Constant fold sqrtf

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28853 91177308-0d34-0410-b5e6-96231b3b80d8

Just a minor tweak so you can run things like:

TestRunner.sh Foo/Bar/test.ll

instead of just

TestRunner.sh test.ll

which is convenient in conjunction with find.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28852 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28851 91177308-0d34-0410-b5e6-96231b3b80d8

Fix IndVarsSimplify/2006-06-16-Indvar-LCSSA-Crash.ll, a case where a
"LCSSA" phi node causes indvars to break dominance properties.  This fixes
causes indvars to avoid inserting aggressive code in this case, instead
indvars should be fixed to be more aggressive in the face of lcssa phi's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28850 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase that crashes indvars

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28849 91177308-0d34-0410-b5e6-96231b3b80d8

A new entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28848 91177308-0d34-0410-b5e6-96231b3b80d8

Implement the getPointerRegClass method, which is required for the ptr_rc
magic to work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28847 91177308-0d34-0410-b5e6-96231b3b80d8

Forgot operands were hard coded for compile unit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28846 91177308-0d34-0410-b5e6-96231b3b80d8

Include information of svn repository for llvm-gcc4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28845 91177308-0d34-0410-b5e6-96231b3b80d8

Include information about svn repository for llvm-gcc4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28844 91177308-0d34-0410-b5e6-96231b3b80d8

Later models likely to have Yonah like attributes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28843 91177308-0d34-0410-b5e6-96231b3b80d8

Do not hardcode random paths into the makefile.  Make the user specify the
SDK to use when building "universal" on Mac OS/X, if they want to use a
specific one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28842 91177308-0d34-0410-b5e6-96231b3b80d8

Upgrade some load/store instructions to use the proper addressing mode stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28841 91177308-0d34-0410-b5e6-96231b3b80d8

In 64-bit mode, addr mode operands use G8RC instead of GPRC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28840 91177308-0d34-0410-b5e6-96231b3b80d8

fix some assumptions that pointers can only be 32-bits.  With this, we can
now compile:

static unsigned long X;
void test1() {
  X = 0;
}

into:

_test1:
        lis r2, ha16(_X)
        li r3, 0
        stw r3, lo16(_X)(r2)
        blr

Totally amazing :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28839 91177308-0d34-0410-b5e6-96231b3b80d8

Split 64-bit instructions out into a separate .td file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28838 91177308-0d34-0410-b5e6-96231b3b80d8

Force 64-bit register availability in 64-bit mode.  For real.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28837 91177308-0d34-0410-b5e6-96231b3b80d8

Document known xcode 2.3 issue

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28836 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the -darwin and -aix llc options, inferring darwinism and aixism from
the target triple & subtarget info.  woo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28835 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing casts. This fixed some regressions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28834 91177308-0d34-0410-b5e6-96231b3b80d8

Improve a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28833 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify TargetData ctor call

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28832 91177308-0d34-0410-b5e6-96231b3b80d8

Don't pass target name into TargetData anymore, it is never used or needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28831 91177308-0d34-0410-b5e6-96231b3b80d8

Don't pass target name into TargetData anymore, it is never used or needed.
Remove explicit casts to std::string now that there is no overload resolution
issues in the TargetData ctors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28830 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify the targetdata ctor by not passing in a "targetname" which is always
ignored.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28829 91177308-0d34-0410-b5e6-96231b3b80d8

Remove ctor with each piece specifyable (which causes overload ambiguities),
add a new init method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28828 91177308-0d34-0410-b5e6-96231b3b80d8

Only count instructions as code size, not constant pools and other per-function stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28827 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify interpreter construction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28826 91177308-0d34-0410-b5e6-96231b3b80d8

Document the subtarget features better, make sure that 64-bit mode, 64-bit
support, and 64-bit register use are all consistent with each other.

