Hi! Without this patch, ifcvt extends lifetime of %eax hard register, which causes reload/LRA ICE later on. Combiner and other passes try hard not to do that, even ifcvt has code for it if x is a hard register a few lines below it, but in this case the hard register is SET_SRC (set_b).
With this patch we just use the pseudo (x) which has been initialized from the hard register before the conditional. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2013-03-05 Jakub Jelinek <ja...@redhat.com> PR rtl-optimization/56484 * ifcvt.c (noce_process_if_block): Before reload if else_bb is NULL, avoid extending lifetimes of hard registers in likely to spilled or small register classes. --- gcc/ifcvt.c.jj 2013-01-11 09:02:48.000000000 +0100 +++ gcc/ifcvt.c 2013-03-05 12:36:19.217251997 +0100 @@ -2491,6 +2491,15 @@ noce_process_if_block (struct noce_if_in || ! noce_operand_ok (SET_SRC (set_b)) || reg_overlap_mentioned_p (x, SET_SRC (set_b)) || modified_between_p (SET_SRC (set_b), insn_b, jump) + /* Avoid extending the lifetime of hard registers on small + register class machines before reload. */ + || (!reload_completed + && REG_P (SET_SRC (set_b)) + && HARD_REGISTER_P (SET_SRC (set_b)) + && (targetm.class_likely_spilled_p + (REGNO_REG_CLASS (REGNO (SET_SRC (set_b)))) + || targetm.small_register_classes_for_mode_p + (GET_MODE (SET_SRC (set_b))))) /* Likewise with X. In particular this can happen when noce_get_condition looks farther back in the instruction stream than one might expect. */ --- gcc/testsuite/gcc.c-torture/compile/pr56484.c.jj 2013-03-05 12:42:24.972220034 +0100 +++ gcc/testsuite/gcc.c-torture/compile/pr56484.c 2013-03-05 12:41:59.000000000 +0100 @@ -0,0 +1,17 @@ +/* PR rtl-optimization/56484 */ + +unsigned char b[4096]; +int bar (void); + +int +foo (void) +{ + int a = 0; + while (bar ()) + { + int c = bar (); + a = a < 0 ? a : c; + __builtin_memset (b, 0, sizeof b); + } + return a; +} Jakub