbx, %rax
; X64-NOBMI-NEXT:    addq $8, %rsp
; X64-NOBMI-NEXT:    popq %rbx
; X64-NOBMI-NEXT:    popq %r14
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_c4_commutative:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    pushq %r14
; X64-BMI1-NEXT:    pushq %rbx
; X64-BMI1-NEXT:    pushq %rax
; X64-BMI1-NEXT:    movq %rsi, %rcx
; X64-BMI1-NEXT:    movq %rdi, %r14
; X64-BMI1-NEXT:    negb %cl
; X64-BMI1-NEXT:    movq $-1, %rbx
; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-BMI1-NEXT:    shrq %cl, %rbx
; X64-BMI1-NEXT:    movq %rbx, %rdi
; X64-BMI1-NEXT:    callq use64@PLT
; X64-BMI1-NEXT:    andq %r14, %rbx
; X64-BMI1-NEXT:    movq %rbx, %rax
; X64-BMI1-NEXT:    addq $8, %rsp
; X64-BMI1-NEXT:    popq %rbx
; X64-BMI1-NEXT:    popq %r14
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_c4_commutative:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    pushq %r14
; X64-BMI2-NEXT:    pushq %rbx
; X64-BMI2-NEXT:    pushq %rax
; X64-BMI2-NEXT:    movq %rsi, %rbx
; X64-BMI2-NEXT:    movq %rdi, %r14
; X64-BMI2-NEXT:    movl %ebx, %eax
; X64-BMI2-NEXT:    negb %al
; X64-BMI2-NEXT:    movq $-1, %rcx
; X64-BMI2-NEXT:    shrxq %rax, %rcx, %rdi
; X64-BMI2-NEXT:    callq use64@PLT
; X64-BMI2-NEXT:    bzhiq %rbx, %r14, %rax
; X64-BMI2-NEXT:    addq $8, %rsp
; X64-BMI2-NEXT:    popq %rbx
; X64-BMI2-NEXT:    popq %r14
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i64 64, %numlowbits
  %mask = lshr i64 -1, %numhighbits
  call void @use64(i64 %mask)
  %masked = and i64 %val, %mask ; swapped order
  ret i64 %masked
}

; 64-bit, but with 32-bit output

; Everything done in 64-bit, truncation happens last.
define i32 @bzhi64_32_c0(i64 %val, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_c0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movb $64, %cl
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl $-1, %eax
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB39_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl $-1, %eax
; X86-NOBMI-NEXT:  .LBB39_2:
; X86-NOBMI-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_c0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb $64, %cl
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl $-1, %eax
; X86-BMI1-NEXT:    shrl %cl, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    jne .LBB39_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl $-1, %eax
; X86-BMI1-NEXT:  .LBB39_2:
; X86-BMI1-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_c0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb $64, %cl
; X86-BMI2-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    movl $-1, %eax
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB39_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    shrxl %ecx, %eax, %eax
; X86-BMI2-NEXT:  .LBB39_2:
; X86-BMI2-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_c0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movq $-1, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shrq %cl, %rax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_c0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_c0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i64 64, %numlowbits
  %mask = lshr i64 -1, %numhighbits
  %masked = and i64 %mask, %val
  %res = trunc i64 %masked to i32
  ret i32 %res
}

; Shifting happens in 64-bit, then truncation. Masking is 32-bit.
define i32 @bzhi64_32_c1(i64 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_c1:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_c1:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_c1:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_c1:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movq %rdi, %rax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %eax
; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_c1:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_c1:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %truncval = trunc i64 %val to i32
  %numhighbits = sub i32 32, %numlowbits
  %mask = lshr i32 -1, %numhighbits
  %masked = and i32 %mask, %truncval
  ret i32 %masked
}

; Shifting happens in 64-bit. Mask is 32-bit, but extended to 64-bit.
; Masking is 64-bit. Then truncation.
define i32 @bzhi64_32_c2(i64 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_c2:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_c2:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_c2:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_c2:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movq %rdi, %rax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %eax
; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_c2:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_c2:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i32 32, %numlowbits
  %mask = lshr i32 -1, %numhighbits
  %zextmask = zext i32 %mask to i64
  %masked = and i64 %zextmask, %val
  %truncmasked = trunc i64 %masked to i32
  ret i32 %truncmasked
}

; Shifting happens in 64-bit. Mask is 32-bit, but calculated in 64-bit.
; Masking is 64-bit. Then truncation.
define i32 @bzhi64_32_c3(i64 %val, i64 %numlowbits) nounwind {
; X86-LABEL: bzhi64_32_c3:
; X86:       # %bb.0:
; X86-NEXT:    movb $64, %cl
; X86-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NEXT:    xorl %eax, %eax
; X86-NEXT:    movl $-1, %edx
; X86-NEXT:    shrdl %cl, %eax, %edx
; X86-NEXT:    testb $32, %cl
; X86-NEXT:    jne .LBB42_2
; X86-NEXT:  # %bb.1:
; X86-NEXT:    movl %edx, %eax
; X86-NEXT:  .LBB42_2:
; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_c3:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    movl $4294967295, %eax # imm = 0xFFFFFFFF
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shrq %cl, %rax
; X64-NOBMI-NEXT:    andl %edi, %eax
; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_c3:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    movq %rsi, %rcx
; X64-BMI1-NEXT:    negb %cl
; X64-BMI1-NEXT:    movl $4294967295, %eax # imm = 0xFFFFFFFF
; X64-BMI1-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-BMI1-NEXT:    shrq %cl, %rax
; X64-BMI1-NEXT:    andl %edi, %eax
; X64-BMI1-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_c3:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    negb %sil
; X64-BMI2-NEXT:    movl $4294967295, %eax # imm = 0xFFFFFFFF
; X64-BMI2-NEXT:    shrxq %rsi, %rax, %rax
; X64-BMI2-NEXT:    andl %edi, %eax
; X64-BMI2-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i64 64, %numlowbits
  %mask = lshr i64 4294967295, %numhighbits
  %masked = and i64 %mask, %val
  %truncmasked = trunc i64 %masked to i32
  ret i32 %truncmasked
}

