| Message ID | 598c816d-e5eb-6c1f-fec4-f2c7ab9e2898@arm.com |
|---|---|
| State | New |
| Headers | show |
| Series | [vect] PR92347: Don't overwrite safelen for epilogues | expand |
On Mon, 11 Nov 2019, Andre Vieira (lists) wrote: > Hi, > > After during my investigation into the first testcase in PR92347 I noticed > that the ICE was arising because the vectorizer was under the assumption it > knew how to vectorize the zp call for both the main and epilogues. By the > time it got to transform the epilogue vectorizable_simd_clone_call would fail, > leading to the eventual ICE. The reason was because the main loop's safelen > was set to MAX_INT at time of analysis, but 0 for the epilogue at the time of > its transformation. This was because we were copying the loop's safelen to > the epilogue after vectorizing the main loop and we set safelen to 0 when we > are done vectorizing. This resulted in the epilogue always having safelen set > to 0 at transformation time. > > I believe its safe to assume that if the main loop can assume that it is safe > to vectorize between 0 and safelen then so is the epilogue in this context. > > Bootstrapped x86_64 and aarch64 and ran libgomp. > > Is this OK for trunk? (If approved can someone commit on my behalf as I'll be > away for two weeks?) OK (and committed). Richard. > Cheers, > Andre > > gcc/ChangeLog: > 2019-11-11 Andre Vieira <andre.simoesdiasvieira@arm.com> > > * tree-vect-loop.c (vect_transform_loop): Don't overwrite epilogues > safelen with 0. > > gcc/testsuite/ChangeLog: > 2019-11-11 Andre Vieira <andre.simoesdiasvieira@arm.com> > > * gcc.dg/vect/pr92347.c: New test. > >
diff --git a/gcc/testsuite/gcc.dg/vect/pr92347.c b/gcc/testsuite/gcc.dg/vect/pr92347.c new file mode 100644 index 0000000000000000000000000000000000000000..f7e6b490c837f209e9a574d6f10e1e2e34e4729e --- /dev/null +++ b/gcc/testsuite/gcc.dg/vect/pr92347.c @@ -0,0 +1,18 @@ +/* Copied from PR 92347. */ +/* { dg-do compile } */ +/* { dg-options "-O1 -fopenmp" } */ +/* { dg-additional-options "-mavx" { target { i?86-*-* x86_64-*-* } } } */ + +int by; + +#pragma omp declare simd +int +zp (int); + +void +qh (int oh) +{ +#pragma omp simd + for (by = 0; by < oh; ++by) + by = zp (by); +} diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c index 78fe6290ee6bf069f7cfaabba466f119e41f39ac..b600d3157457c3180d0456c4f66cbc57012e3c71 100644 --- a/gcc/tree-vect-loop.c +++ b/gcc/tree-vect-loop.c @@ -8587,7 +8587,6 @@ vect_transform_loop (loop_vec_info loop_vinfo) epilogue->simduid = loop->simduid; epilogue->force_vectorize = loop->force_vectorize; - epilogue->safelen = loop->safelen; epilogue->dont_vectorize = false; }
Hi, After during my investigation into the first testcase in PR92347 I noticed that the ICE was arising because the vectorizer was under the assumption it knew how to vectorize the zp call for both the main and epilogues. By the time it got to transform the epilogue vectorizable_simd_clone_call would fail, leading to the eventual ICE. The reason was because the main loop's safelen was set to MAX_INT at time of analysis, but 0 for the epilogue at the time of its transformation. This was because we were copying the loop's safelen to the epilogue after vectorizing the main loop and we set safelen to 0 when we are done vectorizing. This resulted in the epilogue always having safelen set to 0 at transformation time. I believe its safe to assume that if the main loop can assume that it is safe to vectorize between 0 and safelen then so is the epilogue in this context. Bootstrapped x86_64 and aarch64 and ran libgomp. Is this OK for trunk? (If approved can someone commit on my behalf as I'll be away for two weeks?) Cheers, Andre gcc/ChangeLog: 2019-11-11 Andre Vieira <andre.simoesdiasvieira@arm.com> * tree-vect-loop.c (vect_transform_loop): Don't overwrite epilogues safelen with 0. gcc/testsuite/ChangeLog: 2019-11-11 Andre Vieira <andre.simoesdiasvieira@arm.com> * gcc.dg/vect/pr92347.c: New test.