New testcase for PR310

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12596 91177308-0d34-0410-b5e6-96231b3b80d8

Remove some assertions that are now bogus with the last patch I put in

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12595 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12593 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR306: Loop simplify incorrectly updates dominator information
Testcase: LoopSimplify/2004-04-01-IncorrectDomUpdate.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12592 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase for PR306

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12591 91177308-0d34-0410-b5e6-96231b3b80d8

Fix grammar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12590 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for select constant expressions to the CBE, fixing SIOD

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12589 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify code by using the more powerful BuildMI forms.
Implement a small optimization.  In test/Regression/CodeGen/X86/select.ll,
we now generate this for foldSel3:

foldSel3:
        mov %AL, BYTE PTR [%ESP + 4]
        fld DWORD PTR [%ESP + 8]
        fld DWORD PTR [%ESP + 12]
        mov %EAX, DWORD PTR [%ESP + 16]
        mov %ECX, DWORD PTR [%ESP + 20]
        cmp %EAX, %ECX
        fxch %ST(1)
        fcmovae %ST(0), %ST(1)
***     fstp %ST(1)
        ret

Instead of:

foldSel3:
        mov %AL, BYTE PTR [%ESP + 4]
        fld DWORD PTR [%ESP + 8]
        fld DWORD PTR [%ESP + 12]
        mov %EAX, DWORD PTR [%ESP + 16]
        mov %ECX, DWORD PTR [%ESP + 20]
        cmp %EAX, %ECX
        fxch %ST(1)
        fcmovae %ST(0), %ST(1)
***     fxch %ST(1)
***     fstp %ST(0)
        ret

In practice, this only effects code size: performance should be basically
unaffected.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12588 91177308-0d34-0410-b5e6-96231b3b80d8

Wrap at 80 cols

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12587 91177308-0d34-0410-b5e6-96231b3b80d8

Allow converting a builder to an iterator

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12586 91177308-0d34-0410-b5e6-96231b3b80d8

Tests for fp cmov's that I forgot to check in earlier

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12585 91177308-0d34-0410-b5e6-96231b3b80d8

Clear out all of the changes.  Reset version numbers to 1.3
add note about select

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12584 91177308-0d34-0410-b5e6-96231b3b80d8

Add clear() forwarding method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12580 91177308-0d34-0410-b5e6-96231b3b80d8

Generate slightly smaller code, "test R, R" instead of "cmp R, 0"

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12579 91177308-0d34-0410-b5e6-96231b3b80d8

The X86 backend no longer needs the select lowering pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12578 91177308-0d34-0410-b5e6-96231b3b80d8

Codegen FP select instructions into X86 conditional moves.  Annoyingly enough
the X86 does not support a full set of fp cmove instructions, so we can't always
fold the condition into the select.  :(  Yuck.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12577 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for floating point conditional move instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12576 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for FP cmoves

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12575 91177308-0d34-0410-b5e6-96231b3b80d8

Add FP conditional move instructions, which annoyingly have special properties
that require the asmwriter to be extended (printing implicit uses before the
explicit operands)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12574 91177308-0d34-0410-b5e6-96231b3b80d8

Add warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12573 91177308-0d34-0410-b5e6-96231b3b80d8

MBB::remove should not modify the iterator passed in

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12572 91177308-0d34-0410-b5e6-96231b3b80d8

MachineBasicBlock::remove should not modify the iterator passed in

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12571 91177308-0d34-0410-b5e6-96231b3b80d8

Improve description, add warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12570 91177308-0d34-0410-b5e6-96231b3b80d8

MachineBasicBlock::remove should not change the iterator passed into it

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12569 91177308-0d34-0410-b5e6-96231b3b80d8

Factor out getStaticStackSize from InsertPrologCode(), so that I can more
easily steal it for a separate use in the reoptimizer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12568 91177308-0d34-0410-b5e6-96231b3b80d8

Use the true, decoded name of the archive member in getObjectType.