; ---------------------------------------------------------------------------- ;
; Pattern d. 32-bit.
; ---------------------------------------------------------------------------- ;

define i32 @bzhi32_d0(i32 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_d0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_d0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_d0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_d0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl %edi, %eax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_d0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_d0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i32 32, %numlowbits
  %highbitscleared = shl i32 %val, %numhighbits
  %masked = lshr i32 %highbitscleared, %numhighbits
  ret i32 %masked
}

define i32 @bzhi32_d1_indexzext(i32 %val, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_d1_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_d1_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_d1_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_d1_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl %edi, %eax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_d1_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_d1_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i8 32, %numlowbits
  %sh_prom = zext i8 %numhighbits to i32
  %highbitscleared = shl i32 %val, %sh_prom
  %masked = lshr i32 %highbitscleared, %sh_prom
  ret i32 %masked
}

define i32 @bzhi32_d2_load(i32* %w, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_d2_load:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    movl (%eax), %eax
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_d2_load:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    shll $8, %ecx
; X86-BMI1-NEXT:    bextrl %ecx, (%eax), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_d2_load:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    bzhil %ecx, (%eax), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_d2_load:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl (%rdi), %eax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_d2_load:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, (%rdi), %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_d2_load:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, (%rdi), %eax
; X64-BMI2-NEXT:    retq
  %val = load i32, i32* %w
  %numhighbits = sub i32 32, %numlowbits
  %highbitscleared = shl i32 %val, %numhighbits
  %masked = lshr i32 %highbitscleared, %numhighbits
  ret i32 %masked
}

define i32 @bzhi32_d3_load_indexzext(i32* %w, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi32_d3_load_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    movl (%eax), %eax
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi32_d3_load_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    shll $8, %ecx
; X86-BMI1-NEXT:    bextrl %ecx, (%eax), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi32_d3_load_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    bzhil %ecx, (%eax), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi32_d3_load_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movl (%rdi), %eax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %eax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi32_d3_load_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, (%rdi), %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi32_d3_load_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, (%rdi), %eax
; X64-BMI2-NEXT:    retq
  %val = load i32, i32* %w
  %numhighbits = sub i8 32, %numlowbits
  %sh_prom = zext i8 %numhighbits to i32
  %highbitscleared = shl i32 %val, %sh_prom
  %masked = lshr i32 %highbitscleared, %sh_prom
  ret i32 %masked
}

; 64-bit.

define i64 @bzhi64_d0(i64 %val, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_d0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %ebx
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    movb $64, %cl
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl %edx, %esi
; X86-NOBMI-NEXT:    shll %cl, %esi
; X86-NOBMI-NEXT:    shldl %cl, %edx, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    movl %esi, %edi
; X86-NOBMI-NEXT:    jne .LBB47_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %eax, %edi
; X86-NOBMI-NEXT:  .LBB47_2:
; X86-NOBMI-NEXT:    movl %edi, %eax
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    xorl %ebx, %ebx
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    movl $0, %edx
; X86-NOBMI-NEXT:    jne .LBB47_4
; X86-NOBMI-NEXT:  # %bb.3:
; X86-NOBMI-NEXT:    movl %esi, %ebx
; X86-NOBMI-NEXT:    movl %eax, %edx
; X86-NOBMI-NEXT:  .LBB47_4:
; X86-NOBMI-NEXT:    shrdl %cl, %edi, %ebx
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB47_6
; X86-NOBMI-NEXT:  # %bb.5:
; X86-NOBMI-NEXT:    movl %ebx, %eax
; X86-NOBMI-NEXT:  .LBB47_6:
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %edi
; X86-NOBMI-NEXT:    popl %ebx
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_d0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %ebx
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    movb $64, %cl
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl %edx, %esi
; X86-BMI1-NEXT:    shll %cl, %esi
; X86-BMI1-NEXT:    shldl %cl, %edx, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    movl %esi, %edi
; X86-BMI1-NEXT:    jne .LBB47_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edi
; X86-BMI1-NEXT:  .LBB47_2:
; X86-BMI1-NEXT:    movl %edi, %eax
; X86-BMI1-NEXT:    shrl %cl, %eax
; X86-BMI1-NEXT:    xorl %ebx, %ebx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    movl $0, %edx
; X86-BMI1-NEXT:    jne .LBB47_4
; X86-BMI1-NEXT:  # %bb.3:
; X86-BMI1-NEXT:    movl %esi, %ebx
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:  .LBB47_4:
; X86-BMI1-NEXT:    shrdl %cl, %edi, %ebx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    jne .LBB47_6
; X86-BMI1-NEXT:  # %bb.5:
; X86-BMI1-NEXT:    movl %ebx, %eax
; X86-BMI1-NEXT:  .LBB47_6:
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    popl %edi
; X86-BMI1-NEXT:    popl %ebx
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_d0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-BMI2-NEXT:    movb $64, %cl
; X86-BMI2-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    shldl %cl, %eax, %esi
; X86-BMI2-NEXT:    shlxl %ecx, %eax, %edi
; X86-BMI2-NEXT:    xorl %edx, %edx
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB47_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %edi, %esi
; X86-BMI2-NEXT:    movl $0, %edi
; X86-BMI2-NEXT:  .LBB47_2:
; X86-BMI2-NEXT:    shrxl %ecx, %esi, %eax
; X86-BMI2-NEXT:    jne .LBB47_4
; X86-BMI2-NEXT:  # %bb.3:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:  .LBB47_4:
; X86-BMI2-NEXT:    shrdl %cl, %esi, %edi
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    jne .LBB47_6
; X86-BMI2-NEXT:  # %bb.5:
; X86-BMI2-NEXT:    movl %edi, %eax
; X86-BMI2-NEXT:  .LBB47_6:
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    popl %edi
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_d0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movq %rdi, %rax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shrq %cl, %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_d0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, %rdi, %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_d0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhiq %rsi, %rdi, %rax
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i64 64, %numlowbits
  %highbitscleared = shl i64 %val, %numhighbits
  %masked = lshr i64 %highbitscleared, %numhighbits
  ret i64 %masked
}

