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

[uros at gcc dot gnu.org]

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

--- Comment #28 from uros at gcc dot gnu.org ---
Author: uros
Date: Wed May  4 21:13:13 2016
New Revision: 235906

URL: https://gcc.gnu.org/viewcvs?rev=235906&root=gcc&view=rev
Log:
        PR target/70873
        * config/i386/i386.md
        (TARGET_SSE_PARTIAL_REG_DEPENDENCY float_extend sf->df peephole2):
        Change to post-epilogue_completed late splitter.  Use sse_reg_operand
        as operand 0 predicate.
        (TARGET_SSE_PARTIAL_REG_DEPENDENCY float_truncate df->sf peephole2):
        Ditto.
        (TARGET_SSE_PARTIAL_REG_DEPENDENCY float {si,di}->{sf,df} peephole2):
        Ditto.  Emit the pattern using RTX.

        (TARGET_USE_VECTOR_FP_CONVERTS float_extend sf->df splitter):
        Use sse_reg_opreand as operand 0 predicate.  Do not use true_regnum in
        the post-reload splitter.  Use lowpart_subreg instead of gen_rtx_REG.
        (TARGET_USE_VECTOR_FP_CONVERTS float_truncate df->sf splitter):
        Ditto.
        (TARGET_USE_VECTOR_CONVERTS float si->{sf,df} splitter): Use
        sse_reg_operand as operand 0 predicate.

        (TARGET_SPLIT_MEM_OPND_FOR_FP_CONVERTS float_extend sf->df peephole2):
        Use sse_reg_opreand as operand 0 predicate.  Use lowpart_subreg
        instead of gen_rtx_REG.
        (TARGET_SPLIT_MEM_OPND_FOR_FP_CONVERTS float_truncate sf->df
peephole2):
        Ditto.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/i386/i386.md