[PATCH, i386] Unify TARGET_SSE_MATH for trunc* patterns

[Uros Bizjak]

Hello!

This patch fixes TARGET_SSE_MATH for trunc* patterns. It prevents 
generation of cvtsd2ss insn for -mfpmath=sse. Patch was bootstrapped on 
pentium4-pc-linux-gnu, regtesting is in progress for c and c++.

In addition, povray-3.50c was built for various architectures (pentium, 
pentium3, pentium4) with all combinations of -mfpmath. A benchmark.pov 
picture was generated and visually checked for correctnes.

Here are the results of generating benchmark.pov picture for -O3 
-march=pentium4 -ffast-math and various -mfpmath:

pentium4 3.2G
gcc version 4.0.0 20041217 (experimental)
PovRay-3.50c

sse:
Time For Parse:    0 hours  0 minutes   1.0 seconds (1 seconds)
Time For Photon:   0 hours  0 minutes  37.0 seconds (37 seconds)
Time For Trace:    0 hours  5 minutes  31.0 seconds (331 seconds)
    Total Time:    0 hours  6 minutes   9.0 seconds (369 seconds)

387:
Time For Parse:    0 hours  0 minutes   1.0 seconds (1 seconds)
Time For Photon:   0 hours  0 minutes  34.0 seconds (34 seconds)
Time For Trace:    0 hours  5 minutes  16.0 seconds (316 seconds)
    Total Time:    0 hours  5 minutes  51.0 seconds (351 seconds)

mixed:
Time For Parse:    0 hours  0 minutes   2.0 seconds (2 seconds)
Time For Photon:   0 hours  0 minutes  38.0 seconds (38 seconds)
Time For Trace:    0 hours  5 minutes  36.0 seconds (336 seconds)
    Total Time:    0 hours  6 minutes  16.0 seconds (376 seconds)

As it can be seen, mfpmath=387 now beats -mfpmath=sse by 18 seconds, 
that is ~5%. I guess this is quite good :) .

Remaining TARGET_SSE_MATH issues remain with fix* patterns and fp 
compare patterns (this is PR target/19009). These issues will be fixed 
sometime next week.

2004-12-17  Uros Bizjak  <uros@kss-loka.si>

    * config/i386/i386.md (truncdfsf2, *truncdfsf2_2,
    *truncdfsf2_2_nooverlap, *truncdfsf2_sse_only_nooverlap:
     Unify enable constraint with respect to TARGET_SSE2,
    TARGET_80387, TARGET_SSE_MATH and TARGET_MIX_SSE_I387.
    (define_split): Same for correspondig split patterns.
    (truncdfsf2): Use TARGET_SSE_MATH for SSE pattern generation.
    (truncdfsf2_noop): Rename to truncdfsf2_i387_noop.
    (*truncdfsf2_1_sse): Rename to *truncdfsf2_2.
    (*truncdfsf2_1_sse_nooverlap): Rename to *truncdfsf2_2_nooverlap.
    (*truncdfsf2_2): Rename to *truncdfsf2_mixed.
    (*truncdfsf2_2_nooverlap): Rename to *truncdfsf2_mixed_nooverlap.
    (*truncdfsf2_3): Rename to *truncdfsf2_i387.
    (truncdfsf2_sse_only): Rename to truncdfsf2_sse.
    (*truncdfsf2_sse_only_nooverlap): Rename to
    *truncdfsf2_sse_nooverlap.
    (truncxf{s,d}f2_noop) Rename to truncxf{d,s}f2_i387_noop.
    (*truncxfsf2_2): Rename to *truncxfsf2_i387.

    (truncxf{s,d}f2, fmod{s,d}f3, drem{s,d}f3, log1p{s,d}f2,
    rint{s,d}f2, floor{s,d}f2, ceil{s,d}f2, btrunc{s,d}f2,
    nearbyint{s,d}f2): Use renamed patterns.

Uros.
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: fpmath.diff
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20041217/3f758675/attachment.ksh>