define i64 @bzhi64_d1_indexzext(i64 %val, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_d1_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %ebx
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    movb $64, %cl
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl %edx, %esi
; X86-NOBMI-NEXT:    shll %cl, %esi
; X86-NOBMI-NEXT:    shldl %cl, %edx, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    movl %esi, %edi
; X86-NOBMI-NEXT:    jne .LBB48_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %eax, %edi
; X86-NOBMI-NEXT:  .LBB48_2:
; X86-NOBMI-NEXT:    movl %edi, %eax
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    xorl %ebx, %ebx
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    movl $0, %edx
; X86-NOBMI-NEXT:    jne .LBB48_4
; X86-NOBMI-NEXT:  # %bb.3:
; X86-NOBMI-NEXT:    movl %esi, %ebx
; X86-NOBMI-NEXT:    movl %eax, %edx
; X86-NOBMI-NEXT:  .LBB48_4:
; X86-NOBMI-NEXT:    shrdl %cl, %edi, %ebx
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB48_6
; X86-NOBMI-NEXT:  # %bb.5:
; X86-NOBMI-NEXT:    movl %ebx, %eax
; X86-NOBMI-NEXT:  .LBB48_6:
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %edi
; X86-NOBMI-NEXT:    popl %ebx
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_d1_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %ebx
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    movb $64, %cl
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl %edx, %esi
; X86-BMI1-NEXT:    shll %cl, %esi
; X86-BMI1-NEXT:    shldl %cl, %edx, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    movl %esi, %edi
; X86-BMI1-NEXT:    jne .LBB48_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edi
; X86-BMI1-NEXT:  .LBB48_2:
; X86-BMI1-NEXT:    movl %edi, %eax
; X86-BMI1-NEXT:    shrl %cl, %eax
; X86-BMI1-NEXT:    xorl %ebx, %ebx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    movl $0, %edx
; X86-BMI1-NEXT:    jne .LBB48_4
; X86-BMI1-NEXT:  # %bb.3:
; X86-BMI1-NEXT:    movl %esi, %ebx
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:  .LBB48_4:
; X86-BMI1-NEXT:    shrdl %cl, %edi, %ebx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    jne .LBB48_6
; X86-BMI1-NEXT:  # %bb.5:
; X86-BMI1-NEXT:    movl %ebx, %eax
; X86-BMI1-NEXT:  .LBB48_6:
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    popl %edi
; X86-BMI1-NEXT:    popl %ebx
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_d1_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-BMI2-NEXT:    movb $64, %cl
; X86-BMI2-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    shldl %cl, %eax, %esi
; X86-BMI2-NEXT:    shlxl %ecx, %eax, %edi
; X86-BMI2-NEXT:    xorl %edx, %edx
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB48_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %edi, %esi
; X86-BMI2-NEXT:    movl $0, %edi
; X86-BMI2-NEXT:  .LBB48_2:
; X86-BMI2-NEXT:    shrxl %ecx, %esi, %eax
; X86-BMI2-NEXT:    jne .LBB48_4
; X86-BMI2-NEXT:  # %bb.3:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:  .LBB48_4:
; X86-BMI2-NEXT:    shrdl %cl, %esi, %edi
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    jne .LBB48_6
; X86-BMI2-NEXT:  # %bb.5:
; X86-BMI2-NEXT:    movl %edi, %eax
; X86-BMI2-NEXT:  .LBB48_6:
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    popl %edi
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_d1_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movq %rdi, %rax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrq %cl, %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_d1_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, %rdi, %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_d1_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI2-NEXT:    bzhiq %rsi, %rdi, %rax
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i8 64, %numlowbits
  %sh_prom = zext i8 %numhighbits to i64
  %highbitscleared = shl i64 %val, %sh_prom
  %masked = lshr i64 %highbitscleared, %sh_prom
  ret i64 %masked
}

