Using gcc-4.3-20061223 targeting i386-linux and options "-O9
-fomit-frame-pointer -fexpensive-optimizations -S", my test program to detect
unsigned integer overflow:
extern void abort(void);
unsigned foo(unsigned a, unsigned b) {
unsigned sum = a + b;
if (sum < a) abort(); /* check for overflow (wrapping) */
if (sum < b) abort(); /* redundant */
return sum;
}
compiles to:
foo:
subl $12, %esp
movl 16(%esp), %eax
movl 20(%esp), %edx
addl %eax, %edx
cmpl %edx, %eax
ja .L7
cmpl %edx, 20(%esp)
ja .L7
movl %edx, %eax
addl $12, %esp
ret
.L7:
call abort
After the addition, which sets the ALU flags, the compiler issues two compare
instructions and conditional branches. This sequence could be replaced by a
conditional branch following the addl, testing one of the flags (overflow?
carry? I forget which) set by it.
I realize for other processors (e.g., with -march=pentiumpro), the addl may be
replaced by something like leal, which may not set the flags the same way. But
if the flags are set when building for a certain architecture, use them....
And is leal+cmp with data dependence still going to be better than addl?
(Also, I think a different set of register selections could eliminate the last
"movl" instruction, by putting the result of the addition into the return-value
register.)
--
Summary: optimize unsigned-add overflow test on x86 to use cpu
flags from addl
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: target
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: raeburn at raeburn dot org
GCC build triplet: powerpc-apple-darwin
GCC host triplet: powerpc-apple-darwin
GCC target triplet: i386-unknown-linux
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30315
