llvm dev - Mar 2019 - Condition removed? Difference between LLVM and GCC on a small testcase

Hello Dev,

I have a very simple testcase, which shows strange difference between LLVM and
GCC. Does anyone know which optimization pass removes the condition? Thanks!

C code:
extern void bar(int, int);

void foo(int a) {
  int b, d;

  if (a > 114) {
    b = a * 58;
  } else {
    d = a * 51;
  }
  bar(b, d);
}

clang.7.0.1 -O2, LLVM generated assembly:
   0:   6b c7 3a                     imul   $0x3a,%edi,%eax
   3:   6b f7 33                      imul   $0x33,%edi,%esi
   6:   89 c7                           mov    %eax,%edi
   8:   e9 00 00 00 00         jmpq   d <foo+0xd>

GCC.5.2.0 -O3, GCC generated assembly:
   0:   83 ff 72                       cmp    $0x72,%edi
   3:   7f 0b                           jg     10 <foo+0x10>
   5:   6b f7 33                      imul   $0x33,%edi,%esi
   8:   89 c7                           mov    %eax,%edi
   a:   e9 00 00 00 00         jmpq   f <foo+0xf>
   f:   90                                  nop
  10:   6b c7 3a                    imul   $0x3a,%edi,%eax
  13:   89 c7                          mov    %eax,%edi
  15:   e9 00 00 00 00       jmpq   1a <foo+0x1a>



Regards,
Yan
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190301/6998ceb9/attachment-0001.html>

On Fri, 1 Mar 2019 at 14:38, Yan Luo via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> I have a very simple testcase, which shows strange difference between LLVM
and GCC. Does anyone know which optimization pass removes the condition?
SimplifyCFG converts branching conditions into equivalents with the
LLVM "select" instruction (think cmov on x86), and then CSE (common
subexpression elimination) removes the select entirely on the grounds
that the other branch is undefined (it could in principle be anything,
so a*51 or a*58 is as good a value as any to assume).

Cheers.

Tim.

Tim,

Thanks for the reply! It makes sense that LLVM selects a*51 or a*58 for an
undefined value although it makes the arguments to bar()
"predictable".


Regards,
Yan  

-----Original Message-----
From: Tim Northover <t.p.northover at gmail.com> 
Sent: Friday, March 1, 2019 3:54 PM
To: Yan Luo <yan.luo2 at synopsys.com>
Cc: llvm-dev at lists.llvm.org
Subject: Re: [llvm-dev] Condition removed? Difference between LLVM and GCC on a
small testcase

On Fri, 1 Mar 2019 at 14:38, Yan Luo via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> I have a very simple testcase, which shows strange difference between LLVM
and GCC. Does anyone know which optimization pass removes the condition?
SimplifyCFG converts branching conditions into equivalents with the LLVM
"select" instruction (think cmov on x86), and then CSE (common
subexpression elimination) removes the select entirely on the grounds that the
other branch is undefined (it could in principle be anything, so a*51 or a*58 is
as good a value as any to assume).

Cheers.

Tim.