On Thu, Oct 27, 2016 at 2:58 PM, Richard Biener <richard.guent...@gmail.com> wrote: > On Tue, Oct 25, 2016 at 1:21 PM, Bin Cheng <bin.ch...@arm.com> wrote: >> Hi, >> Second patch optimizing (cond (cmp (convert (x), c1), x, c2)) into (minmax >> (x, c)). As commented in patch, this is done if: >> >> + 1) Comparison's operands are promoted from smaller type. >> + 2) Const c1 equals to c2 after canonicalizing comparison. >> + 3) Comparison has tree code LT, LE, GT or GE. >> + This specific pattern is needed when (cmp (convert x) c) may not >> + be simplified by comparison patterns because of multiple uses of >> + x. It also makes sense here because simplifying across multiple >> + referred var is always benefitial for complicated cases. >> >> It also adds call to fold_stmt in tree-if-conv.c so that generated cond_expr >> statement has its chance to be simplified. > > So this sounds like the fold-const.c code. Why only for GIMPLE btw? > Thus, can you make sure you covered > all cases fold-const.c did and remove the corresponding code there? Hmm, no, this patch only adds simplification which is not done by GCC at the moment. I didn't look into existing simplification from fold-const.c when writing this one, however, I will look into it see if it can be moved to match.pd, either along with this pattern or as a separated ones.
Thanks, bin > > Thanks, > Richard. > >> Bootstrap and test on x86_64 and AArch64. It introduces below failure on >> both x86_64 and AArch64: >> FAIL: gcc.dg/vect/slp-cond-3.c >> I believe it reveals defect in vect-slp. In call to fold_stmt in ifcvt, >> canonicalization transforms _145 = _95 <= _96 ? _149 : _147 into _145 = _95 >> > _96 ? _147 : _149. As a result, this stmt has different code to the first >> one of SLP instance. IMO, SLP should be improved to handle operands >> swapping, apparently, current support is not OK. >> >> It also introduces more failures on AArch64(probably other targets) as below: >> FAIL: gcc.dg/vect/pr65947-1.c -flto -ffat-lto-objects scan-tree-dump-times >> vect "LOOP VECTORIZED" 2 >> FAIL: gcc.dg/vect/pr65947-1.c -flto -ffat-lto-objects scan-tree-dump-times >> vect "condition expression based on integer induction." 4 >> FAIL: gcc.dg/vect/pr65947-1.c scan-tree-dump-times vect "LOOP VECTORIZED" 2 >> FAIL: gcc.dg/vect/pr65947-1.c scan-tree-dump-times vect "condition >> expression based on integer induction." 4 >> FAIL: gcc.dg/vect/pr65947-13.c -flto -ffat-lto-objects scan-tree-dump-times >> vect "LOOP VECTORIZED" 2 >> FAIL: gcc.dg/vect/pr65947-13.c scan-tree-dump-times vect "LOOP VECTORIZED" 2 >> FAIL: gcc.dg/vect/pr65947-4.c -flto -ffat-lto-objects scan-tree-dump-times >> vect "LOOP VECTORIZED" 2 >> FAIL: gcc.dg/vect/pr65947-4.c -flto -ffat-lto-objects scan-tree-dump-times >> vect "condition expression based on integer induction." 4 >> FAIL: gcc.dg/vect/pr65947-4.c scan-tree-dump-times vect "LOOP VECTORIZED" 2 >> FAIL: gcc.dg/vect/pr65947-4.c scan-tree-dump-times vect "condition >> expression based on integer induction." 4 >> FAIL: gcc.dg/vect/pr77503.c -flto -ffat-lto-objects scan-tree-dump vect >> "vectorized 1 loops" >> FAIL: gcc.dg/vect/pr77503.c scan-tree-dump vect "vectorized 1 loops" >> FAIL: gcc.dg/vect/vect-pr69848.c -flto -ffat-lto-objects scan-tree-dump >> vect "vectorized 1 loops" >> FAIL: gcc.dg/vect/vect-pr69848.c scan-tree-dump vect "vectorized 1 loops" >> Again, these failures reveal a defect in vectorizer that operand swapping is >> not supported for COND_REDUCTION. >> >> I will send another two patches independent to this patch set resolving >> these failures. Is this OK? >> >> Thanks, >> bin >> >> 2016-10-21 Bin Cheng <bin.ch...@arm.com> >> >> * tree-if-conv.c (ifcvt_follow_ssa_use_edges): New func. >> (predicate_scalar_phi): Call fold_stmt using the new valueize func. >> * match.pd ((cond (cmp (convert (x), c1), x, c2)) -> (minmax (x, >> c))): >> New pattern. >> >> gcc/testsuite/ChangeLog >> 2016-10-21 Bin Cheng <bin.ch...@arm.com> >> >> * gcc.dg/fold-condcmpconv-1.c: New test. >> * gcc.dg/fold-condcmpconv-2.c: New test.
