New testcase.  Believe it or not, this happens a LOT in vortex

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17745 91177308-0d34-0410-b5e6-96231b3b80d8

Fix build on Linux/PowerPC64 using SuSE GCC (#undef PPC)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17744 91177308-0d34-0410-b5e6-96231b3b80d8

Moved from include/llvm/Support/Linker.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17743 91177308-0d34-0410-b5e6-96231b3b80d8

Moved to lib/Bytecode/Archive in preparation for re-write.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17742 91177308-0d34-0410-b5e6-96231b3b80d8

This optimization makes MANY phi nodes that all have the same incoming value.
If this happens, detect it early instead of relying on instcombine to notice
it later.  This can be a big speedup, because PHI nodes can have many
incoming values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17741 91177308-0d34-0410-b5e6-96231b3b80d8

Implement instcombine/phi.ll:test6 - pulling operations through PHI nodes.

This exposes subsequent optimization possiblities and reduces code size.
This triggers 1423 times in spec.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17740 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase: the phi can be eliminated if the casts are sucked into it.
Note that this reduces code size anyway (as well as making further
optimizations simpler) so it's always a win.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17739 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17738 91177308-0d34-0410-b5e6-96231b3b80d8

add a testcase, which we already handle

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17737 91177308-0d34-0410-b5e6-96231b3b80d8

Transform this:
  %X = alloca ...
  %Y = alloca ...
    X == Y

into false.  This allows us to simplify some stuff in eon (and probably
many other C++ programs) where operator= was checking for self assignment.
Folding this allows us to SROA several additional structs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17735 91177308-0d34-0410-b5e6-96231b3b80d8

Remove note to self

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17734 91177308-0d34-0410-b5e6-96231b3b80d8

Fix problem with insertion point for ADJCALLSTACKDOWN.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17733 91177308-0d34-0410-b5e6-96231b3b80d8

Update lists of failing unit tests.
Exclude bigfib, so that we effectively exclude all C++ benchmarks.
Update to-do list: mention va_start.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17732 91177308-0d34-0410-b5e6-96231b3b80d8

Oops, make this test the right thing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17731 91177308-0d34-0410-b5e6-96231b3b80d8

If a function always returns a constant, replace all calls sites with that
constant value.  This makes the return value dead and allows for
simplification in the caller.

This implements IPConstantProp/return-constant.ll

This triggers several dozen times throughout SPEC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17730 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17729 91177308-0d34-0410-b5e6-96231b3b80d8

Fix NotTest - round up extraStack to the nearest doubleword, if it is
not zero.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17728 91177308-0d34-0410-b5e6-96231b3b80d8

Teach SROA how to promote an array index that is variable, if the dimension
of the array is just two.  This occurs 8 times in gcc, 6 times in crafty, and
12 times in 099.go.

This implements ScalarRepl/sroa_two.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17727 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase, SROA with variable array index

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17726 91177308-0d34-0410-b5e6-96231b3b80d8

Update failing Benchmarks; point out that I'm skipping Shootout-C++.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17725 91177308-0d34-0410-b5e6-96231b3b80d8

Rearrange some code, no functionality changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17724 91177308-0d34-0410-b5e6-96231b3b80d8

Update expected UnitTests failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17723 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrite outgoing arg handling to handle more weird corner cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17722 91177308-0d34-0410-b5e6-96231b3b80d8

Support UndefValue emission.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17721 91177308-0d34-0410-b5e6-96231b3b80d8

setting path to prcontext.tcl script. Right now it searches for tclsh in your path, but this should be obtained from configure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17720 91177308-0d34-0410-b5e6-96231b3b80d8

Remove debugging code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17719 91177308-0d34-0410-b5e6-96231b3b80d8

Argument promotion transforms functions to unconditionally load their
argument pointers.  This is only valid to do if the function already
unconditionally loaded an argument or if the pointer passed in is known
to be valid.  Make sure to do the required checks.

This fixed ArgumentPromotion/control-flow.ll and the Burg program.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17718 91177308-0d34-0410-b5e6-96231b3b80d8

Add a testcase we should continue to argpromote

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17717 91177308-0d34-0410-b5e6-96231b3b80d8

Add a testcase for a function we cannot legally promote the argument of.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17716 91177308-0d34-0410-b5e6-96231b3b80d8

Add useful method, minor cleanups

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17715 91177308-0d34-0410-b5e6-96231b3b80d8

Don't print unneeded labels

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17714 91177308-0d34-0410-b5e6-96231b3b80d8

Run prcontext.tcl with tclsh and let it be found in the path. This should be found by configure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17713 91177308-0d34-0410-b5e6-96231b3b80d8

Adding subdirectory dg.exp files in order to be able to use dejagnu to only run specific tests (located in some subdirectory of Regression)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17712 91177308-0d34-0410-b5e6-96231b3b80d8

Changed to use tcl script.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17711 91177308-0d34-0410-b5e6-96231b3b80d8

Hack around stupidity in GCC, fixing Burg with the CBE and
CBackend/2004-11-13-FunctionPointerCast.llx

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17710 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17709 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrote prcontext.py in tcl.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17708 91177308-0d34-0410-b5e6-96231b3b80d8

shld is a very high latency operation. Instead of emitting it for shifts of
two or three, open code the equivalent operation which is faster on athlon
and P4 (by a substantial margin).

