While working on PR 86871, I noticed we were being overly restrictive when handling variable-length vectors. For:
for (i : ...) { res = ...; for (j : ...) res op= ...; a[i] = res; } we don't need a reduction operation (although we do for double reductions like: res = ...; for (i : ...) for (j : ...) res op= ...; a[i] = res; which must still be rejected). Tested on aarch64-linux-gnu (with and without SVE), aarch64_be-elf and x86_64-linux-gnu. OK to install? Richard 2018-08-09 Richard Sandiford <richard.sandif...@arm.com> gcc/ * tree-vect-loop.c (vectorizable_reduction): Allow inner-loop reductions for variable-length vectors. gcc/testsuite/ * gcc.target/aarch64/sve/reduc_8.c: New test. Index: gcc/tree-vect-loop.c =================================================================== --- gcc/tree-vect-loop.c 2018-08-01 16:14:50.227052736 +0100 +++ gcc/tree-vect-loop.c 2018-08-09 15:38:35.230258362 +0100 @@ -6711,6 +6711,7 @@ vectorizable_reduction (stmt_vec_info st } if (reduction_type != EXTRACT_LAST_REDUCTION + && (!nested_cycle || double_reduc) && reduc_fn == IFN_LAST && !nunits_out.is_constant ()) { Index: gcc/testsuite/gcc.target/aarch64/sve/reduc_8.c =================================================================== --- /dev/null 2018-07-26 10:26:13.137955424 +0100 +++ gcc/testsuite/gcc.target/aarch64/sve/reduc_8.c 2018-08-09 15:38:35.230258362 +0100 @@ -0,0 +1,20 @@ +/* { dg-do compile } */ +/* { dg-options "-O2 -ftree-vectorize" } */ + +int +reduc (int *restrict a, int *restrict b, int *restrict c) +{ + for (int i = 0; i < 100; ++i) + { + int res = 0; + for (int j = 0; j < 100; ++j) + if (b[i + j] != 0) + res = c[i + j]; + a[i] = res; + } +} + +/* { dg-final { scan-assembler-times {\tcmpne\tp[0-9]+\.s, } 1 } } */ +/* We ought to use the CMPNE result for the SEL too. */ +/* { dg-final { scan-assembler-not {\tcmpeq\tp[0-9]+\.s, } { xfail *-*-* } } } */ +/* { dg-final { scan-assembler-times {\tsel\tz[0-9]+\.s, } 1 } } */