define i64 @bzhi64_d2_load(i64* %w, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_d2_load:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %ebx
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    movl (%eax), %edx
; X86-NOBMI-NEXT:    movl 4(%eax), %eax
; X86-NOBMI-NEXT:    movb $64, %cl
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl %edx, %esi
; X86-NOBMI-NEXT:    shll %cl, %esi
; X86-NOBMI-NEXT:    shldl %cl, %edx, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    movl %esi, %edi
; X86-NOBMI-NEXT:    jne .LBB49_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %eax, %edi
; X86-NOBMI-NEXT:  .LBB49_2:
; X86-NOBMI-NEXT:    movl %edi, %eax
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    xorl %ebx, %ebx
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    movl $0, %edx
; X86-NOBMI-NEXT:    jne .LBB49_4
; X86-NOBMI-NEXT:  # %bb.3:
; X86-NOBMI-NEXT:    movl %esi, %ebx
; X86-NOBMI-NEXT:    movl %eax, %edx
; X86-NOBMI-NEXT:  .LBB49_4:
; X86-NOBMI-NEXT:    shrdl %cl, %edi, %ebx
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB49_6
; X86-NOBMI-NEXT:  # %bb.5:
; X86-NOBMI-NEXT:    movl %ebx, %eax
; X86-NOBMI-NEXT:  .LBB49_6:
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %edi
; X86-NOBMI-NEXT:    popl %ebx
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_d2_load:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %ebx
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    movl (%eax), %edx
; X86-BMI1-NEXT:    movl 4(%eax), %eax
; X86-BMI1-NEXT:    movb $64, %cl
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl %edx, %esi
; X86-BMI1-NEXT:    shll %cl, %esi
; X86-BMI1-NEXT:    shldl %cl, %edx, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    movl %esi, %edi
; X86-BMI1-NEXT:    jne .LBB49_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edi
; X86-BMI1-NEXT:  .LBB49_2:
; X86-BMI1-NEXT:    movl %edi, %eax
; X86-BMI1-NEXT:    shrl %cl, %eax
; X86-BMI1-NEXT:    xorl %ebx, %ebx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    movl $0, %edx
; X86-BMI1-NEXT:    jne .LBB49_4
; X86-BMI1-NEXT:  # %bb.3:
; X86-BMI1-NEXT:    movl %esi, %ebx
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:  .LBB49_4:
; X86-BMI1-NEXT:    shrdl %cl, %edi, %ebx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    jne .LBB49_6
; X86-BMI1-NEXT:  # %bb.5:
; X86-BMI1-NEXT:    movl %ebx, %eax
; X86-BMI1-NEXT:  .LBB49_6:
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    popl %edi
; X86-BMI1-NEXT:    popl %ebx
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_d2_load:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movl (%eax), %edx
; X86-BMI2-NEXT:    movl 4(%eax), %esi
; X86-BMI2-NEXT:    movb $64, %cl
; X86-BMI2-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    shldl %cl, %edx, %esi
; X86-BMI2-NEXT:    shlxl %ecx, %edx, %edi
; X86-BMI2-NEXT:    xorl %edx, %edx
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB49_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %edi, %esi
; X86-BMI2-NEXT:    movl $0, %edi
; X86-BMI2-NEXT:  .LBB49_2:
; X86-BMI2-NEXT:    shrxl %ecx, %esi, %eax
; X86-BMI2-NEXT:    jne .LBB49_4
; X86-BMI2-NEXT:  # %bb.3:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:  .LBB49_4:
; X86-BMI2-NEXT:    shrdl %cl, %esi, %edi
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    jne .LBB49_6
; X86-BMI2-NEXT:  # %bb.5:
; X86-BMI2-NEXT:    movl %edi, %eax
; X86-BMI2-NEXT:  .LBB49_6:
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    popl %edi
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_d2_load:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movq (%rdi), %rax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shrq %cl, %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_d2_load:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, (%rdi), %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_d2_load:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhiq %rsi, (%rdi), %rax
; X64-BMI2-NEXT:    retq
  %val = load i64, i64* %w
  %numhighbits = sub i64 64, %numlowbits
  %highbitscleared = shl i64 %val, %numhighbits
  %masked = lshr i64 %highbitscleared, %numhighbits
  ret i64 %masked
}