In ReadArchiveBuffer, make sure that MemberName is set in the case where
getObjectType would want to return SVR4LongFilename.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12567 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid TRUE and FALSE which apparently conflict with some macros on OSX

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12566 91177308-0d34-0410-b5e6-96231b3b80d8

Fix linking of constant expr casts due to type resolution changes.  With
this and the other patches 253.perlbmk links again.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12565 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for constant expr casts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12564 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for reading constantexpr select instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12563 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a latent bug in select constantexpr handling that was broke 253.perlbmk

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12562 91177308-0d34-0410-b5e6-96231b3b80d8

Correctly update LiveVariables when an instruction changes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12561 91177308-0d34-0410-b5e6-96231b3b80d8

Fold comparisons into select instructions, making much better code and
using our broad selection of movcc instructions.  :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12560 91177308-0d34-0410-b5e6-96231b3b80d8

Test folding comparisons into select instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12559 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase.  This now codegens to:

max:
        mov %EAX, DWORD PTR [%ESP + 4]
        mov %ECX, DWORD PTR [%ESP + 8]
        cmp %EAX, %ECX
        cmovle %EAX, %ECX
        ret

Someone should really implement passing arguments through registers for
known-internal functions.  :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12558 91177308-0d34-0410-b5e6-96231b3b80d8

Implement spill code folding for all of the conditional move instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12554 91177308-0d34-0410-b5e6-96231b3b80d8

Add direct support for integer select instructions, though we still don't support
folding compares into the select yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12553 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase for select instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12552 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bug: Assembler/2004-03-30-UnclosedFunctionCrash.llx

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12551 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase that crashes the assembler

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12550 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some serious bugs in the cmov descriptions, which didn't cause a problem because
we never generated them

Make indentation a bit more consistent

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12549 91177308-0d34-0410-b5e6-96231b3b80d8

Start cleaning up this pass so that I can debug it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12548 91177308-0d34-0410-b5e6-96231b3b80d8

Oops, actually USE the previously computed value

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12547 91177308-0d34-0410-b5e6-96231b3b80d8

Test general value/value selection which we can do now that we use the
select instruction

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12546 91177308-0d34-0410-b5e6-96231b3b80d8

Now that all the code generators support the select instruction, and the instcombine
pass can eliminate many nasty cases of them, start generating them in the optimizers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12545 91177308-0d34-0410-b5e6-96231b3b80d8

Implement select.ll:test[3-6]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12544 91177308-0d34-0410-b5e6-96231b3b80d8

Add some testcases for select simplification

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12543 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a fairly major performance problem.  If a PHI node had a constant as
an incoming value from a block, the selector would evaluate the constant
at the TOP of the block instead of at the end of the block.  This made the
live range for the constant span the entire block, increasing register
pressure needlessly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12542 91177308-0d34-0410-b5e6-96231b3b80d8

Add the select lowering pass to get initial support for select instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12541 91177308-0d34-0410-b5e6-96231b3b80d8

Add a simple select instruction lowering pass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12540 91177308-0d34-0410-b5e6-96231b3b80d8

Add some new methods

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12539 91177308-0d34-0410-b5e6-96231b3b80d8

Don't warn about a null live range if the Value is a ConstantIntegral.
Otherwise, if you're in debugging mode, you get warnings for (apparently)
every immediate constant in the function during reg. allocation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12538 91177308-0d34-0410-b5e6-96231b3b80d8

Change how the beginnings and ends of MachineFunctions are printed. Get
rid of the funny cast.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12537 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust to new itf

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12534 91177308-0d34-0410-b5e6-96231b3b80d8

Relax the interface a bit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12533 91177308-0d34-0410-b5e6-96231b3b80d8

Merged in RELEASE_12.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12532 91177308-0d34-0410-b5e6-96231b3b80d8

Merged in RELEASE_12.
This fixes the zlib linking problems.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12531 91177308-0d34-0410-b5e6-96231b3b80d8

Handle -0.0 correctly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12530 91177308-0d34-0410-b5e6-96231b3b80d8

Add a faq entry

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12529 91177308-0d34-0410-b5e6-96231b3b80d8

Add a bunch of methods that should have been added a long time ago.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12526 91177308-0d34-0410-b5e6-96231b3b80d8

Add another check

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12525 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an assertion
Contributed by Reid Spencer

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12524 91177308-0d34-0410-b5e6-96231b3b80d8

Make error message a bit nicer.
Contributed by Reid Spencer

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12523 91177308-0d34-0410-b5e6-96231b3b80d8

Add two methods which have been needed for a long time: Type::get(Un)signedVersion

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12522 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure to get the headers from zlib correctly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12521 91177308-0d34-0410-b5e6-96231b3b80d8

X % -1 == X % 1 == 0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12520 91177308-0d34-0410-b5e6-96231b3b80d8

Two changes:
#1 is to unconditionally strip constantpointerrefs out of
instruction operands where they are absolutely pointless and inhibit
optimization.  GRRR!

