------- Comment #4 from ajrobb at bigfoot dot com 2008-08-31 07:48 -------
I should point out that for operations (x * y / z) where y > z, there is
nothing to be gained on an Intel Core 2 duo (as extra adds and a shrd are
required for the scaling to work over the whole domain: 0..(1<<32)/y. However,
there are still gains to be had where y < z and a simple 32-bit scale factor
can be used.
In the following example, fast is at least 20% faster than std with gcc 3.4
(Cygwin) when compiled without inlining (gcc -O2 -fomit-frame-pointer).
uint32_t std(uint32_t x) {
return x * 40 / 47;
}
uint32_t fast(uint32_t x) {
static const uint64_t fact = ((((uint64_t)40) << 32) + 46) / 47;
return (x * fact) >> 32;
}
.def _std; .scl 2; .type 32; .endef
_std:
movl 4(%esp), %edx
movl $-1370734243, %eax
leal (%edx,%edx,4), %edx
sall $3, %edx
mull %edx
shrl $5, %edx
movl %edx, %eax
ret
.p2align 4,,15
.globl _fast
.def _fast; .scl 2; .type 32; .endef
_fast:
movl $-639675980, %eax
mull 4(%esp)
movl %edx, %eax
xorl %edx, %edx
ret
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=20283
