On 07/27/2013 06:51 PM, Bill Schmidt wrote:
PR57993 reports a scenario where a conditional candidate is incorrectly replaced when its putative "hidden basis" (the basis hidden by one or more PHI nodes) does not dominate all of the PHI nodes on which the candidate depends.This indicates that the hidden basis was incorrectly identified as a useful basis for the candidate. There are two ways we can fix this. One would be to add more complexity to the code that determines the hidden basis. Currently that code does not recursively follow the PHI definitions backwards to ensure that the basis dominates all the PHIs. Adding code to do this would ensure we didn't identify an incorrect basis in the first place, but would duplicate quite a bit of existing logic in the later analysis phase, and increase compile time. What I've done here is to instead delay detection of the problem until the analysis phase. If we detect that we chose an invalid basis, we assign an "infinite" cost to the replacement so that we won't pursue it further. This requires that we know which basic block contains the basis statement. The basis statement may have itself been replaced by another statement earlier, so we need to keep the candidate table up to date with respect to such replacements. Bootstrapped and tested on powerpc64-linux-gnu with no new regressions. Ok for trunk? Thanks, Bill gcc: 2013-07-27 Bill Schmidt <wschm...@vnet.linux.ibm.com> * gimple-ssa-strength-reduction.c (replace_mult_candidate): Record replaced statement in the candidate table. (phi_add_costs): Return infinite cost when the hidden basis does not dominate all phis on which the candidate is dependent. (replace_one_candidate): Record replaced statement in the candidate table. gcc/testsuite: 2013-07-27 Bill Schmidt <wschm...@vnet.linux.ibm.com> * gcc.dg/torture/pr57993.c: New test.
Looks good. Please install! Thanks, Jeff
