This is similar to Mark Gilsse's patch in the OP, except that it ensures that the expression will fold back to a single condition. I did include Richi's patch from #c6 to make it more likely to trigger asap.
I'd appreciate feedback on the match.pd changes; it's my first time looking into this new(ish) mini-language. r~
* gimplify.c (gimplify_expr) [VEC_COND_EXPR]: Allow the first argument to be is_gimple_condexpr. * match.pd: Simplify and + ior of vec_cond. diff --git a/gcc/gimplify.c b/gcc/gimplify.c index 7be6bd7..c434e55 100644 --- a/gcc/gimplify.c +++ b/gcc/gimplify.c @@ -10773,8 +10773,21 @@ gimplify_expr (tree *expr_p, gimple_seq *pre_p, gimple_seq *post_p, goto expr_2; } - case FMA_EXPR: case VEC_COND_EXPR: + { + gimplify_status r0, r1, r2; + r0 = gimplify_expr (&TREE_OPERAND (*expr_p, 0), pre_p, + post_p, is_gimple_condexpr, fb_rvalue); + r1 = gimplify_expr (&TREE_OPERAND (*expr_p, 1), pre_p, + post_p, is_gimple_val, fb_rvalue); + r2 = gimplify_expr (&TREE_OPERAND (*expr_p, 2), pre_p, + post_p, is_gimple_val, fb_rvalue); + ret = MIN (MIN (r0, r1), r2); + recalculate_side_effects (*expr_p); + break; + } + + case FMA_EXPR: case VEC_PERM_EXPR: /* Classified as tcc_expression. */ goto expr_3; diff --git a/gcc/match.pd b/gcc/match.pd index 5903782..4192d29 100644 --- a/gcc/match.pd +++ b/gcc/match.pd @@ -30,6 +30,7 @@ along with GCC; see the file COPYING3. If not see real_zerop real_onep real_minus_onep zerop CONSTANT_CLASS_P + COMPARISON_CLASS_P tree_expr_nonnegative_p integer_valued_real_p integer_pow2p @@ -2452,6 +2453,35 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT) replace if (x == 0) with tem = ~x; if (tem != 0) which is clearly less optimal and which we'll transform again in forwprop. */ +/* Simplification of vec_cond. */ + +(for bcode (bit_and bit_ior) + (simplify + (bcode (vec_cond COMPARISON_CLASS_P@0 integer_all_onesp@2 integer_zerop@3) + (vec_cond COMPARISON_CLASS_P@1 @2 @3)) + (with + { + tree al = TREE_OPERAND (@0, 0); + tree ar = TREE_OPERAND (@0, 1); + tree bl = TREE_OPERAND (@1, 0); + tree br = TREE_OPERAND (@1, 1); + } + (if (operand_equal_p (al, bl, 0) && operand_equal_p (ar, br, 0)) + (with + { + tree_code tcode + = bcode == BIT_AND_EXPR ? TRUTH_AND_EXPR : TRUTH_OR_EXPR; + tree folded + = combine_comparisons (UNKNOWN_LOCATION, tcode, TREE_CODE (@0), + TREE_CODE (@1), TREE_TYPE (@0), al, ar); + } + (if (folded) + (switch + (if (integer_truep (folded)) @2) + (if (integer_zerop (folded)) @3) + (if (COMPARISON_CLASS_P (folded)) + { build3 (VEC_COND_EXPR, TREE_TYPE (@2), folded, @2, @3); }))) + ))))) /* Simplification of math builtins. These rules must all be optimizations as well as IL simplifications. If there is a possibility that the new