https://bugs.llvm.org/show_bug.cgi?id=34565
Bug ID: 34565
Summary: X86CmovConversion ineffective when debug info is
enabled
Product: libraries
Version: trunk
Hardware: PC
OS: Linux
Status: NEW
Severity: normal
Priority: P
Component: Backend: X86
Assignee: unassignedb...@nondot.org
Reporter: matthew.v...@sony.com
CC: llvm-bugs@lists.llvm.org
rL308142 added a new optimization that converts cmov instructions to
combinations of
branch and mov instructions when the branch is well predicted
Some of our internal testing highlighted a case where this optimization is
ineffective
when debug info is generated. That is, when debug info is enabled, the cmov
instruction
is still generated.
A reduced IR repro is attached, both with and without debug info. Also, I've
included the
C++ file that was used to generate both of them.
Reproduce using:
$ clang -O2 -c g.ll ; objdump -d --no-show-raw-insn g.o > g.dis
$ clang -O2 -c no_g.ll ; objdump -d --no-show-raw-insn no_g.o > no_g.dis
$ sdiff -w74 g.dis no_g.dis
g.o: file format elf64-x86-64 | no_g.o: file format elf64-x86-
Disassembly of section .text.main: Disassembly of section .text.main:
0000000000000000 <main>: 0000000000000000 <main>:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: push %r15 4: push %r15
6: push %r14 6: push %r14
8: push %r12 8: push %r12
a: push %rbx a: push %rbx
b: mov $0xa,%r14d | b: mov $0xa,%r15d
11: xor %ebx,%ebx | 11: xor %r12d,%r12d
13: lea 0x0(%rip),%r15 | 14: lea 0x0(%rip),%r14
1a: nopw 0x0(%rax,%rax,1) | 1b: nopl 0x0(%rax,%rax,1)
20: lea (%r14,%rbx,1),%r12d | 20: lea (%r15,%r12,1),%ebx
24: callq 29 <main+0x29> 24: callq 29 <main+0x29>
29: cltq 29: cltq
>>> 2b: cmp %rax,(%r15,%r12,8) | 2b: cmp %rax,(%r14,%rbx,8)
>>> 2f: cmovae %r12d,%r14d | 2f: jb 40 <main+0x40>
>>> 33: cmovb %r12d,%ebx | 31: mov %ebx,%r15d
37: cmp %r14d,%ebx | 34: cmp %r15d,%r12d
3a: jb 20 <main+0x20> | 37: jb 20 <main+0x20>
3c: xor %eax,%eax | 39: jmp 48 <main+0x48>
3e: pop %rbx | 3b: nopl 0x0(%rax,%rax,1)
3f: pop %r12 | 40: mov %ebx,%r12d
41: pop %r14 | 43: cmp %r15d,%r12d
43: pop %r15 | 46: jb 20 <main+0x20>
45: pop %rbp | 48: xor %eax,%eax
46: retq | 4a: pop %rbx
> 4b: pop %r12
> 4d: pop %r14
> 4f: pop %r15
> 51: pop %rbp
> 52: retq
$
--
You are receiving this mail because:
You are on the CC list for the bug._______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
