https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86829
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |missed-optimization, patch
Status|UNCONFIRMED |NEW
Last reconfirmed| |2018-08-02
CC| |rguenth at gcc dot gnu.org
Version|unknown |9.0
Ever confirmed|0 |1
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
+ /* Simplify sin(atan(x)) -> x / sqrt(x*x + 1). */
+ (for sins (SIN)
+ atans (ATAN)
+ sqrts (SQRT)
+ (simplify
+ (sins (atans @0))
+ (rdiv @0 ( sqrts (plus (mult @0 @0)
+ { build_one_cst (float_type_node);})))))
+
+
+ /* Simplify cos(atan(x)) -> 1 / sqrt(x*x + 1). */
+ (for coss (COS)
+ atans (ATAN)
+ sqrts (SQRT)
+ (simplify
+ (coss (atans @0))
+ (rdiv { build_one_cst (float_type_node);}
+ ( sqrts (plus (mult @0 @0) { build_one_cst (float_type_node);})))))
likely happens because you use float_type_node instead of 'type' for
build_one_cst. I think that if the atan intermediate result has additional
uses then the substitution may not be beneficial? Thus you should
write (atans:s @0) in both places.
Otherwise this looks OK. Can you please post the patch to gcc-patches
and add testcases?
