[llvm-bugs] [Bug 36811] New: Inline assembly input operand inefficiency

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=36811

            Bug ID: 36811
           Summary: Inline assembly input operand inefficiency
           Product: libraries
           Version: 6.0
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: nruslan_devel at yahoo.com
                CC: llvm-bugs at lists.llvm.org

I have noticed that whenever "mr" is specified for an assembly input operand
which may accept both memory and register, clang/llvm does not seem to generate
efficient code.

(-O2 is used for all examples)

For example,

1.

unsigned long func(unsigned long x)
{
    unsigned long r;
    asm ("bsf %1, %0"
        : "=r" (r)
        : "mr" (x)
        : "cc");
    return r;
}

generates code which unnecessarily moves %rdi to memory

func:                                   # @func
    .cfi_startproc
# %bb.0:
    movq    %rdi, -8(%rsp)
    #APP
    bsfq    -8(%rsp), %rax
    #NO_APP
    retq

2. whereas, if we change "mr" to simply "r" (for x)

we get optimal code

func:                                   # @func
    .cfi_startproc
# %bb.0:
    #APP
    bsfq    %rdi, %rax
    #NO_APP
    retq


3. gcc generates optimal code in both cases

func:
.LFB0:
    .cfi_startproc
#APP
# 4 "1.c" 1
    bsf %rdi, %rax
# 0 "" 2
#NO_APP
    ret

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20180320/8e9e467a/attachment.html>