[Bug rtl-optimization/70873] [7 Regressio] 20% performance regression at 482.sphinx3 after r235442 with -O2 -m32 on Haswell.

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70873

--- Comment #18 from H.J. Lu <hjl.tools at gmail dot com> ---
For float_truncate,

(define_split
  [(set (match_operand:SF 0 "register_operand")
        (float_truncate:SF
          (match_operand:DF 1 "nonimmediate_operand")))]
  "TARGET_USE_VECTOR_FP_CONVERTS
   && optimize_insn_for_speed_p ()
   && reload_completed && SSE_REG_P (operands[0])
   && (!EXT_REX_SSE_REG_P (operands[0])
       || TARGET_AVX512VL)"

;; It's more profitable to split and then extend in the same register.
(define_peephole2
  [(set (match_operand:SF 0 "register_operand")
        (float_truncate:SF
          (match_operand:DF 1 "memory_operand")))]
  "TARGET_SPLIT_MEM_OPND_FOR_FP_CONVERTS
   && optimize_insn_for_speed_p ()
   && SSE_REG_P (operands[0])"
  [(set (match_dup 2) (match_dup 1))
   (set (match_dup 0) (float_truncate:SF (match_dup 2)))]
  "operands[2] = gen_rtx_REG (DFmode, REGNO (operands[0]));")