Hi, With PowerPC GCC tool chain v4.6.1 (for e500mc), i have a test case which shows regression at -O3 with the patch [IRA-based register pressure calculation for RTL loop invariant motion] listed below:
http://gcc.gnu.org/ml/gcc-patches/2009-09/msg01889.html Reduced Test Case: /***************************************************************************************/ extern int a; int asdf (int x); char buf[32]; int test (void) { int i, x; for (i = 0 ; i < a; i++) { asm ("stfd %0,0(%1)" : : "d" (8.9731144674927890e20), "b" (buf)); x += asdf(i); } return x; /****************************************************************************************/ Code Generated: test: .... ble- 7,.L2 lis 26,.LC0@ha lis 30,buf@ha lis 27,.LC1@ha li 31,0 la 26,.LC0@l(26) la 30,buf@l(30) la 27,.LC1@l(27) la 28,a@l(28) .p2align 4,,15 .L3: lfd 0,0(26) ----------------------(A) #APP # 9 "test.c" 1 stfd 0,0(30) # 0 "" 2 #NO_APP lis 30,buf@ha ---------------(B) lfd 0,0(27) ---------------(C) la 30,buf@l(30) ----------------(D) #APP # 10 "test.c" 1 stfd 0,8(30) # 0 "" 2 #NO_APP mr 3,31 bl asdf lwz 0,0(28) addi 31,31,1 add 29,29,3 cmpw 7,0,31 bgt+ 7,.L3 .L2: lwz 0,36(1) ... Here we can see that the load of the constant values (A, B, C, D) are inside the loop. The ‘loop invariant code motion’ pass identifies that these instructions (const value load instructions) can be moved outside the loop. But with flag ‘-fira-loop-pressure’ ON [IRA-based register pressure calculation for RTL loop invariant motion], the heuristic determines that this is not profitable. And with ‘-O3’ and above, we cannot switch OFF the flag ‘-fira-loop-pressure’, because it has been explicitly overridden in the PPC backend target files. 1) Is it possible to update the heuristic to allow this optimization. 2) If not, the user should have control over the '-fira-loop-pressure' flag when passed explicitly. File: rs6000.c if (optimize >= 3 && global_init_p && !global_options_set.x_flag_ira_loop_pressure) flag_ira_loop_pressure = 1; Regards, Rohit
