I found this in ppc32 3.3.3 (Debian), it also appears in 3.4.0 20040215. If the loop count is large enough, it wont use the count register in the second loop. eg: #if 1 #define LOOPS 100000 #else #define LOOPS 10 #endif int main() { unsigned long i; i = LOOPS; while (i--) { asm volatile("":::"memory"); } i = LOOPS; while (i--) { asm volatile("":::"memory"); } } produces the following code: lis 0,0x1 ori 0,0,34464 mtctr 0 .L13: bdnz .L13 lis 0,0x1 ori 0,0,34464 .L12: addic. 0,0,-1 bne+ 0,.L12 blr If you reduce COUNT it will use the count register for both loops as expected
I can confirm this on the mainline and the tree-ssa, I have not looked to see if the doloop rewrite fixes it as I do not have the lno branch on the computer which I have access to right now.
I thought this looked familiar. There's an old patch of mine at http://gcc.gnu.org/ml/gcc-patches/2002-09/msg01194.html
Fixed on the mainline, most likely by the rewrite of doloop optimization.