PRs 77536 and 68212 document cases where the estimated frequencies become unreliable to make appropriate decisions based on them. This patch removes the optimize_insn_for_speed_p check on the splitter so that the transformation will take place when -ffast-math is specified.
Bootstrap/regtest on powerpc64le, ok for trunk? -Pat 2016-09-13 Pat Haugen <pthau...@us.ibm.com> * config/rs6000/rs6000.md (div->recip splitter): Remove optimize_insn_for_speed_p condition. Index: config/rs6000/rs6000.md =================================================================== --- config/rs6000/rs6000.md (revision 240092) +++ config/rs6000/rs6000.md (working copy) @@ -4323,13 +4323,15 @@ (define_expand "recip<mode>3" ;; Split to create division from FRE/FRES/etc. and fixup instead of the normal ;; hardware division. This is only done before register allocation and with ;; -ffast-math. This must appear before the divsf3/divdf3 insns. +;; We used to also check optimize_insn_for_speed_p () but problems with guessed +;; frequencies (pr68212/pr77536) yields that unreliable so it was removed. (define_split [(set (match_operand:RECIPF 0 "gpc_reg_operand" "") (div:RECIPF (match_operand 1 "gpc_reg_operand" "") (match_operand 2 "gpc_reg_operand" "")))] "RS6000_RECIP_AUTO_RE_P (<MODE>mode) - && can_create_pseudo_p () && optimize_insn_for_speed_p () - && flag_finite_math_only && !flag_trapping_math && flag_reciprocal_math" + && can_create_pseudo_p () && flag_finite_math_only + && !flag_trapping_math && flag_reciprocal_math" [(const_int 0)] { rs6000_emit_swdiv (operands[0], operands[1], operands[2], true);