define i64 @bzhi64_d3_load_indexzext(i64* %w, i8 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_d3_load_indexzext:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %ebx
; X86-NOBMI-NEXT:    pushl %edi
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    movl (%eax), %edx
; X86-NOBMI-NEXT:    movl 4(%eax), %eax
; X86-NOBMI-NEXT:    movb $64, %cl
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl %edx, %esi
; X86-NOBMI-NEXT:    shll %cl, %esi
; X86-NOBMI-NEXT:    shldl %cl, %edx, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    movl %esi, %edi
; X86-NOBMI-NEXT:    jne .LBB50_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %eax, %edi
; X86-NOBMI-NEXT:  .LBB50_2:
; X86-NOBMI-NEXT:    movl %edi, %eax
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    xorl %ebx, %ebx
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    movl $0, %edx
; X86-NOBMI-NEXT:    jne .LBB50_4
; X86-NOBMI-NEXT:  # %bb.3:
; X86-NOBMI-NEXT:    movl %esi, %ebx
; X86-NOBMI-NEXT:    movl %eax, %edx
; X86-NOBMI-NEXT:  .LBB50_4:
; X86-NOBMI-NEXT:    shrdl %cl, %edi, %ebx
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB50_6
; X86-NOBMI-NEXT:  # %bb.5:
; X86-NOBMI-NEXT:    movl %ebx, %eax
; X86-NOBMI-NEXT:  .LBB50_6:
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    popl %edi
; X86-NOBMI-NEXT:    popl %ebx
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_d3_load_indexzext:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %ebx
; X86-BMI1-NEXT:    pushl %edi
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    movl (%eax), %edx
; X86-BMI1-NEXT:    movl 4(%eax), %eax
; X86-BMI1-NEXT:    movb $64, %cl
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl %edx, %esi
; X86-BMI1-NEXT:    shll %cl, %esi
; X86-BMI1-NEXT:    shldl %cl, %edx, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    movl %esi, %edi
; X86-BMI1-NEXT:    jne .LBB50_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %eax, %edi
; X86-BMI1-NEXT:  .LBB50_2:
; X86-BMI1-NEXT:    movl %edi, %eax
; X86-BMI1-NEXT:    shrl %cl, %eax
; X86-BMI1-NEXT:    xorl %ebx, %ebx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    movl $0, %edx
; X86-BMI1-NEXT:    jne .LBB50_4
; X86-BMI1-NEXT:  # %bb.3:
; X86-BMI1-NEXT:    movl %esi, %ebx
; X86-BMI1-NEXT:    movl %eax, %edx
; X86-BMI1-NEXT:  .LBB50_4:
; X86-BMI1-NEXT:    shrdl %cl, %edi, %ebx
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    jne .LBB50_6
; X86-BMI1-NEXT:  # %bb.5:
; X86-BMI1-NEXT:    movl %ebx, %eax
; X86-BMI1-NEXT:  .LBB50_6:
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    popl %edi
; X86-BMI1-NEXT:    popl %ebx
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_d3_load_indexzext:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    pushl %edi
; X86-BMI2-NEXT:    pushl %esi
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movl (%eax), %edx
; X86-BMI2-NEXT:    movl 4(%eax), %esi
; X86-BMI2-NEXT:    movb $64, %cl
; X86-BMI2-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    shldl %cl, %edx, %esi
; X86-BMI2-NEXT:    shlxl %ecx, %edx, %edi
; X86-BMI2-NEXT:    xorl %edx, %edx
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB50_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %edi, %esi
; X86-BMI2-NEXT:    movl $0, %edi
; X86-BMI2-NEXT:  .LBB50_2:
; X86-BMI2-NEXT:    shrxl %ecx, %esi, %eax
; X86-BMI2-NEXT:    jne .LBB50_4
; X86-BMI2-NEXT:  # %bb.3:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:  .LBB50_4:
; X86-BMI2-NEXT:    shrdl %cl, %esi, %edi
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    jne .LBB50_6
; X86-BMI2-NEXT:  # %bb.5:
; X86-BMI2-NEXT:    movl %edi, %eax
; X86-BMI2-NEXT:  .LBB50_6:
; X86-BMI2-NEXT:    popl %esi
; X86-BMI2-NEXT:    popl %edi
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_d3_load_indexzext:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movq (%rdi), %rax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrq %cl, %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_d3_load_indexzext:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, (%rdi), %rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_d3_load_indexzext:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    # kill: def $esi killed $esi def $rsi
; X64-BMI2-NEXT:    bzhiq %rsi, (%rdi), %rax
; X64-BMI2-NEXT:    retq
  %val = load i64, i64* %w
  %numhighbits = sub i8 64, %numlowbits
  %sh_prom = zext i8 %numhighbits to i64
  %highbitscleared = shl i64 %val, %sh_prom
  %masked = lshr i64 %highbitscleared, %sh_prom
  ret i64 %masked
}

