test/CodeGen/X86/atomic-cmp-fold.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu | FileCheck %s --check-prefix=X64
; RUN: llc < %s -mtriple=i686-unknown-linux-gnu | FileCheck %s --check-prefix=X86

; Compares of atomic loads should fold the load into the compare's memory
; operand, like compares of regular loads do: the folded form still performs
; a single full-width load, and on x86 an ordinary load already provides
; every ordering up to seq_cst.

define i1 @cmp8_eq0_monotonic(ptr %p) {
; X64-LABEL: cmp8_eq0_monotonic:
; X64:       # %bb.0:
; X64-NEXT:    cmpb $0, (%rdi)
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp8_eq0_monotonic:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    cmpb $0, (%eax)
; X86-NEXT:    sete %al
; X86-NEXT:    retl
  %v = load atomic i8, ptr %p monotonic, align 1
  %c = icmp eq i8 %v, 0
  ret i1 %c
}

define i1 @cmp16_eq0_acquire(ptr %p) {
; X64-LABEL: cmp16_eq0_acquire:
; X64:       # %bb.0:
; X64-NEXT:    cmpw $0, (%rdi)
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp16_eq0_acquire:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    cmpw $0, (%eax)
; X86-NEXT:    sete %al
; X86-NEXT:    retl
  %v = load atomic i16, ptr %p acquire, align 2
  %c = icmp eq i16 %v, 0
  ret i1 %c
}

define i1 @cmp32_eq0_seq_cst(ptr %p) {
; X64-LABEL: cmp32_eq0_seq_cst:
; X64:       # %bb.0:
; X64-NEXT:    cmpl $0, (%rdi)
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp32_eq0_seq_cst:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    cmpl $0, (%eax)
; X86-NEXT:    sete %al
; X86-NEXT:    retl
  %v = load atomic i32, ptr %p seq_cst, align 4
  %c = icmp eq i32 %v, 0
  ret i1 %c
}

define i1 @cmp64_eq0_unordered(ptr %p) {
; X64-LABEL: cmp64_eq0_unordered:
; X64:       # %bb.0:
; X64-NEXT:    cmpq $0, (%rdi)
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp64_eq0_unordered:
; X86:       # %bb.0:
; X86-NEXT:    subl $12, %esp
; X86-NEXT:    .cfi_def_cfa_offset 16
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    fildll (%eax)
; X86-NEXT:    fistpll (%esp)
; X86-NEXT:    movl (%esp), %eax
; X86-NEXT:    orl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    sete %al
; X86-NEXT:    addl $12, %esp
; X86-NEXT:    .cfi_def_cfa_offset 4
; X86-NEXT:    retl
  %v = load atomic i64, ptr %p unordered, align 8
  %c = icmp eq i64 %v, 0
  ret i1 %c
}

define i1 @cmp32_imm_acquire(ptr %p) {
; X64-LABEL: cmp32_imm_acquire:
; X64:       # %bb.0:
; X64-NEXT:    cmpl $42, (%rdi)
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp32_imm_acquire:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    cmpl $42, (%eax)
; X86-NEXT:    sete %al
; X86-NEXT:    retl
  %v = load atomic i32, ptr %p acquire, align 4
  %c = icmp eq i32 %v, 42
  ret i1 %c
}

define i1 @cmp32_imm_large(ptr %p) {
; X64-LABEL: cmp32_imm_large:
; X64:       # %bb.0:
; X64-NEXT:    cmpl $305419896, (%rdi) # imm = 0x12345678
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp32_imm_large:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    cmpl $305419896, (%eax) # imm = 0x12345678
; X86-NEXT:    sete %al
; X86-NEXT:    retl
  %v = load atomic i32, ptr %p monotonic, align 4
  %c = icmp eq i32 %v, 305419896 ; 0x12345678
  ret i1 %c
}

