Hi,
While trying to figure out why the PowerMac mesh scsi driver will work
when compiled with 4.1.2 but not with 4.3.5 I noticed some strange
behavior with literal strings.
#include <stdio.h>
const char *getMsg(unsigned int u)
{
const char *p;
switch(u)
{
case 0:
p="This is message 0"; break;
case 1:
p="This is message 1"; break;
case 2:
p="This is message 2"; break;
case 3:
p="This is message 3"; break;
case 4:
p="This is message 4"; break;
case 5:
p="This is message 5"; break;
default:
p=NULL;
}
return p;
}
This function when compiled with 3.4.6 produces pretty good code.
4.3.5 produces lesser good code. Can someone suggest what I search the
bug database for to see if this has been reported and what the status
is? Is it PowerPC specific?
In both the 4.2.4 and 4.3.5 versions of mesh, the mesh_start() routine
has a lengthy list of register assignments near the start of the
routine:
.L241:
lis 9,.LANCHOR0@ha # tmp459,
lis 11,.LC47@ha # tmp461,
la 23,.LANCHOR0@l(9) # tmp460,, tmp459
lis 9,.LC48@ha # tmp463,
la 19,.LC47@l(11) # tmp462,, tmp461
la 18,.LC48@l(9) # tmp464,, tmp463
lis 11,.LC16@ha # tmp465,
lis 9,.LC50@ha # tmp467,
la 17,.LC16@l(11) # tmp466,, tmp465
la 22,.LC50@l(9) # tmp468,, tmp467
lis 11,.LC51@ha # tmp469,
lis 9,.LC57@ha # tmp471,
la 21,.LC51@l(11) # tmp470,, tmp469
la 20,.LC57@l(9) # tmp472,, tmp471
lis 11,.LC58@ha # tmp473,
lis 9,.LC59@ha # tmp475,
la 16,.LC58@l(11) # tmp474,, tmp473
la 15,.LC59@l(9) # tmp476,, tmp475
lis 11,.LC60@ha # tmp477,
lis 9,.LC53@ha # tmp480,
la 14,.LC60@l(11) # tmp478,, tmp477
la 24,.LC53@l(9) # tmp482,, tmp480
Later, when one of these is used you'll see:
mr 3,19 #, tmp462
mr 4,31 #, cmd
mr 6,28 #, id
bl printk #
for .LC47, which is "<4>mesh_start: %p ser=%lu tgt=%d cmd="
These all end up in:
.section .rodata.str1.4,"aMS",@progbits,1
and will have addresses close to each other, right? I don't understand
why they can't do something like:
lis 15,.LANCHOR0@ha
la 15,.LANCHOR0@l(15)
and then do something like:
la 3,<some offset>(15)
when each string is used. What am I missing?
The above register loads also result in:
stmw 14,8(1) #,
and
lmw 14,8(1) #,
for 4.3.5. A smaller set is saved in 4.1.2
stmw 25,20(1) #,
and
lmw 25,20(1) #,
Won't this lead to unneeded memory traffic and/or cache churning?
kevin
P.S.: For all those who send email from their Blackberry's and iphones:
Sent from my 33 MHz 386 using Firefox 2!
Attachment:
gcctest1-3.4.6.s
Description: Binary data
Attachment:
gcctest1-4.3.5.s
Description: Binary data
