The IFN_GOMP_SIMD_LANE handling in vectorizable_store tries to use MEM_REF
offsets to maintain pointer disambiguation info. This patch makes sure
that we don't try to do the same optimisation for IFN_MASK_STOREs, which
have no similar offset argument.
The patch fixes libgomp.c-c++-common/pr66199-*.c for SVE. Previously
we had an ncopies==2 store and stored both halves to the same address.
Tested on aarch64-linux-gnu (with and without SVE), aarch64_be-elf
and x86_64-linux-gnu. OK to install?
Richard
2019-01-07 Richard Sandiford <richard.sandif...@arm.com>
gcc/
* tree-vect-stmts.c (vectorizable_store): Don't use the dataref_offset
optimization for masked stores.
Index: gcc/tree-vect-stmts.c
===================================================================
--- gcc/tree-vect-stmts.c 2019-01-04 11:39:27.190246648 +0000
+++ gcc/tree-vect-stmts.c 2019-01-07 13:23:28.048406652 +0000
@@ -7059,6 +7059,7 @@ vectorizable_store (stmt_vec_info stmt_i
bool simd_lane_access_p
= STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info);
if (simd_lane_access_p
+ && !loop_masks
&& TREE_CODE (DR_BASE_ADDRESS (first_dr_info->dr)) == ADDR_EXPR
&& VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr_info->dr), 0))
&& integer_zerop (DR_OFFSET (first_dr_info->dr))