define i1 @cmp64_imm_sext32(ptr %p) {
; X64-LABEL: cmp64_imm_sext32:
; X64:       # %bb.0:
; X64-NEXT:    cmpq $-100, (%rdi)
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp64_imm_sext32:
; X86:       # %bb.0:
; X86-NEXT:    subl $12, %esp
; X86-NEXT:    .cfi_def_cfa_offset 16
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    fildll (%eax)
; X86-NEXT:    fistpll (%esp)
; X86-NEXT:    movl (%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    notl %ecx
; X86-NEXT:    xorl $-100, %eax
; X86-NEXT:    orl %ecx, %eax
; X86-NEXT:    sete %al
; X86-NEXT:    addl $12, %esp
; X86-NEXT:    .cfi_def_cfa_offset 4
; X86-NEXT:    retl
  %v = load atomic i64, ptr %p monotonic, align 8
  %c = icmp eq i64 %v, -100
  ret i1 %c
}

; The immediate does not fit CMP64mi32, so the load stays in a register.
define i1 @cmp64_imm_too_wide(ptr %p) {
; X64-LABEL: cmp64_imm_too_wide:
; X64:       # %bb.0:
; X64-NEXT:    movabsq $4886718345, %rax # imm = 0x123456789
; X64-NEXT:    cmpq %rax, (%rdi)
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp64_imm_too_wide:
; X86:       # %bb.0:
; X86-NEXT:    subl $12, %esp
; X86-NEXT:    .cfi_def_cfa_offset 16
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    fildll (%eax)
; X86-NEXT:    fistpll (%esp)
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $591751049, %ecx # imm = 0x23456789
; X86-NEXT:    xorl (%esp), %ecx
; X86-NEXT:    xorl $1, %eax
; X86-NEXT:    orl %ecx, %eax
; X86-NEXT:    sete %al
; X86-NEXT:    addl $12, %esp
; X86-NEXT:    .cfi_def_cfa_offset 4
; X86-NEXT:    retl
  %v = load atomic i64, ptr %p monotonic, align 8
  %c = icmp eq i64 %v, 4886718345 ; 0x123456789
  ret i1 %c
}

define i1 @cmp32_reg(ptr %p, i32 %x) {
; X64-LABEL: cmp32_reg:
; X64:       # %bb.0:
; X64-NEXT:    cmpl %esi, (%rdi)
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp32_reg:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl (%eax), %eax
; X86-NEXT:    cmpl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    sete %al
; X86-NEXT:    retl
  %v = load atomic i32, ptr %p acquire, align 4
  %c = icmp eq i32 %v, %x
  ret i1 %c
}

; The loaded value sits on the compare's RHS.
define i1 @cmp32_reg_rhs(ptr %p, i32 %x) {
; X64-LABEL: cmp32_reg_rhs:
; X64:       # %bb.0:
; X64-NEXT:    cmpl (%rdi), %esi
; X64-NEXT:    setl %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp32_reg_rhs:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl (%eax), %eax
; X86-NEXT:    cmpl %eax, {{[0-9]+}}(%esp)
; X86-NEXT:    setl %al
; X86-NEXT:    retl
  %v = load atomic i32, ptr %p acquire, align 4
  %c = icmp slt i32 %x, %v
  ret i1 %c
}

; Relaxed load of a level, compare, branch (the log-level-check shape).
define i32 @cmp32_sge_branch(ptr %p, i32 %lvl) {
; X64-LABEL: cmp32_sge_branch:
; X64:       # %bb.0:
; X64-NEXT:    cmpl %esi, (%rdi)
; X64-NEXT:    jl .LBB10_2
; X64-NEXT:  # %bb.1: # %log
; X64-NEXT:    movl $1, %eax
; X64-NEXT:    retq
; X64-NEXT:  .LBB10_2: # %skip
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    retq
;
; X86-LABEL: cmp32_sge_branch:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl (%eax), %eax
; X86-NEXT:    cmpl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    jl .LBB10_2
; X86-NEXT:  # %bb.1: # %log
; X86-NEXT:    movl $1, %eax
; X86-NEXT:    retl
; X86-NEXT:  .LBB10_2: # %skip
; X86-NEXT:    xorl %eax, %eax
; X86-NEXT:    retl
  %v = load atomic i32, ptr %p monotonic, align 4
  %c = icmp sge i32 %v, %lvl
  br i1 %c, label %log, label %skip
log:
  ret i32 1
skip:
  ret i32 0
}