#2 is to implement InstCombine/getelementptr_const.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12519 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12518 91177308-0d34-0410-b5e6-96231b3b80d8

Minor efficiency improvement, finegrainify namespacification

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12517 91177308-0d34-0410-b5e6-96231b3b80d8

Add note about setting up linux systems to directly execute LLVM bytecode files
Fixes PR303

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12516 91177308-0d34-0410-b5e6-96231b3b80d8

Hrm, we were leaking ~1M of garbage that valgrind never told us about because
it was "reachable".  Cute.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12515 91177308-0d34-0410-b5e6-96231b3b80d8

Add a dose of reality into the release notes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12510 91177308-0d34-0410-b5e6-96231b3b80d8

Teach the optimizer to delete zero sized alloca's (but not mallocs!)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12507 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase, the optimizer can delete zero sized allocas

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12506 91177308-0d34-0410-b5e6-96231b3b80d8

Perhaps the last bugfix before the release?  HOPEFULLY!  :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12505 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR299: [Sparc] Code generator asserts on alloc of zero size type

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12504 91177308-0d34-0410-b5e6-96231b3b80d8

Malloc doesn't kill a load.  This patch need not go into 1.2 though.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12500 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a minor wording bug in the manual

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12496 91177308-0d34-0410-b5e6-96231b3b80d8

This is the last remaining bug that I know of in the loop extractor.  The loop
extractor has extracted hundreds of loops from the SPEC benchmarks without crashing
and without misoptimizing the programs.

This testcase could be reduced substantially more by hand, but I don't have time
to work on it right now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12495 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a really nasty bug that was breaking ijpeg in LLC mode.  We were incorrectly
folding load instructions into other instructions across free instruction
boundaries.  Perhaps this will also fix the other strange failures?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12494 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bug: CodeExtractor/2004-03-17-MissedLiveIns.ll
With this fix we now successfully extract all 149 loops from 256.bzip2 without
crashing or miscompiling the program!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12493 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase, reduced from 256.bzip2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12492 91177308-0d34-0410-b5e6-96231b3b80d8

Add statistics to the loop extractor.  The loop extractor has successfully
extracted all 63 loops for Olden/bh without crashing and without
miscompiling the program!!!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12491 91177308-0d34-0410-b5e6-96231b3b80d8

Fix problem with PHI nodes having multiple predecessors from different
exit nodes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12490 91177308-0d34-0410-b5e6-96231b3b80d8

Fix CodeExtractor/2004-03-17-UpdatePHIsOutsideRegion.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12489 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12488 91177308-0d34-0410-b5e6-96231b3b80d8

Seriously simplify and correct the PHI node handling code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12487 91177308-0d34-0410-b5e6-96231b3b80d8

Fix CodeExtractor/2004-03-17-OutputMismatch.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12486 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase that crashes the code extractor

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12485 91177308-0d34-0410-b5e6-96231b3b80d8

Fix several bugs in the extractor:
1. Names were not put on the new arguments created (ok, this just helps sanity :)
2. Fix outgoing pointer values
3. Do not insert stores for values that had not been computed
4. Fix some wierd problems with the outset calculation

This fixes CodeExtractor/2004-03-14-DominanceProblem.ll, making the extractor
work on at least one simple case!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12484 91177308-0d34-0410-b5e6-96231b3b80d8

The code extractor needs dominator info.  Provide it

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12483 91177308-0d34-0410-b5e6-96231b3b80d8

The code extractor needs dominator information, so we provide it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12482 91177308-0d34-0410-b5e6-96231b3b80d8

Prune #includes, moving the module interface to the front.  Note that this
exposed the fact that the header was not self-contained.  There is a reason
we do things :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12481 91177308-0d34-0410-b5e6-96231b3b80d8

Make this header file self-contained

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12480 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a HORRIBLY NASTY bug that caused siod to stop working last night.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12479 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a tag-o

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12477 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrite the second on AnalysisUsage usage.  This documents the new
addRequiredTransitive member that Misha added, and explains the whole
concept a lot better.  Also, the document used incorrect "subsubsection"
tags instead of "doc_subsubsection" which this fixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12474 91177308-0d34-0410-b5e6-96231b3b80d8

When loop extraction succeeds, make sure to map the function pointers over
to avoid dangling references.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12470 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an inverted condition that causes us to think that loop extraction
accomplished something when it really did not.  This does not fix the bigger problem tho.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12469 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an iterator invalidation problem in a "buggy" pass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12468 91177308-0d34-0410-b5e6-96231b3b80d8