------- Comment #3 from zackw at panix dot com 2006-07-12 22:42 -------
I should mention that the exact command line flags were -O2
-fomit-frame-pointer -march=pentium4, and that I hand-tweaked the label numbers
for ease of reading.
Also, -fno-tree-ch does suppress this bad optimization, but in exchange we get
mildly worse code from the loop optimizer proper - it uses [reg+reg] indexing
and a 0..n count instead of [reg] indexing and a base..limit count. The code
is pretty short so I'll just paste it here (meaningless labels removed):
_Z17has_bad_chars_newRKSs:
pushl %ebx
movl 8(%esp), %eax
movl (%eax), %eax
xorl %ecx, %ecx
movl -12(%eax), %ebx
.L2:
cmpl %ecx, %ebx
je .L10
movzbl (%ecx,%eax), %edx
cmpb $31, %dl
jbe .L4
cmpb $92, %dl
je .L4
addl $1, %ecx
cmpb $127, %dl
jne .L2
.L4:
movl $1, %eax
popl %ebx
.p2align 4,,2
ret
.L10:
xorl %eax, %eax
popl %ebx
.p2align 4,,2
ret
Looking at the code, I see that the entire purpose of tree-ch is to duplicate
loop bodies in this fashion, and the justification given is that it "increases
effectiveness of code motion and reduces the need for loop preconditioning",
which I take to cover the above degradation in addressing mode choice. I'm not
an optimizer expert, but surely there is a way to get the best of both worlds
here...?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28364