; 64-bit, but with 32-bit output

; Everything done in 64-bit, truncation happens last.
define i32 @bzhi64_32_d0(i64 %val, i64 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_d0:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    pushl %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    movb $64, %cl
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    movl %esi, %edx
; X86-NOBMI-NEXT:    shll %cl, %edx
; X86-NOBMI-NEXT:    shldl %cl, %esi, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    je .LBB51_2
; X86-NOBMI-NEXT:  # %bb.1:
; X86-NOBMI-NEXT:    movl %edx, %eax
; X86-NOBMI-NEXT:    xorl %edx, %edx
; X86-NOBMI-NEXT:  .LBB51_2:
; X86-NOBMI-NEXT:    shrdl %cl, %eax, %edx
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    testb $32, %cl
; X86-NOBMI-NEXT:    jne .LBB51_4
; X86-NOBMI-NEXT:  # %bb.3:
; X86-NOBMI-NEXT:    movl %edx, %eax
; X86-NOBMI-NEXT:  .LBB51_4:
; X86-NOBMI-NEXT:    popl %esi
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_d0:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    pushl %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-BMI1-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    movb $64, %cl
; X86-BMI1-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI1-NEXT:    movl %esi, %edx
; X86-BMI1-NEXT:    shll %cl, %edx
; X86-BMI1-NEXT:    shldl %cl, %esi, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    je .LBB51_2
; X86-BMI1-NEXT:  # %bb.1:
; X86-BMI1-NEXT:    movl %edx, %eax
; X86-BMI1-NEXT:    xorl %edx, %edx
; X86-BMI1-NEXT:  .LBB51_2:
; X86-BMI1-NEXT:    shrdl %cl, %eax, %edx
; X86-BMI1-NEXT:    shrl %cl, %eax
; X86-BMI1-NEXT:    testb $32, %cl
; X86-BMI1-NEXT:    jne .LBB51_4
; X86-BMI1-NEXT:  # %bb.3:
; X86-BMI1-NEXT:    movl %edx, %eax
; X86-BMI1-NEXT:  .LBB51_4:
; X86-BMI1-NEXT:    popl %esi
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_d0:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-BMI2-NEXT:    movb $64, %cl
; X86-BMI2-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-BMI2-NEXT:    shldl %cl, %eax, %edx
; X86-BMI2-NEXT:    shlxl %ecx, %eax, %eax
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB51_2
; X86-BMI2-NEXT:  # %bb.1:
; X86-BMI2-NEXT:    movl %eax, %edx
; X86-BMI2-NEXT:    xorl %eax, %eax
; X86-BMI2-NEXT:  .LBB51_2:
; X86-BMI2-NEXT:    shrdl %cl, %edx, %eax
; X86-BMI2-NEXT:    testb $32, %cl
; X86-BMI2-NEXT:    je .LBB51_4
; X86-BMI2-NEXT:  # %bb.3:
; X86-BMI2-NEXT:    shrxl %ecx, %edx, %eax
; X86-BMI2-NEXT:  .LBB51_4:
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_d0:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movq %rsi, %rcx
; X64-NOBMI-NEXT:    movq %rdi, %rax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shlq %cl, %rax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $rcx
; X64-NOBMI-NEXT:    shrq %cl, %rax
; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_d0:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrq %rsi, %rdi, %rax
; X64-BMI1-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_d0:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhiq %rsi, %rdi, %rax
; X64-BMI2-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-BMI2-NEXT:    retq
  %numhighbits = sub i64 64, %numlowbits
  %highbitscleared = shl i64 %val, %numhighbits
  %masked = lshr i64 %highbitscleared, %numhighbits
  %res = trunc i64 %masked to i32
  ret i32 %res
}

; Shifting happens in 64-bit, then truncation. Masking is 32-bit.
define i32 @bzhi64_32_d1(i64 %val, i32 %numlowbits) nounwind {
; X86-NOBMI-LABEL: bzhi64_32_d1:
; X86-NOBMI:       # %bb.0:
; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NOBMI-NEXT:    xorl %ecx, %ecx
; X86-NOBMI-NEXT:    subb {{[0-9]+}}(%esp), %cl
; X86-NOBMI-NEXT:    shll %cl, %eax
; X86-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X86-NOBMI-NEXT:    shrl %cl, %eax
; X86-NOBMI-NEXT:    retl
;
; X86-BMI1-LABEL: bzhi64_32_d1:
; X86-BMI1:       # %bb.0:
; X86-BMI1-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI1-NEXT:    shll $8, %eax
; X86-BMI1-NEXT:    bextrl %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI1-NEXT:    retl
;
; X86-BMI2-LABEL: bzhi64_32_d1:
; X86-BMI2:       # %bb.0:
; X86-BMI2-NEXT:    movb {{[0-9]+}}(%esp), %al
; X86-BMI2-NEXT:    bzhil %eax, {{[0-9]+}}(%esp), %eax
; X86-BMI2-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_32_d1:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movl %esi, %ecx
; X64-NOBMI-NEXT:    movq %rdi, %rax
; X64-NOBMI-NEXT:    negb %cl
; X64-NOBMI-NEXT:    shll %cl, %eax
; X64-NOBMI-NEXT:    # kill: def $cl killed $cl killed $ecx
; X64-NOBMI-NEXT:    shrl %cl, %eax
; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1-LABEL: bzhi64_32_d1:
; X64-BMI1:       # %bb.0:
; X64-BMI1-NEXT:    shll $8, %esi
; X64-BMI1-NEXT:    bextrl %esi, %edi, %eax
; X64-BMI1-NEXT:    retq
;
; X64-BMI2-LABEL: bzhi64_32_d1:
; X64-BMI2:       # %bb.0:
; X64-BMI2-NEXT:    bzhil %esi, %edi, %eax
; X64-BMI2-NEXT:    retq
  %truncval = trunc i64 %val to i32
  %numhighbits = sub i32 32, %numlowbits
  %highbitscleared = shl i32 %truncval, %numhighbits
  %masked = lshr i32 %highbitscleared, %numhighbits
  ret i32 %masked
}

; ---------------------------------------------------------------------------- ;
; Constant mask
; ---------------------------------------------------------------------------- ;

