http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16541
Steven Bosscher <steven at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |RESOLVED
Resolution| |FIXED
--- Comment #6 from Steven Bosscher <steven at gcc dot gnu.org> 2011-05-22
15:21:05 UTC ---
(In reply to comment #2)
> Here is a simplified test case that shows a code quality regression:
>
> extern unsigned char first_one[65536];
> int FirstOnet(unsigned long long arg1)
> {
> if (arg1 >> 48)
> return (first_one[arg1 >> 48]);
> return 0;
> }
>
> the code generated by gcc-3.0 -O2 -fomit-frame-pointer is:
>
> movl 8(%esp), %edx
> movl %edx, %eax
> shrl $16, %eax
> xorl %edx, %edx
> movl %eax, %ecx
> orl %edx, %ecx
> je .L3
> movzbl first_one(%eax), %eax
> ret
> .p2align 2
> .L3:
> xorl %eax, %eax
> ret
>
> and by mainline (a bit worse):
>
> pushl %ebx <- using a callee saved register
> xorl %ecx, %ecx
> movl 12(%esp), %edx
> movl %edx, %eax <- why not load directly to eax?
> xorl %edx, %edx
> shrl $16, %eax
> movl %edx, %ebx
> orl %eax, %ebx
> je .L4
> movzbl first_one(%eax), %ecx
> .L4:
> popl %ebx
> movl %ecx, %eax
> ret
>
>
> Here is what Intel's compiler generates:
>
> movzwl 10(%esp), %edx #28.5
> xorl %eax, %eax #30.3
> orl %edx, %eax #30.3
> je ..B1.3 # Prob 50% #30.3
> # LOE edx ebx ebp esi edi
> ..B1.2: # Preds ..B1.1
> movzbl first_one(%edx), %eax #31.13
> ret #31.13
> # LOE
> ..B1.3: # Preds ..B1.1
> xorl %eax, %eax #32.10
> ret #32.10
>
>
>
$ cc1 -quiet -m32 -O2 t.c -fdump-tree-optimized
$ cat t.s
.file "t.c"
.text
.p2align 4,,15
.globl FirstOnet
.type FirstOnet, @function
FirstOnet:
.LFB0:
.cfi_startproc
movzwl 10(%esp), %edx
xorl %eax, %eax
testl %edx, %edx
je .L2
movzbl first_one(%edx), %eax
.L2:
rep
ret
.cfi_endproc
.LFE0:
.size FirstOnet, .-FirstOnet
.ident "GCC: (GNU) 4.6.0 20110312 (experimental) [trunk revision
170907]"
.section .note.GNU-stack,"",@progbits
