[1/2] Rename across ebbs

[Bernd Schmidt]

While working on a C6X scheduling patch, I found myself wondering - what
would be involved in making the register renamer operate on extended
basic blocks rather than simple bbs? Somewhat surprisingly, the answer
turns out to be "not much". After the last rewrite, all the conflict
tests are based on id numbers and bitmaps, not live ranges.

The following two patches add support for saving and restoring renamer
state at basic block boundaries. Below is the initial patch, which just
reorders the code a little without changing functionality. Besides
moving and splitting out functions, I've only removed one redundant
structure field and some #if 0'ed code.

Bootstrapped and regression tested (with both patches applied) on
i686-linux, with

@@ -5106,6 +5106,9 @@ static const struct default_options ix86
+#ifdef INSN_SCHEDULING
+    { OPT_LEVELS_2_PLUS, OPT_frename_registers, NULL, 1 },
+#endif

to ensure it gets tested. Also tested with a 4.5 c6x-elf toolchain,
which has -frename-registers enabled by default. With a few other
changes to make full use of it, I get up to 10% speedup on certain
testcases on c6x. It would be nice if someone could run benchmarks on
other targets where this could affect performance. I may or may not
manage an i686 SPEC run over the weekend; it's unlikely to be affected
much due to not using sched_ebb.


Bernd
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: rnreg-prelim.diff
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20110527/f0d80fca/attachment.ksh>