>>Sorry this is probably a very basic question, as I don't have a strong knowledge in the field, but here is my question.
>>
>>I use Shark to optimize some of my code (so far, ~3-4 fold increase, after ~ half a dozen hot spots were fixed). However, I found in many places the comment 'Invariant load', typically when looping over an array and using the loop variable as the index (this particular code is not worth any optimization, it is just the simplest example I could find in my code):
>>
>>/* int i ; double *dsdrate,*dsdr */
>>for (i=0;i<countSpecies;i++) /* Invariant load, Early exit */
>>dsdrate[i]=dsdr[i]; /* Invariant load, Unroll, Unaligned loop start */
>>
>>So what exactly is causing these 'Invariant load' warnings from Shark? It seems it disappears when I increment a pointer inside the loop instead, like so (I did not test that particular code in Shark, though):
>>double *p1=dsdrate;
>>double *p2=dsdr;
>>for (i=0;i<countSpecies;i++) {
>>*p1=*p2;
>>p1++;
>>p2++;
>>}
>>
>>There are maybe a couple of places where I could do some optimization on those 'Invariant load' warnings. Thanks for your input!
>
>This frequently happens when the array in question is of global or external scope. The compiler might be reloading the pointer to the array each time through the loop, just in case someone else changed it, perhaps in another thread.
>
>We'd have to see the disassembly for the loop showing what was loaded and perhaps also the exact source showing how dsdrate and dsdr are declared to be sure.
>
>Ian
Thanks for the quick and helpful response!
I wanted to mention it and then forgot, but yes, those variables are of external scope, as they are instance variables of an objc object. They are thus members of a struct, which is passed by ref to the function (as objc objects are pointers to struct). So yes, this is probably the problem. I never thought of it in this particular configuration... Is the fix to use a local variable instead? Something like:
double *p1=dsdrate;
double *p2=dsdr;
for (i=0;i<countSpecies;i++)
p1[i]=p2[i];
Just in case, here is the assembly I should have added in the initial post (line 0x264d8 is not part of the loop per se; the loop is at the very beginning of the function; this is the assembly just for the loop):
0x264d0 lwz r2,100(r3) 3:1
0x264d4 li r6,0 1:1
0x264d8 mr r12,r5 1:1
0x264dc cmplw r6,r2 1:1
0x264e0 bge $+36 <-[BCKCalculatorGeneric getdsdratedt:atTime:values:] + 52> 1:1
0x264e4 slwi r0,r6,3 1:1
0x264e8 lwz r5,124(r3) 3:1 !Invariant load
0x264ec addi r6,r6,1 1:1
0x264f0 lfdx f1,r8,r0 4:1
0x264f4 stfdx f1,r5,r0 3:3
0x264f8 lwz r4,100(r3) 3:1 !Invariant load
0x264fc cmplw r6,r4 1:1
0x26500 blt $-28 <-[BCKCalculatorGeneric getdsdratedt:atTime:values:] + 20> 1:1 Loop end[1]
The assembly for the line INSIDE the loop is:
0x264e4 slwi r0,r6,3 1:1
0x264e8 lwz r5,124(r3) 3:1
0x264f0 lfdx f1,r8,r0 4:1
0x264f4 stfdx f1,r5,r0 3:3
thanks!
--
Charles Parnot
email@hidden
Help science go fast forward:
http://cmgm.stanford.edu/~cparnot/xgrid-stanford/
Room B157 in Beckman Center
279, Campus Drive
Stanford University
Stanford, CA 94305 (USA)
Tel +1 650 725 7754
Fax +1 650 725 8021
_______________________________________________
Do not post admin requests to the list. They will be ignored.
PerfOptimization-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
http://lists.apple.com/mailman/options/perfoptimization-dev/email@hidden
This email sent to email@hidden
References:
>(no subject) (From: Charles PARNOT <email@hidden>)
