Hi Kewen, 在 2022/11/22 11:11, Kewen.Lin 写道: > Maybe we can adjust prepare_cmp_insn to fail if the constructed cbranchcc4 > pattern doesn't satisfy the predicate of operand 0 rather than to assert. > It's something like: > > if (!insn_operand_matches (icode, 0, test)) > goto fail; > > or only assign and return if insn_operand_matches (icode, 0, test). > > The code makes the assumption that all this kind of cbranchcc4 patterns > should match what target defines for cbranchcc4 optab, but unfortunately > it's not sure for our port and I don't see how it should be.
Thanks for your comments. I just drafted a patch to let it go to "fail" when predicate of operand 0 is not satisfied. It works and passed bootstrap on ppc64le. But my concern is prepare_cmp_insn is a generic function and is used to create a cmp rtx. It is not only called by emit_conditional* (finally called by ifcvt) but other functions (even new functions). If we change the logical in prepare_cmp_insn, we may lost some potential optimization. After all, the branch_2insn is a valid insn. I think the essential of the problem is we want to exclude those comparisons (from cbranchcc4 used in ifcvt) which need two CC bits. So, we can change the logical of ifcvt - add an additional check with predicate of operand 0 when checking the have_cbranchcc4 flag in ifcvt. What's your opinion? Thanks Gui Haochen