For example, instead of compiling this:

long long X2(long long Y) { return Y << 2; }

to:

X3_2:
        movl 4(%esp), %eax
        movl 8(%esp), %edx
        shldl $2, %eax, %edx
        shll $2, %eax
        ret

Compile it to:

X2:
        movl 4(%esp), %eax
        movl 8(%esp), %ecx
        movl %eax, %edx
        shrl $30, %edx
        leal (%edx,%ecx,4), %edx
        shll $2, %eax
        ret

Likewise, for << 3, compile to:

X3:
        movl 4(%esp), %eax
        movl 8(%esp), %ecx
        movl %eax, %edx
        shrl $29, %edx
        leal (%edx,%ecx,8), %edx
        shll $3, %eax
        ret

This matches icc, except that icc open codes the shifts as adds on the P4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17707 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing check

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17706 91177308-0d34-0410-b5e6-96231b3b80d8

Compile:

long long X3_2(long long Y) { return Y+Y; }
int X(int Y) { return Y+Y; }

into:

X3_2:
        movl 4(%esp), %eax
        movl 8(%esp), %edx
        addl %eax, %eax
        adcl %edx, %edx
        ret
X:
        movl 4(%esp), %eax
        addl %eax, %eax
        ret

instead of:

X3_2:
        movl 4(%esp), %eax
        movl 8(%esp), %edx
        shldl $1, %eax, %edx
        shll $1, %eax
        ret

X:
        movl 4(%esp), %eax
        shll $1, %eax
        ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17705 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify handling of shifts to be the same as we do for adds.  Add support
for (X * C1) + (X * C2) (where * can be mul or shl), allowing us to fold:

   Y+Y+Y+Y+Y+Y+Y+Y

into
         %tmp.8 = shl long %Y, ubyte 3           ; <long> [#uses=1]

instead of

        %tmp.4 = shl long %Y, ubyte 2           ; <long> [#uses=1]
        %tmp.12 = shl long %Y, ubyte 2          ; <long> [#uses=1]
        %tmp.8 = add long %tmp.4, %tmp.12               ; <long> [#uses=1]

This implements add.ll:test25

Also add support for (X*C1)-(X*C2) -> X*(C1-C2), implementing sub.ll:test18

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17704 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17703 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for shifts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17702 91177308-0d34-0410-b5e6-96231b3b80d8

Fold:
   (X + (X << C2)) --> X * ((1 << C2) + 1)
   ((X << C2) + X) --> X * ((1 << C2) + 1)

This means that we now canonicalize "Y+Y+Y" into:

        %tmp.2 = mul long %Y, 3         ; <long> [#uses=1]

instead of:

        %tmp.10 = shl long %Y, ubyte 1          ; <long> [#uses=1]
        %tmp.6 = add long %Y, %tmp.10               ; <long> [#uses=1]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17701 91177308-0d34-0410-b5e6-96231b3b80d8

Lazily create the abort message, so only translation units that use unwind
will actually get it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17700 91177308-0d34-0410-b5e6-96231b3b80d8

Fix: CodeExtractor/2004-11-12-InvokeExtract.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17699 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17698 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug where the code extractor would get a bit confused handling invoke
instructions, setting DefBlock to a block it did not have dom info for.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17697 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify handling of constant initializers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17696 91177308-0d34-0410-b5e6-96231b3b80d8

Makefile for lib/Linker

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17695 91177308-0d34-0410-b5e6-96231b3b80d8

This file originated in lib/VMCore/Linker.cpp but now lives in
lib/Linker/LinkModules.cpp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17694 91177308-0d34-0410-b5e6-96231b3b80d8

This file originated in tools/gccld/Linker.cpp but now lives in
lib/Linker/LinkArchives.cpp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17693 91177308-0d34-0410-b5e6-96231b3b80d8

* Clean up all the shared library output on uninstall
* Provide the correct set of input directories to the TAGS target
* Provide a CTAGS target for building Vi style ctags files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17688 91177308-0d34-0410-b5e6-96231b3b80d8

Document the new llvm-ranlib command.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17687 91177308-0d34-0410-b5e6-96231b3b80d8

Correctly terminate a list.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17686 91177308-0d34-0410-b5e6-96231b3b80d8

Document the modifiers and the file format.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17685 91177308-0d34-0410-b5e6-96231b3b80d8

Make this build in release mode

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17684 91177308-0d34-0410-b5e6-96231b3b80d8

Add llvm-ar to the index.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17682 91177308-0d34-0410-b5e6-96231b3b80d8

First attempt at llvm-ar documentation. Modifiers need a little more
explanation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17681 91177308-0d34-0410-b5e6-96231b3b80d8

Actually, leave the check in.  This prevents us from counting dead arguments
as IPCP opportunities.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17680 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bug: IPConstantProp/deadarg.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17679 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17678 91177308-0d34-0410-b5e6-96231b3b80d8

