[llvm-bugs] [Bug 34282] New: Failure to remove redundant check

via llvm-bugs llvm-bugs at lists.llvm.org
Tue Aug 22 08:52:37 PDT 2017


https://bugs.llvm.org/show_bug.cgi?id=34282

            Bug ID: 34282
           Summary: Failure to remove redundant check
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: aivchenk at gmail.com
                CC: llvm-bugs at lists.llvm.org

For the following function:
int foo(char* C,  int a, int b, int c, int d, int e, int f) {
  int y1 = C[0] ? a : b;
  int y2 = C[0] || C[1] ? c : d;
  int y3 = C[0] || (C[1] && C[2]) ? e : f;
  return C[0] ? y1 + y2 : C[1] ? y2 : y3;
}
> 
.. we generate bad code:

# BB#0:
        cmpb    $0, (%rdi)
        je      .LBB0_1
# BB#5:
        addl    %esi, %ecx
        movl    %ecx, %eax
        retq
.LBB0_1:
        movl    8(%rsp), %eax
        movb    1(%rdi), %dl
        testb   %dl, %dl
        je      .LBB0_3
# BB#2:
        cmpb    $0, 2(%rdi)
        cmovnel %r9d, %eax
.LBB0_3:
        testb   %dl, %dl
        cmovel  %eax, %ecx
        movl    %ecx, %eax
        retq

Just for comparison, here is GCC version:
        cmpb    $0, (%rdi)                                                      
        leal    (%rsi,%rcx), %eax                                               
        jne     .L1                                                             
        cmpb    $0, 1(%rdi)                                                     
        movl    %ecx, %eax                                                      
        cmove   8(%rsp), %eax                                                   
.L1:                                                                            
        rep ret

The main problem with LLVM code is that we do redundant "cmpb    $0, 2(%rdi)"
check (while GCC removed that): C[2] does not influence the function return
value. As far as I can say, all other problems with our code, like doing testb
twice for cmov condition, are the result of not deleting the C[2] check

-- 
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/20170822/46f1063f/attachment.html>


More information about the llvm-bugs mailing list