This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[BENCHMARK]-mfpmath=sse should disable x387 intrinsics
- From: Uros Bizjak <uros at kss-loka dot si>
- To: gcc-patches at gcc dot gnu dot org, Roger Sayle <roger at eyesopen dot com>
- Date: Thu, 25 Nov 2004 10:18:27 +0100
- Subject: [BENCHMARK]-mfpmath=sse should disable x387 intrinsics
Hello Roger!
I have done a couple of whetstone benchmarks with your patch to disable
x387 intrinsics on pentium4, 3.2 GHz. As it can be seen from attached
results, the best results can be obtained with the combination of sse
and i387 math. This combination is the fastest one, acheiving more than
8% gain, comparing to the default of i387 only.
-mfpmath=sse is the worst choice in case of pentium4. The result is
lower by 18%, comparing to the default. That is, -mfpmath=sse,387 is
faster by 28%, comparing to -mfpmath=sse on pentium4.
For example, the code for
double test(double a, double b) {
return sin(a*b);
}
with -mfpmath=sse looks like:
pushl %ebp
movl %esp, %ebp
movsd 16(%ebp), %xmm0
mulsd 8(%ebp), %xmm0
movsd %xmm0, 8(%ebp)
popl %ebp
jmp sin
However, libm's sin still does (that is also the case with x86_64's
"long double" functions):
fldl 4(%esp)
fsin
...
So, the parameter still goes thru stack, as in case of using i387
intrinsic. I guess there is no libm for i386 using sse-only code.
However, the benefits of -ffast-math would be lost here, because libm
provides full version of sin code, including all checking of infinity, etc..
With "#include <math.h>", this code is produced, even with -mfpmath=sse:
test:
pushl %ebp
movl %esp, %ebp
subl $8, %esp
movsd 16(%ebp), %xmm0
mulsd 8(%ebp), %xmm0
movsd %xmm0, -8(%ebp)
fldl -8(%ebp)
#APP
fsin
#NO_APP
leave
ret
Another problem is ix86 calling convention, which requires FP return
value in FP stack. Consider this code:
double test(double a, double b) {
return a * sin(b);
}
With -O2 -march=pentium4 -ffast-math -mfpmath=sse the result is:
test:
pushl %ebp
movl %esp, %ebp
subl $24, %esp
movsd 16(%ebp), %xmm0
movsd %xmm0, (%esp)
call sin
fstpl -8(%ebp)
movsd -8(%ebp), %xmm0
mulsd 8(%ebp), %xmm0
movsd %xmm0, 8(%ebp)
fldl 8(%ebp)
leave
ret
and with -mfpmath=sse,387:
pushl %ebp
movl %esp, %ebp
fldl 16(%ebp)
fsin
fmull 8(%ebp)
popl %ebp
ret
However, I think it is a good idea to have -mfpmath=sse disable i387
intrinsics, as it means more choices for user. If someone wants to use
i387 intrinsics (which is indeed 387-only feature :), -mfpmath=sse,387
should be specified.
Uros.
gcc version 4.0.0 20041125 (experimental)
gcc -O2 -march=pentium4 -ffast-math -mfpmath=sse -D __NO_MATH_INLINES -lm whetss.c
##########################################
Single Precision C/C++ Whetstone Benchmark
Calibrate
0.00 Seconds 1 Passes (x 100)
0.04 Seconds 5 Passes (x 100)
0.20 Seconds 25 Passes (x 100)
0.97 Seconds 125 Passes (x 100)
5.16 Seconds 625 Passes (x 100)
Use 12121 passes (x 100)
Single Precision C/C++ Whetstone Benchmark
Loop content Result MFLOPS MOPS Seconds
N1 floating point -1.12475013732910156 748.306 0.311
N2 floating point -1.12274742126464844 598.920 2.720
N3 if then else 1.00000000000000000 1348.950 0.930
N4 fixed point 12.00000000000000000 23863.242 0.160
N5 sin,cos etc. 0.49911010265350342 35.673 28.270
N6 floating point 0.99999982118606567 230.457 28.370
N7 assignments 3.00000000000000000 380.946 5.880
N8 exp,sqrt etc. 0.75110864639282227 12.916 34.910
MWIPS 1193.588 101.551
--
gcc -O2 -march=pentium4 -ffast-math -mfpmath=387 -D __NO_MATH_INLINES -lm whetss.c
(-mfpmath=387 is the default on pentium4)
##########################################
Single Precision C/C++ Whetstone Benchmark
Calibrate
0.00 Seconds 1 Passes (x 100)
0.03 Seconds 5 Passes (x 100)
0.16 Seconds 25 Passes (x 100)
0.82 Seconds 125 Passes (x 100)
4.35 Seconds 625 Passes (x 100)
Use 14367 passes (x 100)
Single Precision C/C++ Whetstone Benchmark
Loop content Result MFLOPS MOPS Seconds
N1 floating point -1.12441420555114746 590.678 0.467
N2 floating point -1.12241148948669434 526.137 3.670
N3 if then else 1.00000000000000000 1315.915 1.130
N4 fixed point 12.00000000000000000 23818.940 0.190
N5 sin,cos etc. 0.49907428026199341 54.432 21.960
N6 floating point 0.99999988079071045 189.060 40.990
N7 assignments 3.00000000000000000 381.469 6.960
N8 exp,sqrt etc. 0.75095528364181519 20.267 26.370
MWIPS 1412.171 101.737
--
gcc -O2 -march=pentium4 -ffast-math -mfpmath=sse,387 -D __NO_MATH_INLINES -lm whetss.c
##########################################
Single Precision C/C++ Whetstone Benchmark
Calibrate
0.00 Seconds 1 Passes (x 100)
0.03 Seconds 5 Passes (x 100)
0.15 Seconds 25 Passes (x 100)
0.75 Seconds 125 Passes (x 100)
4.01 Seconds 625 Passes (x 100)
Use 15601 passes (x 100)
Single Precision C/C++ Whetstone Benchmark
Loop content Result MFLOPS MOPS Seconds
N1 floating point -1.12475013732910156 745.122 0.402
N2 floating point -1.12274742126464844 599.078 3.500
N3 if then else 1.00000000000000000 1356.894 1.190
N4 fixed point 12.00000000000000000 23401.496 0.210
N5 sin,cos etc. 0.49907428026199341 54.332 23.890
N6 floating point 0.99999982118606567 230.553 36.500
N7 assignments 3.00000000000000000 380.854 7.570
N8 exp,sqrt etc. 0.75095528364181519 20.229 28.690
MWIPS 1530.230 101.952