https://gcc.gnu.org/g:f6586e410cf92aa17b535c42b3b54e6030395e70
commit r16-9906-gf6586e410cf92aa17b535c42b3b54e6030395e70 Author: Richard Biener <[email protected]> Date: Tue Nov 3 13:29:18 2026 +0100 tree-optimization/127924 - use correct loop mask for multi-part simd clone masks When an inbranch simd clone call in a fully-masked loop needs multiple mask arguments per call (for example simdlen(32) with 16-lane AVX512 masks), vectorizable_simd_clone_call iterates M over the mask parts but fetched the loop mask with the call copy index J. All parts of a call were thus ANDed with the first loop mask, enabling lanes beyond the iteration count for the later parts and calling the clone on inactive lanes. Use M as the loop mask index instead. PR tree-optimization/127924 * tree-vect-stmts.cc (vectorizable_simd_clone_call): Use the mask part index M rather than the copy index J when getting the loop mask for explicit mask arguments. * gcc.target/i386/vect-simd-clone-avx512-4.c: New test. (cherry picked from commit 1b3a0deba8302912cfbc5faf850a6abcdb3e984a) Diff: --- .../gcc.target/i386/vect-simd-clone-avx512-4.c | 37 ++++++++++++++++++++++ gcc/tree-vect-stmts.cc | 4 +-- 2 files changed, 39 insertions(+), 2 deletions(-) diff --git a/gcc/testsuite/gcc.target/i386/vect-simd-clone-avx512-4.c b/gcc/testsuite/gcc.target/i386/vect-simd-clone-avx512-4.c new file mode 100644 index 000000000000..e23bdd427bbd --- /dev/null +++ b/gcc/testsuite/gcc.target/i386/vect-simd-clone-avx512-4.c @@ -0,0 +1,37 @@ +/* { dg-do run } */ +/* { dg-require-effective-target avx512bw } */ +/* { dg-options "-O2 -fopenmp-simd -mavx512bw -mprefer-vector-width=512 --param vect-partial-vector-usage=2" } */ + +#include "avx512bw-check.h" + +#pragma omp declare simd simdlen(32) inbranch +int __attribute__((noinline)) baz (int x) { return 1000 / x; } + +short d[1024]; +int b[1024]; + +void __attribute__((noipa)) +foo (int n) +{ + for (int i = 0; i < n; ++i) + { + d[i] = (short) (d[i] + 1); + if ((i & 1) == 0) + b[i] = baz (b[i]); + } +} + +static void +avx512bw_test (void) +{ + for (int i = 0; i < 1024; i++) + { + b[i] = i + 1; + d[i] = 0; + } + foo (17); + for (int i = 0; i < 64; i++) + if (b[i] != ((i < 17 && (i & 1) == 0) ? 1000 / (i + 1) : i + 1) + || d[i] != (i < 17)) + abort (); +} diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc index 9305bc6e3651..b13185e5a4b9 100644 --- a/gcc/tree-vect-stmts.cc +++ b/gcc/tree-vect-stmts.cc @@ -4720,7 +4720,7 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info, tree loop_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, ncopies_in, - vectype, j); + vectype, m); vec_oprnd0 = prepare_vec_mask (loop_vinfo, TREE_TYPE (loop_mask), @@ -4781,7 +4781,7 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info, tree loop_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, ncopies_in, - vectype, j); + vectype, m); vec_oprnd0 = prepare_vec_mask (loop_vinfo, TREE_TYPE (loop_mask),
