https://gcc.gnu.org/g:9b840b6380426a8e760076f6541c6aa74c7a6e5c
commit r16-9908-g9b840b6380426a8e760076f6541c6aa74c7a6e5c Author: Richard Biener <[email protected]> Date: Tue Oct 20 14:17:01 2026 +0200 tree-optimization/127975 - mixed AVX512 / AVX2 style masks with loop masking loop masking with AVX512 assumes integer type masks throughout, but when only a limited subset of AVX512 is enabled we can end up with mixed style masks which we do not support. Verify this doesn't happen and disable loop masking in that case instead of ICEing at transform time. PR tree-optimization/127975 * tree-vect-loop.cc (vect_verify_full_masking_avx512): Require all loop masks to be integer mode masks. * gcc.dg/vect/vect-simd-22.c: New testcase. (cherry picked from commit f4b64250a70482fbb51d6bb97fae481f90aa9649) Diff: --- gcc/testsuite/gcc.dg/vect/vect-simd-22.c | 14 ++++++++++++++ gcc/tree-vect-loop.cc | 29 ++++++++++++++++++----------- 2 files changed, 32 insertions(+), 11 deletions(-) diff --git a/gcc/testsuite/gcc.dg/vect/vect-simd-22.c b/gcc/testsuite/gcc.dg/vect/vect-simd-22.c new file mode 100644 index 000000000000..bbeb23f9e0db --- /dev/null +++ b/gcc/testsuite/gcc.dg/vect/vect-simd-22.c @@ -0,0 +1,14 @@ +/* { dg-do compile } */ +/* { dg-additional-options "-fopenmp-simd" } */ +/* { dg-additional-options "-mavx512vl -mno-avx512bw --param vect-partial-vector-usage=2" { target { avx512vl } } } */ + +#pragma omp declare simd simdlen(2) +float f(double); + +void g(float *d, double *a, int n) +{ +#pragma omp simd + for (int i = 0; i < n; i++) + d[i] = f(a[i]); + +} diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc index efe9dd7266a7..67b36d02adfe 100644 --- a/gcc/tree-vect-loop.cc +++ b/gcc/tree-vect-loop.cc @@ -1146,6 +1146,8 @@ vect_verify_full_masking_avx512 (loop_vec_info loop_vinfo) if (!iv_type) return false; + bool ok = true; + /* Produce the rgroup controls. */ for (auto const &mask : LOOP_VINFO_MASKS (loop_vinfo).mask_set) { @@ -1166,37 +1168,42 @@ vect_verify_full_masking_avx512 (loop_vec_info loop_vinfo) masks->rgc_vec.safe_grow_cleared (nscalars_per_iter, true); rgroup_controls *rgm = &(*masks).rgc_vec[nscalars_per_iter - 1]; + /* For now vect_get_loop_mask only supports integer mode masks + when we need to split it. */ + tree mask_type = truth_type_for (vectype); + if (GET_MODE_CLASS (TYPE_MODE (mask_type)) != MODE_INT + || TYPE_PRECISION (TREE_TYPE (mask_type)) != 1) + { + ok = false; + break; + } + if (!rgm->type || rgm->factor > nvectors) { - rgm->type = truth_type_for (vectype); + rgm->type = mask_type; rgm->compare_type = NULL_TREE; rgm->max_nscalars_per_iter = nscalars_per_iter; rgm->factor = nvectors; rgm->bias_adjusted_ctrl = NULL_TREE; } } + if (!ok) + { + release_vec_loop_controls (&LOOP_VINFO_MASKS (loop_vinfo).rgc_vec); + return false; + } /* There is no fixed compare type we are going to use but we have to be able to get at one for each mask group. */ unsigned int min_ni_width = wi::min_precision (vect_max_vf (loop_vinfo), UNSIGNED); - bool ok = true; for (auto &rgc : LOOP_VINFO_MASKS (loop_vinfo).rgc_vec) { tree mask_type = rgc.type; if (!mask_type) continue; - /* For now vect_get_loop_mask only supports integer mode masks - when we need to split it. */ - if (GET_MODE_CLASS (TYPE_MODE (mask_type)) != MODE_INT - || TYPE_PRECISION (TREE_TYPE (mask_type)) != 1) - { - ok = false; - break; - } - /* If iv_type is usable as compare type use that - we can elide the saturation in that case. */ if (TYPE_PRECISION (iv_type) >= min_ni_width)
