https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97909
Bug ID: 97909
Summary: expr_not_equal_to (mainly in match.pd) vs. ranger
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: jakub at gcc dot gnu.org
Target Milestone: ---
Something I've noticed today and filing this so that it isn't forgotten.
Various places in match.pd use expr_not_equal_to to decide based on value
ranges whether to perform something or not.
This uses just SSA_NAME_RANGE_INFO under the hood right now.
Would be nice to use range for this; bet we'd need a way for match.pd to
provide
extra information on where it is, so either a gimple *, or perhaps for match.pd
it would be easier to provide a tree, which would be usually SSA_NAME of the
lhs of the statement we want to query the range at, and then the function would
check if the tree is an SSA_NAME and in that case talk to ranger and ask about
range at the start of the SSA_NAME_DEF_STMT.
E.g. we have:
/* X % -Y is the same as X % Y. */
(simplify
(trunc_mod @0 (convert? (negate @1)))
(if (INTEGRAL_TYPE_P (type)
&& !TYPE_UNSIGNED (type)
&& !TYPE_OVERFLOW_TRAPS (type)
&& tree_nop_conversion_p (type, TREE_TYPE (@1))
/* Avoid this transformation if X might be INT_MIN or
Y might be -1, because we would then change valid
INT_MIN % -(-1) into invalid INT_MIN % -1. */
&& (expr_not_equal_to (@0, wi::to_wide (TYPE_MIN_VALUE (type)))
|| expr_not_equal_to (@1, wi::minus_one (TYPE_PRECISION
(TREE_TYPE (@1))))))
(trunc_mod @0 (convert @1))))
so if expr_not_equal_to has a defaulted = NULL_TREE extra argument, the second
expr_not_equal_to could pass as the third argument @2 where it would be
(negate@2 @1). Unclear if we have something to pass for the trunc_mod's LHS or
how to identify that statement.
Testcase for exactly this could be:
int x = foo (), y;
if (x != -1)
y = z % -x;
else
y = -1;
where we could optimize the y = z % -x; to y = z % x; but don't currently,
because we know that x is not -1 only in certain paths and SSA_NAME_RANGE_INFO
can't reflect that.