; 32-bit

define i32 @bzhi32_constant_mask32(i32 %val) nounwind {
; X86-LABEL: bzhi32_constant_mask32:
; X86:       # %bb.0:
; X86-NEXT:    movl $2147483647, %eax # imm = 0x7FFFFFFF
; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    retl
;
; X64-LABEL: bzhi32_constant_mask32:
; X64:       # %bb.0:
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    andl $2147483647, %eax # imm = 0x7FFFFFFF
; X64-NEXT:    retq
  %masked = and i32 %val, 2147483647
  ret i32 %masked
}

define i32 @bzhi32_constant_mask32_load(i32* %val) nounwind {
; X86-LABEL: bzhi32_constant_mask32_load:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movl $2147483647, %eax # imm = 0x7FFFFFFF
; X86-NEXT:    andl (%ecx), %eax
; X86-NEXT:    retl
;
; X64-LABEL: bzhi32_constant_mask32_load:
; X64:       # %bb.0:
; X64-NEXT:    movl $2147483647, %eax # imm = 0x7FFFFFFF
; X64-NEXT:    andl (%rdi), %eax
; X64-NEXT:    retq
  %val1 = load i32, i32* %val
  %masked = and i32 %val1, 2147483647
  ret i32 %masked
}

define i32 @bzhi32_constant_mask16(i32 %val) nounwind {
; X86-LABEL: bzhi32_constant_mask16:
; X86:       # %bb.0:
; X86-NEXT:    movl $32767, %eax # imm = 0x7FFF
; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    retl
;
; X64-LABEL: bzhi32_constant_mask16:
; X64:       # %bb.0:
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    andl $32767, %eax # imm = 0x7FFF
; X64-NEXT:    retq
  %masked = and i32 %val, 32767
  ret i32 %masked
}

define i32 @bzhi32_constant_mask16_load(i32* %val) nounwind {
; X86-LABEL: bzhi32_constant_mask16_load:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movl $32767, %eax # imm = 0x7FFF
; X86-NEXT:    andl (%ecx), %eax
; X86-NEXT:    retl
;
; X64-LABEL: bzhi32_constant_mask16_load:
; X64:       # %bb.0:
; X64-NEXT:    movl $32767, %eax # imm = 0x7FFF
; X64-NEXT:    andl (%rdi), %eax
; X64-NEXT:    retq
  %val1 = load i32, i32* %val
  %masked = and i32 %val1, 32767
  ret i32 %masked
}

define i32 @bzhi32_constant_mask8(i32 %val) nounwind {
; X86-LABEL: bzhi32_constant_mask8:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    andl $127, %eax
; X86-NEXT:    retl
;
; X64-LABEL: bzhi32_constant_mask8:
; X64:       # %bb.0:
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    andl $127, %eax
; X64-NEXT:    retq
  %masked = and i32 %val, 127
  ret i32 %masked
}

define i32 @bzhi32_constant_mask8_load(i32* %val) nounwind {
; X86-LABEL: bzhi32_constant_mask8_load:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl (%eax), %eax
; X86-NEXT:    andl $127, %eax
; X86-NEXT:    retl
;
; X64-LABEL: bzhi32_constant_mask8_load:
; X64:       # %bb.0:
; X64-NEXT:    movl (%rdi), %eax
; X64-NEXT:    andl $127, %eax
; X64-NEXT:    retq
  %val1 = load i32, i32* %val
  %masked = and i32 %val1, 127
  ret i32 %masked
}

; 64-bit

define i64 @bzhi64_constant_mask64(i64 %val) nounwind {
; X86-LABEL: bzhi64_constant_mask64:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $1073741823, %edx # imm = 0x3FFFFFFF
; X86-NEXT:    andl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_constant_mask64:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movabsq $4611686018427387903, %rax # imm = 0x3FFFFFFFFFFFFFFF
; X64-NOBMI-NEXT:    andq %rdi, %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1NOTBM-LABEL: bzhi64_constant_mask64:
; X64-BMI1NOTBM:       # %bb.0:
; X64-BMI1NOTBM-NEXT:    movl $15872, %eax # imm = 0x3E00
; X64-BMI1NOTBM-NEXT:    bextrq %rax, %rdi, %rax
; X64-BMI1NOTBM-NEXT:    retq
;
; X64-BMI1TBM-LABEL: bzhi64_constant_mask64:
; X64-BMI1TBM:       # %bb.0:
; X64-BMI1TBM-NEXT:    bextrq $15872, %rdi, %rax # imm = 0x3E00
; X64-BMI1TBM-NEXT:    retq
;
; X64-BMI2TBM-LABEL: bzhi64_constant_mask64:
; X64-BMI2TBM:       # %bb.0:
; X64-BMI2TBM-NEXT:    bextrq $15872, %rdi, %rax # imm = 0x3E00
; X64-BMI2TBM-NEXT:    retq
;
; X64-BMI2NOTBM-LABEL: bzhi64_constant_mask64:
; X64-BMI2NOTBM:       # %bb.0:
; X64-BMI2NOTBM-NEXT:    movb $62, %al
; X64-BMI2NOTBM-NEXT:    bzhiq %rax, %rdi, %rax
; X64-BMI2NOTBM-NEXT:    retq
  %masked = and i64 %val, 4611686018427387903
  ret i64 %masked
}

