[Bug rtl-optimization/53705] wrong code with custom flags - stores to memory are lost
amylaar at gcc dot gnu.org
gcc-bugzilla@gcc.gnu.org
Tue Jul 30 13:37:00 GMT 2013
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53705
Jorn Wolfgang Rennecke <amylaar at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amylaar at gcc dot gnu.org
--- Comment #3 from Jorn Wolfgang Rennecke <amylaar at gcc dot gnu.org> ---
(In reply to Zdenek Sojka from comment #0)
> Created attachment 27640 [details]
> reduced testcase (from testsuite/gcc.c-torture/execute/loop-2e.c)
>
> GCC 4.3-4.8 fails, while with GCC 4.2, the code is reduced just to return 0.
> I am not sure if this should be marked as a regression.
>
> Compiler output:
> $ gcc -O2 -fno-omit-frame-pointer -fpeel-loops -fsched2-use-superblocks
> -fno-tree-loop-optimize -fno-web --param=max-completely-peel-times=256
> testcase.c
> $ ./a.out
> Aborted
>
> Looking at the assembly:
> ...
> sub rsp, 320 #,
> mov rdx, QWORD PTR p[rip] # p.0, p
> lea rcx, [rbp-320] # q,
> ...
> mov QWORD PTR [rcx+240], rdi # *q_19, tmp73
> lea rax, [rdx+144] # tmp73,
> lea rsi, [rdx+148] # tmp73,
> lea rdi, [rdx+152] # tmp73,
> lea rdx, [rdx+156] # tmp73,
> cmp QWORD PTR [rbp-8], rdx # q, tmp73
> mov QWORD PTR [rcx+280], rax # *q_19, tmp73
> mov QWORD PTR [rcx+248], r8 # *q_19, tmp73
> mov QWORD PTR [rcx+256], r9 # *q_19, tmp73
> mov QWORD PTR [rcx+264], r10 # *q_19, tmp73
> mov QWORD PTR [rcx+272], r11 # *q_19, tmp73
> mov QWORD PTR [rcx+288], rsi # *q_19, tmp73
> mov QWORD PTR [rcx+296], rdi # *q_19, tmp73
> mov QWORD PTR [rcx+304], rdx # *q_19, tmp73
> jne .L46 #,
> ...
> It seems the dependency between storing to *q++ (esp. q[39]) in foo() and
> verifying it in main() is lost. Furthermore, it seems q[39] ([rcx+312]) is
> not stored to at all.
I can't reproduce the unrolling with an i686-pc-linux-gnu X x86_64-pc-linux-gnu
compiler, neither for 4.8.2 20130729 (prerelease) , nor for 4.9.0 20130729
(experimental) sources.
Is that an issuse with the cross-compiler, or with the compiler version?
More information about the Gcc-bugs
mailing list