------- Comment #43 from pthaugen at gcc dot gnu dot org 2008-04-30 18:49
-------
Created an attachment (id=15553)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=15553&action=view)
Testcase
I tried a mainline with the latest patch. While we no longer have problems
with the prior testcases, there is no improvement for leslie3d on ppc64. I can
still double the performance of the benchmark by specifying --param
max-aliased-vops=10000.
Including a new trimmed down testcase from the benchmark where I'm still seeing
poor code when max-aliased-vops is not increased, compiled with 'gfortran -m32
-O2'.
Refer to the first nested loop in procedure FLUXK():
DO I = I1, I2
QS(I) = WAV(I,J,K) * ZAREA
END DO
Base:
.L150:
lwz 0,24(18) # <variable>.stride, <variable>.stride
lwz 9,36(18) # <variable>.stride, <variable>.stride
lwz 11,12(18) # <variable>.stride, <variable>.stride
lwz 10,4(18) # wav.offset, wav.offset
mullw 0,17,0 # tmp660, ivtmp.602, <variable>.stride
lwz 8,0(18) # wav.data, wav.data
mullw 9,30,9 # tmp666, ivtmp.590, <variable>.stride
mullw 11,6,11 # tmp670, i, <variable>.stride
add 0,0,9 # tmp672, tmp660, tmp666
addi 6,6,1 # i, i,
add 0,0,11 # tmp673, tmp672, tmp670
add 0,0,10 # tmp674, tmp673, wav.offset
slwi 0,0,3 # tmp676, tmp674,
lfdx 0,8,0 #, tmp678
fmul 0,0,8 # tmp679, tmp678, zarea.64
stfdx 0,15,7 #* ivtmp.589, tmp679
addi 7,7,8 # ivtmp.589, ivtmp.589,
bdnz .L150 #
With --param max-aliased-vops=1000:
.L150:
lfd 0,0(11) #* ivtmp.599, tmp739
add 11,11,30 # ivtmp.599, ivtmp.599, D.2783
fmul 0,0,8 # tmp740, tmp739, zarea.64
stfdx 0,22,9 #* ivtmp.602, tmp740
addi 9,9,8 # ivtmp.602, ivtmp.602,
bdnz .L150 #
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32921