define i1 @test32_mask_bit(ptr %p) {
; X64-LABEL: test32_mask_bit:
; X64:       # %bb.0:
; X64-NEXT:    movl (%rdi), %eax
; X64-NEXT:    andl $8, %eax
; X64-NEXT:    shrl $3, %eax
; X64-NEXT:    # kill: def $al killed $al killed $eax
; X64-NEXT:    retq
;
; X86-LABEL: test32_mask_bit:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl (%eax), %eax
; X86-NEXT:    andl $8, %eax
; X86-NEXT:    shrl $3, %eax
; X86-NEXT:    # kill: def $al killed $al killed $eax
; X86-NEXT:    retl
  %v = load atomic i32, ptr %p monotonic, align 4
  %a = and i32 %v, 8
  %c = icmp ne i32 %a, 0
  ret i1 %c
}

; The mask fits neither i8 nor i16 and is not a shifted mask, so the
; masked compare cannot be narrowed and keeps its full width.
define i1 @test32_mask_wide(ptr %p) {
; X64-LABEL: test32_mask_wide:
; X64:       # %bb.0:
; X64-NEXT:    movl (%rdi), %eax
; X64-NEXT:    testl $65792, %eax # imm = 0x10100
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: test32_mask_wide:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl (%eax), %eax
; X86-NEXT:    testl $65792, %eax # imm = 0x10100
; X86-NEXT:    sete %al
; X86-NEXT:    retl
  %v = load atomic i32, ptr %p monotonic, align 4
  %a = and i32 %v, 65792 ; 0x10100
  %c = icmp eq i32 %a, 0
  ret i1 %c
}

; The loaded value has a second use: the load must stay.
define i32 @no_fold_multiuse(ptr %p) {
; X64-LABEL: no_fold_multiuse:
; X64:       # %bb.0:
; X64-NEXT:    movl (%rdi), %ecx
; X64-NEXT:    cmpl $7, %ecx
; X64-NEXT:    movl $100, %eax
; X64-NEXT:    cmovnel %ecx, %eax
; X64-NEXT:    retq
;
; X86-LABEL: no_fold_multiuse:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl (%eax), %ecx
; X86-NEXT:    cmpl $7, %ecx
; X86-NEXT:    movl $100, %eax
; X86-NEXT:    je .LBB13_2
; X86-NEXT:  # %bb.1:
; X86-NEXT:    movl %ecx, %eax
; X86-NEXT:  .LBB13_2:
; X86-NEXT:    retl
  %v = load atomic i32, ptr %p acquire, align 4
  %c = icmp eq i32 %v, 7
  %r = select i1 %c, i32 100, i32 %v
  ret i32 %r
}

; Two atomic loads: at most one side can fold.
define i1 @cmp_two_loads(ptr %p, ptr %q) {
; X64-LABEL: cmp_two_loads:
; X64:       # %bb.0:
; X64-NEXT:    movl (%rdi), %eax
; X64-NEXT:    cmpl (%rsi), %eax
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp_two_loads:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movl (%ecx), %ecx
; X86-NEXT:    cmpl (%eax), %ecx
; X86-NEXT:    sete %al
; X86-NEXT:    retl
  %a = load atomic i32, ptr %p acquire, align 4
  %b = load atomic i32, ptr %q acquire, align 4
  %c = icmp eq i32 %a, %b
  ret i1 %c
}

; Volatile atomic load: folding keeps the single access, matching how
; compares of volatile non-atomic loads are handled.
define i1 @cmp32_volatile(ptr %p) {
; X64-LABEL: cmp32_volatile:
; X64:       # %bb.0:
; X64-NEXT:    cmpl $0, (%rdi)
; X64-NEXT:    sete %al
; X64-NEXT:    retq
;
; X86-LABEL: cmp32_volatile:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    cmpl $0, (%eax)
; X86-NEXT:    sete %al
; X86-NEXT:    retl
  %v = load atomic volatile i32, ptr %p acquire, align 4
  %c = icmp eq i32 %v, 0
  ret i1 %c
}