This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [patch RFC] SH: Use FRAME_GROWS_DOWNWARD


Kaz Kojima wrote:



CSiBE 2.1.0 Runtime result (sec):

		original compiler		patched compiler
		no-omit-fp	omit-fp		no-omit-fp	omit-fp	
bzip2.c 	48.3267		47.03		44.1133		43.6933	
bzip2.d 	10.1933		10.22		10.2633		10.24	
bzip2recover 	5.01667		4.93		5.01667		4.94	
vam.fact 	1.64667		1.67667		1.65333		1.66333	
vam.fib 	1.74667		1.76333		1.76333		1.75667	
vam.test2 	0.24		0.233333	0.233333	0.23	
flex		15.1033		15.0367		15.08		15.06	
jikespg 	1.67667		1.64333		(*)		(*)
cjpeg2		3.63333		3.59		3.64		3.48	
cjpeg0		0.583333	0.58		0.586667	0.56	
cjpeg1		3.67		3.58667		3.61333		3.55667	
djpeg0		0.386667	0.383333	0.39		0.386667
djpeg1		2.81333		2.58		2.63333		2.42667	
djpeg2		2.77		2.63333		2.67333		2.48667	
jpegtran0 	0.313333	0.296667	0.303333	0.276667
jpegtran1 	2.15667		2.11333		2.08333		2.04	
jpegtran2 	5.37		5.22		4.77333		4.76333	
png2pnm0 	1.12667		1.08333		1.09		1.06333	
png2pnm1 	7.98667		7.62		8.02667		7.72667	
pnm2png0 	7.18		7.14333		7.16667		7.14667	
pnm2png1 	40.42		40.2033		40.2433		40.12	
dehex0		1.77667		1.75		1.77667		1.78	
dehex1		11.64		11.6633		11.66		11.54	
enhex0		5.52667		5.53333		5.53		5.54333	
enhex1		36.31		36.18		36.3533		36.1767	
minigzip 	6.34667		6.19		6.16333		6.06667	
minigzip0 	55.0333		53.3133		53.1767		51.97	

(*) runtime error with
*** glibc detected *** double free or corruption (out): 0x004a4600 ***

Regards,
kaz


That sounds promising. If I replace the missing numbers with 1.68, awk calculates about 1.8%
improvement on the geometric mean:
bash-2.05b$ cat tmp|awk 'BEGIN {FS=" ";prod = 1 } {prod*=$2;num++} END {print exp(log(prod)/num)}'
3.96185
bash-2.05b$ cat tmp|awk 'BEGIN {FS=" ";prod = 1 } {prod*=$3;num++} END {print exp(log(prod)/num)}'
3.88543
bash-2.05b$ cat tmp|awk 'BEGIN {FS=" ";prod = 1 } {prod*=$4;num++} END {print exp(log(prod)/num)}'
3.8922
bash-2.05b$ cat tmp|awk 'BEGIN {FS=" ";prod = 1 } {prod*=$5;num++} END {print exp(log(prod)/num)}'
3.81453


However, the jikespg failure needs to be investigated. Is this a latent bug in the benchmark,
or is it a compiler bug?



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]