On Thu, Nov 24, 2016 at 4:22 PM, Bin Cheng <bin.ch...@arm.com> wrote: > Hi, > This patch fixes two issues in newly introduced pattern: (cond (cmp > (convert1? @1) @3) (convert2? @1) @2). > For PR78507, we need to check if from_type is INTEGRAL_TYPE_P explicitly, > this patch adds check for that. > Note we don't check c1_type/c2_type for that because it's guarded by > INTEGER_CST@. > For PR78510, the ICE is covered by revision 242831, but underlying issue is > when the block of conditions > is not satisfied, the last case simplification is applied anyway since code > is initialized to cmp at the > first place. This patch fixes that by setting code to proper value only when > transformation is valid, it > also incorporates Marc's suggestion by using cmp directly. Two tests are > added too. > Bootstrap and test on x86_64 and AArch64 ongoing, is it OK if no failures?
Ok. Btw, - (cond (cmp@0 (convert1? @1) ... ... - enum tree_code code = TREE_CODE (@0)... shows another misconception (I didn't notice...) in that TREE_CODE of a captured expression results in the expression code. On GENERIC that works but on GIMPLE @0 will be a SSA_NAME and thus 'code' was SSA_NAME ... Thanks, Richard. > Thanks, > bin > > 2016-11-24 Bin Cheng <bin.ch...@arm.com> > > PR middle-end/78507 > PR middle-end/78510 > * match.pd ((cond (cmp (convert1? @1) @3) (convert2? @1) @2)): Use > cmp directly, rather than cmp_code. Initialize code to ERROR_MARK > and set it to result code if transformation is valid. Use code EQ > directly in last simplification case. > > gcc/testsuite/ChangeLog > 2016-11-24 Bin Cheng <bin.ch...@arm.com> > > PR middle-end/78507 > PR middle-end/78510 > * g++.dg/torture/pr78507.C: New test. > * gcc.dg/torture/pr78510.c: New test.
