http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54272
Oleg Endo <olegendo at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Status|UNCONFIRMED |ASSIGNED Last reconfirmed| |2012-08-16 Ever Confirmed|0 |1 --- Comment #5 from Oleg Endo <olegendo at gcc dot gnu.org> 2012-08-16 20:34:53 UTC --- Regarding saturating arithmetic, I took the function FIXED_SSADD from libgcc/fixed-bit.c and substituted FIXED_C_TYPE with int. int test_01 (int a, int b) { int c; int x, y, z; x = a; y = b; z = x + y; if ((((x ^ y) >> 31) & 1) == 0) { if (((z ^ x) >> 31) & 1) { z = 1; z = z << 31; if (x >= 0) z--; } } c = z; return c; } compiled with -O2 -m4 -ml: div0s r4,r5 mov r4,r0 bt/s .L4 add r5,r0 div0s r0,r4 bt/s .L6 cmp/pz r4 .L4: rts nop .align 1 .L6: mov.l .L7,r0 movt r1 rts sub r1,r0 .L8: .align 2 .L7: .long -2147483648 In this case combine can successfully match the div0s sign comparison patterns that I added in PR 52933. However, after matching those, it does not try to combine the patterns any further (probably because of the conditional branches). I guess this would be a job for an optimization at the GIMPLE level.