Take: ``` int test(int a, int b) { if (a <= b) return a < b ? a : b; return 0; } ``` This used to be optimized in GCC 11 and before via EVRP: pushing new range for b_3(D): int [a_2(D), +INF] EQUIVALENCES: { b_3(D) } (1 elements) evrp visiting stmt _5 = MIN_EXPR <a_2(D), b_3(D)>; extract_range_from_stmt visiting: _5 = MIN_EXPR <a_2(D), b_3(D)>; Intersecting int VARYING and int VARYING to int VARYING Folding statement: _5 = MIN_EXPR <a_2(D), b_3(D)>; Folded into: _5 = a_2(D); But in GCC 12 and above it is missed: Folding statement: _5 = MIN_EXPR <a_2(D), b_3(D)>; folding with relation a_2(D) <= b_3(D) Not folded
I suspect this was moving over to ranger and somehow this transformation was lost (maybe due to a missing testcase?)
I Kinda see how to implement this by creating operator_min::fold_range/operator_max::fold_range but I am still new on using these interfaces so I am not 100% sure how to use them.
(In reply to Andrew Pinski from comment #2) > I Kinda see how to implement this by creating > operator_min::fold_range/operator_max::fold_range but I am still new on > using these interfaces so I am not 100% sure how to use them. Actually, on ranger, we'd be able to make the range choice of the range of a_2 or b_3, but it can't rewrite the IL... and since the range of both is varying, fold_range would still return varying. Unless we indicate there are relations. fodl_range itself only takes what it is given, so we have to query the relations first. In theory all that is missing is to teach simplification about relation queries. For instance, in simplify_using_ranges::fold_cond_with_ops, we are invoking the range-op handler without any relations.. we query the ranges, but not the relation. If we add something like this (and make sure both operands are symbolic) diff --git a/gcc/vr-values.cc b/gcc/vr-values.cc index ecb294131b0..ad2c2d6c090 100644 --- a/gcc/vr-values.cc +++ b/gcc/vr-values.cc @@ -315,10 +315,17 @@ simplify_using_ranges::fold_cond_with_ops (enum tree_code code, || !query->range_of_expr (r1, op1, s)) return NULL_TREE; + relation_kind rel = VREL_VARYING; + if (gimple_range_ssa_p (op0) && gimple_range_ssa_p (op1)) + rel = query->query_relation (s, op0, op1); + // Create a trio with the relation set between op0 and op2 for folding. + // TRIOS are lhs-op0, lhs-op1, op0-op1 relations. + relation_trio trio (VREL_VARYING, VREL_VARYING, rel); + tree type = TREE_TYPE (op0); int_range<1> res; range_op_handler handler (code); - if (handler && handler.fold_range (res, type, r0, r1)) + if (handler && handler.fold_range (res, type, r0, r1, trio)) { if (res == range_true (type)) return boolean_true_node; This should do what you want I think... fold_range should use the relation passed in to determine that the condition is always true or false. I have not fully tested this patch, fwiw.
Just looking at the generated code of #c0 with -O2 on x86_64, this regressed with r13-3596-ge7310e24b1c0ca67b1bb507c1330b2bf39e59e32 Andrew, are you going to address this for GCC 14, or defer to GCC 15?
(In reply to Jakub Jelinek from comment #4) > Just looking at the generated code of #c0 with -O2 on x86_64, this regressed > with > r13-3596-ge7310e24b1c0ca67b1bb507c1330b2bf39e59e32 > Andrew, are you going to address this for GCC 14, or defer to GCC 15? Id prefer to defer it I think. Although we can run that thru the testings if anyone really wants it. Maybe in GCC 15 someone can add relations in general to simplifications
I think this is trivial to do in DOM and not handling these cases could easily be seen as an oversight. When we fail to find an expression in the hash table of available expressions, we have a bit of existing code that can ask about a relation between two operands of a binary operator and based on that relation possibly simplify the original expression. So for example, if we have: _4 = MIN_EXPR <a_2(D), b_3(D)>; And the MIN_EXPR expression isn't in the hash table, we look to see if we have recorded a_2 == b_3 and if so we simplify the MIN_EXPR into a copy. So this is just a matter of extending that code ever so slightly to do an additional lookup.
The master branch has been updated by Jeff Law <law@gcc.gnu.org>: https://gcc.gnu.org/g:8fe27ed193d60f6cd8b34761858a720c95eabbdb commit r14-9419-g8fe27ed193d60f6cd8b34761858a720c95eabbdb Author: jlaw <jeffreyalaw@gmail.com> Date: Sun Mar 10 11:58:00 2024 -0600 [committed] [PR tree-optimization/110199] Simplify MIN/MAX more often So as I mentioned in the BZ, the case of t = MIN_EXPR (A, B) where we know something about the relationship between A and B can be trivially handled by some existing code in DOM. That existing code would simplify when A == B. But by testing GE and LE instead of EQ we can cover more cases with minimal effort. When applicable the MIN/MAX turns into a simple copy. I made one other change. We have other binary operations that we simplify when we know something about the relationship between the operands. That code was not canonicalizing the order of operands when building the expression to lookup in the hash tables to discover that relationship. Since those paths are only testing for equality, we can trivially reverse them and not have to worry about changing codes or anything like that. So extremely safe and avoids having to come back and fix that code to match the MIN_EXPR/MAX_EXPR case later. Bootstrapped on x86 and also tested on the crosses. I briefly thought there was an sh regression, but that was actually the recent fwprop changes twiddling code generation for one test. PR tree-optimization/110199 gcc/ * tree-ssa-scopedtables.cc (avail_exprs_stack::simplify_binary_operation): Generalize handling of MIN_EXPR/MAX_EXPR to allow additional simplifications. Canonicalize comparison operands for other cases. gcc/testsuite * gcc.dg/tree-ssa/minmax-27.c: New test. * gcc.dg/tree-ssa/minmax-28.c: New test.