llvm-svn: 91823

the underlying PHI node insertion issue in SSAUpdate is fixed.

llvm-svn: 91821

'GetValueInMiddleOfBlock' case, instead of inserting 
duplicates.

A similar fix is almost certainly needed by the machine-level
SSAUpdate implementation.

llvm-svn: 91820

implement some optimizations for MIN(MIN()) and MAX(MAX()) and 
MIN(MAX()) etc.  This substantially improves the code in PR5822 but
doesn't kick in much elsewhere.  2 max's were optimized in 
pairlocalalign and one in smg2000.

llvm-svn: 91814

Use the presence of NSW/NUW to fold "icmp (x+cst), x" to a constant in
cases where it would otherwise be undefined behavior.

Surprisingly (to me at least), this triggers hundreds of the times in
a few benchmarks: lencode, ldecode, and 466.h264ref seem to *really*
like this.

llvm-svn: 91812

a bunch in lencode, ldecod, spass, 176.gcc, 252.eon, among others.  It is 
also the first part of PR5822

llvm-svn: 91811

cache a pointer as being unavailable due to phi trans in the
wrong place.  This would cause later queries to fail even when
they didn't involve phi trans.

llvm-svn: 91787

llvm-svn: 91780

llvm-svn: 91778

llvm-svn: 91777

rename dprintf to dbgpritnf, in order to fix build with glibc (which already defines dprintf in stdio.h

llvm-svn: 91775

1. In indirect load/store insns , the name of fsr should be emitted as INDF.
2. include standard asmbly headers in generated asmbly.

llvm-svn: 91768

llvm-svn: 91764

where instcombine would have to split a critical edge due to a
phi node of an invoke.  Since instcombine can't change the CFG,
it has to bail out from doing the transformation.

llvm-svn: 91763

* change FindElementAndOffset to return a uint64_t instead of unsigned, and
  to identify the type to be used for that result in a GEP instruction.
* move "isa<ConstantInt>" to be first in conditional.
* replace some dyn_casts with casts.
* add a comment about handling mem intrinsics.

llvm-svn: 91762

llvm-svn: 91754

incarnations), integrated into the MC framework.  

The disassembler is table-driven, using a custom TableGen backend to 
generate hierarchical tables optimized for fast decode.  The disassembler 
consumes MemoryObjects and produces arrays of MCInsts, adhering to the 
abstract base class MCDisassembler (llvm/MC/MCDisassembler.h).

The disassembler is documented in detail in

- lib/Target/X86/Disassembler/X86Disassembler.cpp (disassembler runtime)
- utils/TableGen/DisassemblerEmitter.cpp (table emitter)

You can test the disassembler by running llvm-mc -disassemble for i386
or x86_64 targets.  Please let me know if you encounter any problems
with it.

llvm-svn: 91749

Based on patch by Michael Beck!

llvm-svn: 91745

llvm-svn: 91744

llvm-svn: 91742

Make some methods const.  The only interesting change here is that
it changes raw_fd_ostream::preferred_buffer_size to return zero on
a scary stat failure instead of setting the stream to an error state.
This method really should not mutate the stream.

llvm-svn: 91740

llvm-svn: 91736

the templates in TypesContext.h can find them.  Caught by clang++.

llvm-svn: 91735

llvm-svn: 91729

- Move DisableScheduling flag into TargetOption.h
- Move SDNodeOrdering into its own header file. Give it a minimal interface that
  doesn't conflate construction with storage.
- Move assigning the ordering into the SelectionDAGBuilder.

This isn't used yet, so there should be no functional changes.

llvm-svn: 91727

llvm-svn: 91722

llvm-svn: 91717

llvm-svn: 91715

debugging some leaks (PR5770 in particular).

llvm-svn: 91713

bootstrap.  This also replaces the WeakVH references that Chris objected to
with normal Value references.

llvm-svn: 91711

- an MDNode is designated as function-local when created, and continues to be even if its operands are modified not to refer to function-local IR
- function-localness is designated via lowest bit in SubclassData
- getLocalFunction() descends MDNode tree to see if it is consistently function-local

Add verification of MDNodes to checks that MDNodes are consistently function-local.
Update AsmWriter to use isFunctionLocal().

llvm-svn: 91708

llvm-svn: 91706

associative either.

llvm-svn: 91701

MI was not being used but it was also not being deleted, so it was kept in the garbage list. The memory itself was freed once the function code gen was done.

Once in a while the codegen of another function would create an instruction on the same address. Adding it to the garbage group would work once, but when another pointer was added it would cause an assert as "Cache" was about to be pushed to Ts.

For a patch that make us detect problems like this earlier, take a look at

http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20091214/092758.html

With that patch we assert as soon and the new instruction is added to the garbage set.

llvm-svn: 91691

Patch contributed by Ken Werner of IBM!

llvm-svn: 91681

Patch contributed by Ken Werner of IBM!

llvm-svn: 91680

multiple uses.  (The construct in question was found in gcc.)

llvm-svn: 91675

be non-optimal. To be precise, we should avoid folding loads if the instructions
only update part of the destination register, and the non-updated part is not
needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these instructions breaks
the partial register dependency and it can improve performance. e.g.

movss (%rdi), %xmm0
cvtss2sd %xmm0, %xmm0

instead of
cvtss2sd (%rdi), %xmm0

An alternative method to break dependency is to clear the register first. e.g.
xorps %xmm0, %xmm0
cvtss2sd (%rdi), %xmm0

llvm-svn: 91672

is not reassociative.

llvm-svn: 91667

llvm-svn: 91666