[SH4] splits after reload

["Rakesh Kumar - Software, Noida" <rakeshku at noida dot hcltech dot com>]

Hi,

This mail is related to the recombination of address arithmetic insns
generated by DFmode splits after reload. The attached patch addresses 
PR 5360. Patch needs some improvement, but I would like to know the
viewpoint on applied strategy. The previous discussion of the problem
is available on the following thread:

http://gcc.gnu.org/ml/gcc/2003-12/msg00700.html

Toshi suggested that try_split () should have recombination
mechanism for combining split insns with the current instruction
stream. But as Jim Wilson says, it won't be that useful 
for many targets. The problem seems to be more acute for SH4.
So, I thought, it might be a good idea to fix this problem in a 
target dependent manner.

I have prepared a target dependent patch for SH4, attached with
the mail. When any DFmode Load/Store insn in split (in flow2), the
address arithmetic insn is recorded in an array. A marker
(UNSPECV_DF_SPLIT) is put in the insn stream on first split.
Then, in peep2 pass, these address arithmetic insns are recombined
with other ADD incsns in the same basic block if possible. The care
has been taken of deleted insns and blocks during flow2 pass. Following
is the overview of changes made:

  config/sh/sh.c  (fix_df_splits) : New function
                  (try_recombine) : New function
                  (split_array, split_index) : New variables defined
  config/sh/sh.md : New define_peephole2
                  : Modifications in define_split for DFmode load/store

Here are the size improvements I got through this little patch. The 
stress benchmarks are taken from stress-1.17 suite.

Size Gains (with compiler options -O2 -ml -m4)
==============================================
  Test-case name			GCC 3.4		With patch
%age gain
  --------------			-------		----------
---------
  slalom.c				11544			11416
1.11
  layer3.i				21840			21488
1.61
  map_fog.i				31840			31840
0
  revolt.i				6532			6532
0
  l3bitstream.i			7104			7072
0.45
  navion_aero.i			1664			1664
0

I am seeing some good performance gains as well. Will mail you the results
soon.

Please have a look at the patch and suggest me if it is OK to
work in this direction.

Thanks and Regards,
Rakesh Kumar

Attachment: patch-df-split-fix
Description: Binary data