In some rare cases, the register allocator can spill registers but end up not utilizing registers at all. The fundamental problem is linearscan's backtracking can end up freeing more than one allocated registers. However,  reloads and restores might be folded into uses / defs and freed registers might not be used at all.

VirtRegMap keeps track of allocations so it knows what's not used. As a horrible hack, the stack coloring can color spill slots with *free* registers. That is, it replace reload and spills with copies from and to the free register. It unfold instructions that load and store the spill slot and replace them with register using variants.

Not yet enabled. This is part 1. More coming.

llvm-svn: 70787

register destinations that are tied to source operands.  The
TargetInstrDescr::findTiedToSrcOperand method silently fails for inline
assembly.  The existing MachineInstr::isRegReDefinedByTwoAddr was very
close to doing what is needed, so this revision makes a few changes to
that method and also renames it to isRegTiedToUseOperand (for consistency
with the very similar isRegTiedToDefOperand and because it handles both
two-address instructions and inline assembly with tied registers).

llvm-svn: 68714

Remove the "fast" cases for spill and restore point determination, as these were subtlely wrong in obscure cases.  Patch the testcase
to account for this change.

llvm-svn: 68093

useful with it at the moment, but it will in the future.

llvm-svn: 67012

llvm-svn: 66158

Be more careful about choosing restore points when doing restore folding.  This fixes some subtle miscompilations.

llvm-svn: 66147

Add a restore folder, which shaves a dozen or so machineinstrs off oggenc.  Update a testcase to check this.

llvm-svn: 66029

llvm-svn: 65121

Add TargetInstrInfo::isSafeToMoveRegisterClassDefs. It returns true if it's safe to move an instruction which defines a value in the register class. Replace pre-splitting specific IgnoreRegisterClassBarriers with this new hook.

llvm-svn: 63936

between call frame setup/restore points.  Unfortunately, this regresses 
code size a bit, but at least it's correct now!

llvm-svn: 63837

Live interval reconstruction needs to account for this, and scour its maps to
prevent dangling references.

llvm-svn: 63558

llvm-svn: 63536

llvm-svn: 63492

Fix an issue in PHI construction that was exposed by GCC 4.2 producing a different set iteration order for the reg_iterator.

llvm-svn: 63490

Correct the algorithms for choosing spill and restore points so that we don't try to insert loads/stores between call frame setup and the actual call.

This fixes the last known failure for the pre-alloc-splitter.

llvm-svn: 63339

and an iterator invalidation issue.

FreeBench/pifft no longer miscompiles with these fixes!

llvm-svn: 63293

llvm-svn: 63276

vast majority of code size regressions introduced by pre-alloc-splitting.

llvm-svn: 63274

llvm-svn: 63091

llvm-svn: 63049

llvm-svn: 63041

llvm-svn: 63040

llvm-svn: 63026

markers, and ended up foiling the interval reconstruction.

This allows us to turn on reconstruction in the pre alloc splitter, which
fixes a number of miscompilations.

llvm-svn: 63025

llvm-svn: 63021

llvm-svn: 62917

going to be re-spilled again.

This also helps performance.  Pre-alloc-splitting now seems to be an overall win on SPEC.

llvm-svn: 62834

llvm-svn: 62821

I accidentally removed this check in an earlier commit, which cause breakage in the pre alloc splitter.

llvm-svn: 62678

llvm-svn: 62639

llvm-svn: 62600

llvm-svn: 62073

two address instructions.  We need to keep track of things we've processed AS USES
independetly of whether we've processed them as defs.

This fixes all known miscompilations when reconstruction is turned on.

llvm-svn: 61802

problem, rather than fixing it.  The problem has now been fixed the right way.

llvm-svn: 61723

llvm-svn: 61514

llvm-svn: 61465

llvm-svn: 61460

Add prototype code for recomputing a live interval's ranges and valnos through recursive phi construction.

llvm-svn: 61458

llvm-svn: 61182

llvm-svn: 61159