On Wed, Sep 05, 2018 at 03:14:40PM +0200, Jakub Jelinek wrote:
> It isn't really a GIMPLE transform, just needs to provide the callers
> with a tree. Specifically, if it passes several initial checks, it
> expands the X expression unconditionally and then just uses the result of
> that. What I could and probably should do is to just emit the cheaper
> of the two sequences and make_tree the result thereof, so the caller
> would emit only the actual comparison, rather than the modulo or its
> replacement again too as the patch does.
In particular I meant this incremental patch, where the effect is really
that the first argument of the modulo is expanded unconditionally once,
the two possible variants are then each expanded into a sequence, then based
on cost we decide which one to use, emit it (the other one is GCed) and
let the callers emit the comparison.
--- gcc/expr.c.jj 2018-09-05 15:34:10.395601918 +0200
+++ gcc/expr.c 2018-09-05 15:47:00.609834755 +0200
@@ -11716,12 +11716,13 @@ maybe_optimize_mod_cmp (enum tree_code c
if (mocost <= mucost)
{
- *arg0 = build2_loc (loc, TRUNC_MOD_EXPR, TREE_TYPE (treeop0),
- treeop0, treeop1);
+ emit_insn (moinsns);
+ *arg0 = make_tree (TREE_TYPE (*arg0), mor);
return code;
}
- *arg0 = t;
+ emit_insn (muinsns);
+ *arg0 = make_tree (type, mur);
*arg1 = c4;
return code == EQ_EXPR ? LE_EXPR : GT_EXPR;
}
Jakub