Fix documentation for Makefile target name change. install-bytecode is now
just "install" in the runtime directory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17677 91177308-0d34-0410-b5e6-96231b3b80d8

Make IP Constant prop more aggressive about handling self recursive calls.
This implements IPConstantProp/recursion.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17666 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17665 91177308-0d34-0410-b5e6-96231b3b80d8

Correct the name of stosd for the AT&T syntax:
It's stosl (l for long == 32 bit).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17658 91177308-0d34-0410-b5e6-96231b3b80d8

Do not let dead constant expressions hanging off of functions prevent IPCP.
This allows to elimination of a bunch of global pool descriptor args from
programs being pool allocated (and is also generally useful!)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17657 91177308-0d34-0410-b5e6-96231b3b80d8

Provide conversion from posix time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17656 91177308-0d34-0410-b5e6-96231b3b80d8

Fix isBytecodeFile to correctly recognized compressed bytecode too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17655 91177308-0d34-0410-b5e6-96231b3b80d8

* Implement getStatusInfo for getting stat(2) like information
* Implement createTemporaryFile for mkstemp(3) functionality
* Fix isBytecodeFile to accept llvc magic # (compressed) as bytecode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17654 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure llee can deal with compressed bytecode too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17652 91177308-0d34-0410-b5e6-96231b3b80d8

Recognize compressed LLVM bytecode files.
This should fix the problem of not being able to link compressed LLVM
bytecode files from LLVM libraries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17648 91177308-0d34-0410-b5e6-96231b3b80d8

Tune compression:
bzip2: block size 9 -> 5, reduces memory by 400Kbytes, doesn't affect speed
       or compression ratio on all but the largest bytecode files (>1MB)
zip:   level 9 -> 6, this speeds up compression time by ~30% but only
       degrades the compressed size by a few bytes per megabyte. Those few
       bytes aren't worth the effort.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17647 91177308-0d34-0410-b5e6-96231b3b80d8

Change this back so that I get stable numbers to reflect the change from the
nightly testers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17646 91177308-0d34-0410-b5e6-96231b3b80d8

Document quick-test target.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17644 91177308-0d34-0410-b5e6-96231b3b80d8

Add a quick-test target that uses QUICKTEST variable to quickly run a
portion of the test suite. e.g.:

make quck-test QUICKTEST=Regression/Bytecode

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17643 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bug: 2004-11-08-FreeUseCrash.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17642 91177308-0d34-0410-b5e6-96231b3b80d8

Name file properly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17641 91177308-0d34-0410-b5e6-96231b3b80d8

Hrm, don't ask how I ran into this bug

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17640 91177308-0d34-0410-b5e6-96231b3b80d8

* Convert tabs to spaces
* Order #includes according to style guide
* Remove extraneous blank lines

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17639 91177308-0d34-0410-b5e6-96231b3b80d8

Output the program name (in this case, gccld) with warning about invalid files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17638 91177308-0d34-0410-b5e6-96231b3b80d8

Allow hbd to be bugpointable on darwin by fixing common and linkonce codegen

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17637 91177308-0d34-0410-b5e6-96231b3b80d8

Handle headers for compressed bytecode files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17634 91177308-0d34-0410-b5e6-96231b3b80d8

Don't silently ignore invalid files: tell the user!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17633 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug that was preventing povray and namd from pool allocating correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17632 91177308-0d34-0410-b5e6-96231b3b80d8

Handle assert_fail special

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17631 91177308-0d34-0410-b5e6-96231b3b80d8

Remove extra spaces

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17629 91177308-0d34-0410-b5e6-96231b3b80d8

This file is gone

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17627 91177308-0d34-0410-b5e6-96231b3b80d8

Add a --disable-compression option to gccas so the default compression of
bytecode can be defeated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17626 91177308-0d34-0410-b5e6-96231b3b80d8

Make writing compressed bytecode the default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17625 91177308-0d34-0410-b5e6-96231b3b80d8

Implement and document the TOOL_VERBOSE option that asks each tool invoked
to be verbose about its actions too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17624 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17623 91177308-0d34-0410-b5e6-96231b3b80d8

Document compressed bytecode details.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17622 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo: isvalid -> isValid

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17621 91177308-0d34-0410-b5e6-96231b3b80d8

Move path override in test/Makefile.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17613 91177308-0d34-0410-b5e6-96231b3b80d8

Override PATH to submake so that all tools are found if the PATH is
not set up properly (it also avoids using different tools that happen
to be in the path).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17612 91177308-0d34-0410-b5e6-96231b3b80d8

* Add support for f2c and the -f2c switch to enable Fortran benchmarks
* Remove spurious spaces between variable names and `=' (they're not lined up
  anyway and there's no hope of doing that)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17611 91177308-0d34-0410-b5e6-96231b3b80d8

Now using path to python found by configure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17610 91177308-0d34-0410-b5e6-96231b3b80d8

First draft of compendium of documents relevant to writers of LLVM machine
target backends

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17609 91177308-0d34-0410-b5e6-96231b3b80d8

Put int the getReg cast optimization from x86 so that we generate fewer
move instructions for the register allocator to coalesce.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@17608 91177308-0d34-0410-b5e6-96231b3b80d8