Hi All, In a process of testing omp-simd tests we found out that couple important benchmark did not show any speed-up with aggressive if-conversion since the hottest loops have not been vectorized. Here is a simple patch that resolves vectorization issue related to aggressive if-conversion: (1) Use the only mask for load/store of given size; (2) Use iterative algorithm to remove all multiple predicate uses in pattern tree - splitting of predicate definition may introduce new variable with multiple uses.
Reproducer is included to patch. Bootstrap and regression testing did not show any new failures. Is it OK for trunk? 2015-01-23 Yuri Rumyantsev <ysrum...@gmail.com> PR tree-optimization/64746 * tree-if-conv.c (mask_exists): New function. (predicate_mem_writes): Save created mask with given size for further use. (stmt_is_root_of_bool_pattern): Remove argument VAR and store to it. (ifcvt_repair_bool_pattern): Collect all statements that are root of bool pattern and use iterative algorithm to remove multiple uses of predicates, display number of required iterations. gcc/testsuite/ChangeLog * gcc.target/i386/avx2-vect-aggressive-1.c: New test.
64746.patch
Description: Binary data