define i64 @bzhi64_constant_mask64_load(i64* %val) nounwind {
; X86-LABEL: bzhi64_constant_mask64_load:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movl (%ecx), %eax
; X86-NEXT:    movl $1073741823, %edx # imm = 0x3FFFFFFF
; X86-NEXT:    andl 4(%ecx), %edx
; X86-NEXT:    retl
;
; X64-NOBMI-LABEL: bzhi64_constant_mask64_load:
; X64-NOBMI:       # %bb.0:
; X64-NOBMI-NEXT:    movabsq $4611686018427387903, %rax # imm = 0x3FFFFFFFFFFFFFFF
; X64-NOBMI-NEXT:    andq (%rdi), %rax
; X64-NOBMI-NEXT:    retq
;
; X64-BMI1NOTBM-LABEL: bzhi64_constant_mask64_load:
; X64-BMI1NOTBM:       # %bb.0:
; X64-BMI1NOTBM-NEXT:    movl $15872, %eax # imm = 0x3E00
; X64-BMI1NOTBM-NEXT:    bextrq %rax, (%rdi), %rax
; X64-BMI1NOTBM-NEXT:    retq
;
; X64-BMI1TBM-LABEL: bzhi64_constant_mask64_load:
; X64-BMI1TBM:       # %bb.0:
; X64-BMI1TBM-NEXT:    bextrq $15872, (%rdi), %rax # imm = 0x3E00
; X64-BMI1TBM-NEXT:    retq
;
; X64-BMI2TBM-LABEL: bzhi64_constant_mask64_load:
; X64-BMI2TBM:       # %bb.0:
; X64-BMI2TBM-NEXT:    bextrq $15872, (%rdi), %rax # imm = 0x3E00
; X64-BMI2TBM-NEXT:    retq
;
; X64-BMI2NOTBM-LABEL: bzhi64_constant_mask64_load:
; X64-BMI2NOTBM:       # %bb.0:
; X64-BMI2NOTBM-NEXT:    movb $62, %al
; X64-BMI2NOTBM-NEXT:    bzhiq %rax, (%rdi), %rax
; X64-BMI2NOTBM-NEXT:    retq
  %val1 = load i64, i64* %val
  %masked = and i64 %val1, 4611686018427387903
  ret i64 %masked
}

define i64 @bzhi64_constant_mask32(i64 %val) nounwind {
; X86-LABEL: bzhi64_constant_mask32:
; X86:       # %bb.0:
; X86-NEXT:    movl $2147483647, %eax # imm = 0x7FFFFFFF
; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    xorl %edx, %edx
; X86-NEXT:    retl
;
; X64-LABEL: bzhi64_constant_mask32:
; X64:       # %bb.0:
; X64-NEXT:    movq %rdi, %rax
; X64-NEXT:    andl $2147483647, %eax # imm = 0x7FFFFFFF
; X64-NEXT:    retq
  %masked = and i64 %val, 2147483647
  ret i64 %masked
}

define i64 @bzhi64_constant_mask32_load(i64* %val) nounwind {
; X86-LABEL: bzhi64_constant_mask32_load:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movl $2147483647, %eax # imm = 0x7FFFFFFF
; X86-NEXT:    andl (%ecx), %eax
; X86-NEXT:    xorl %edx, %edx
; X86-NEXT:    retl
;
; X64-LABEL: bzhi64_constant_mask32_load:
; X64:       # %bb.0:
; X64-NEXT:    movq (%rdi), %rax
; X64-NEXT:    andl $2147483647, %eax # imm = 0x7FFFFFFF
; X64-NEXT:    retq
  %val1 = load i64, i64* %val
  %masked = and i64 %val1, 2147483647
  ret i64 %masked
}

define i64 @bzhi64_constant_mask16(i64 %val) nounwind {
; X86-LABEL: bzhi64_constant_mask16:
; X86:       # %bb.0:
; X86-NEXT:    movl $32767, %eax # imm = 0x7FFF
; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    xorl %edx, %edx
; X86-NEXT:    retl
;
; X64-LABEL: bzhi64_constant_mask16:
; X64:       # %bb.0:
; X64-NEXT:    movq %rdi, %rax
; X64-NEXT:    andl $32767, %eax # imm = 0x7FFF
; X64-NEXT:    retq
  %masked = and i64 %val, 32767
  ret i64 %masked
}

define i64 @bzhi64_constant_mask16_load(i64* %val) nounwind {
; X86-LABEL: bzhi64_constant_mask16_load:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movl $32767, %eax # imm = 0x7FFF
; X86-NEXT:    andl (%ecx), %eax
; X86-NEXT:    xorl %edx, %edx
; X86-NEXT:    retl
;
; X64-LABEL: bzhi64_constant_mask16_load:
; X64:       # %bb.0:
; X64-NEXT:    movq (%rdi), %rax
; X64-NEXT:    andl $32767, %eax # imm = 0x7FFF
; X64-NEXT:    retq
  %val1 = load i64, i64* %val
  %masked = and i64 %val1, 32767
  ret i64 %masked
}

define i64 @bzhi64_constant_mask8(i64 %val) nounwind {
; X86-LABEL: bzhi64_constant_mask8:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    andl $127, %eax
; X86-NEXT:    xorl %edx, %edx
; X86-NEXT:    retl
;
; X64-LABEL: bzhi64_constant_mask8:
; X64:       # %bb.0:
; X64-NEXT:    movq %rdi, %rax
; X64-NEXT:    andl $127, %eax
; X64-NEXT:    retq
  %masked = and i64 %val, 127
  ret i64 %masked
}

define i64 @bzhi64_constant_mask8_load(i64* %val) nounwind {
; X86-LABEL: bzhi64_constant_mask8_load:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl (%eax), %eax
; X86-NEXT:    andl $127, %eax
; X86-NEXT:    xorl %edx, %edx
; X86-NEXT:    retl
;
; X64-LABEL: bzhi64_constant_mask8_load:
; X64:       # %bb.0:
; X64-NEXT:    movq (%rdi), %rax
; X64-NEXT:    andl $127, %eax
; X64-NEXT:    retq
  %val1 = load i64, i64* %val
  %masked = and i64 %val1, 127
  ret i64 %masked
}