quality hit. will look at this soon.

llvm-svn: 37748

some issues)

llvm-svn: 37747

llvm-svn: 37746

llvm-svn: 37745

llvm-svn: 37743

2. When calculating ANTIC_IN, only iterate the changed blocks.  For most average
inputs this is a small speedup, but for cases with unusual CFGs, this can be a significant win.

llvm-svn: 37742

llvm-svn: 37740

Properly handle kills of a physical register which has sub-registers that are read by later instructions.

llvm-svn: 37739

llvm-svn: 37737

llvm-svn: 37731

instruction flag, and use the flag along with a virtual member function
hook for targets to override if there are instructions that are only
trivially rematerializable with specific operands (i.e. constant pool
loads).

llvm-svn: 37728

llvm-svn: 37725

and code quality by ~2% on my tests.

A big thank you to Roman Levenstein for this patch! See
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070618/050717.html
for more details.

llvm-svn: 37724

with weak linkage.

llvm-svn: 37723

llvm-svn: 37721

extended vector types. Remove the special SDNode opcodes used for pre-legalize
vector operations, and the special MVT::Vector type used with them. Adjust
lowering and legalize to work with the normal SDNode kinds instead, and to
use the normal MVT functions to work with vector types instead of using the
two special operands that the pre-legalize nodes held.

This allows pre-legalize and post-legalize DAGs, and the code that operates
on them, to be more consistent. Pre-legalize vector operators can be handled
more consistently with scalar operators. And, -view-dag-combine1-dags and
-view-legalize-dags now look prettier for vector code.

llvm-svn: 37719

between SSE1 instructions and their respective SSE2 analogues.

llvm-svn: 37718

llvm-svn: 37717

llvm-svn: 37716

llvm-svn: 37715

2) Remove some maximal-set computing code that is no longer used.
3) Use a post-order CFG traversal to compute ANTIC_IN instead of a postdom traversal.
This causes the ANTIC_IN calculation to converge much faster.  Thanks to Daniel Berlin for suggesting this.

With this patch, the time to optimize 403.gcc decreased from 17.5s to 7.5s, and Anton's huge
testcase decreased from 62 minutes to 38 seconds.

llvm-svn: 37714

llvm-svn: 37713

llvm-svn: 37712

llvm-svn: 37710

blocks, and optimizing within one basic block.

llvm-svn: 37709

Rework topo_sort so eliminate some behavior that scaled terribly.  This reduces the time to optimize 403.gcc from 18.2s to 17.5s,
and has an even larger effect on larger testcases.

llvm-svn: 37708

Perform fewer set insertions while calculating ANTIC_IN.  This reduces the amount of time to optimize 403.gcc from 21.9s to 18.2s.

llvm-svn: 37707

Remove some code that I was using for collecting performance information that should not have been committed.

llvm-svn: 37706

llvm-svn: 37705

TargetLowering to SelectionDAG so that they have more convenient
access to the current DAG, in preparation for the ValueType routines
being changed from standalone functions to members of SelectionDAG for
the pre-legalize vector type changes.

llvm-svn: 37704

llvm-svn: 37703

Avoid excessive calls to find_leader when calculating AVAIL_OUT.  This reduces the time to optimize 403.gcc from 23.5s to 21.9s.

llvm-svn: 37702

std::set is really really terrible. Switch to SmallPtrSet to reduce compile time. For Duraid's example. The overall isel time is reduced from 0.6255 sec to 0.1876 sec.

llvm-svn: 37701

llvm-svn: 37700

Reserve space in vectors before topologically sorting into them.  This improves the time to optimize 403.gcc from 28s to 23.5s.

llvm-svn: 37699

Make a bunch of optimizations for compile time to GVNPRE, including smarter set unions, deferring blocks rather than computing maximal sets, and smarter use of sets.  With these enhancements, the time to optimize 273.perlbmk goes from 5.3s to 2.7s.

llvm-svn: 37698

llvm-svn: 37697

 1. Make SmallPtrSet::erase faster in the small case by replacing a memmove
    with a pointer copy.
 2. Fix a bug where the null terminator at the end of the array in the small
    case was not copied

llvm-svn: 37696

Use schedulePass() to accomodate these requirement instead of directly
assigning a manager to new CallGraph PassManager.

llvm-svn: 37695

transformation.  Also, keep track of which end of the integer interval overflows
occur on.  This fixes Transforms/InstCombine/2007-06-21-DivCompareMiscomp.ll
and rdar://5278853, a miscompilation of perl.

llvm-svn: 37692