Add a new "IsPPC" feature, to distinguish ppc32 vs ppc64 targets, use this
to configure TargetData differently.  This not makes ppc64 blow up on lots
of stuff :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28825 91177308-0d34-0410-b5e6-96231b3b80d8

Rename some subtarget features.  A CPU now can *have* 64-bit instructions,
can in 32-bit mode we can choose to optionally *use* 64-bit registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28824 91177308-0d34-0410-b5e6-96231b3b80d8

apple's compiler works too

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28823 91177308-0d34-0410-b5e6-96231b3b80d8

PR# not associated with XFAIL

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28822 91177308-0d34-0410-b5e6-96231b3b80d8

debug info is alive again

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28821 91177308-0d34-0410-b5e6-96231b3b80d8

This test isn't implemented yet

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28820 91177308-0d34-0410-b5e6-96231b3b80d8

Add a error message to cbu to match bu

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28819 91177308-0d34-0410-b5e6-96231b3b80d8

move header

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28818 91177308-0d34-0410-b5e6-96231b3b80d8

add decimal form of LLVMDebugVersion

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28817 91177308-0d34-0410-b5e6-96231b3b80d8

1. Revise vector debug support.

2. Update docs for vector debug support and new version control.

3. Simplify serialization of DebugDescInfo subclasses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28816 91177308-0d34-0410-b5e6-96231b3b80d8

More libcall transformations:
printf("%s\n", str) -> puts(str)
printf("%c", c) -> putchar(c)
Also fixed fprintf(file, "%c", c) -> fputc(c, file)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28815 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify fprintf(file, "%s", str) to fputs(str, file).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28814 91177308-0d34-0410-b5e6-96231b3b80d8

First baby step towards ppc64 support.  This adds a new -march=ppc64 backend
that is currently just like ppc32 :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28813 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Regression/Linker/2006-06-15-GlobalVarAnment.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28812 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase, the linker is not merging alignments right.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28811 91177308-0d34-0410-b5e6-96231b3b80d8

Only print the stack trace if it was requested. Previously, any call into
the Signals module that registered the handlers would cause the stack trace
to be generated. Now, you must explicitly call PrintStackTraceOnErrorSignal
in order for that to happen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28810 91177308-0d34-0410-b5e6-96231b3b80d8

Teach the local allocator to know that live-in values (e.g. arguments) are
live at function entry.  This prevents it from using arg registers for other
purposes before the arguments are used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28809 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note that Nate noticed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28808 91177308-0d34-0410-b5e6-96231b3b80d8

1. Support standard dwarf format (was bootstrapping in Apple format.)

2. Add vector support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28807 91177308-0d34-0410-b5e6-96231b3b80d8

Was pointed out that structure alignment and type alignment are not the same
thing.  Doubles still need to be special cased.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28806 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure global doubles and vectors are aligned properly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28805 91177308-0d34-0410-b5e6-96231b3b80d8

Add some more matcher classes for shifts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28804 91177308-0d34-0410-b5e6-96231b3b80d8

Implement Transforms/InstCombine/bswap.ll, turning common shift/and/or bswap
idioms into bswap intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28803 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase, instcombine should turn these into llvm.bswap intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28802 91177308-0d34-0410-b5e6-96231b3b80d8

Fix building on case-sensitive file systems, grr :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28801 91177308-0d34-0410-b5e6-96231b3b80d8

Actually add instructions to the list of defined values so it gets
recognized as such!  This prevents the CppWriter from treating every
operand as a forward reference and making a mess of the output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28800 91177308-0d34-0410-b5e6-96231b3b80d8

Alignment of globals has not been quite right.  Needed to drop the pointer type
to get the alignment of the element type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28799 91177308-0d34-0410-b5e6-96231b3b80d8

Vector extract / insert index operand should have ptr type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28798 91177308-0d34-0410-b5e6-96231b3b80d8

Type of extract_element index operand should be iPTR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28797 91177308-0d34-0410-b5e6-96231b3b80d8

Type of vector extract / insert index operand should be iPTR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28796 91177308-0d34-0410-b5e6-96231b3b80d8

Consistency. EXTRACT_ELEMENT index operand should have ptr type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28795 91177308-0d34-0410-b5e6-96231b3b80d8

Assert. Rather than silently stop printing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28794 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid undesirable behavior when assert is not enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28793 91177308-0d34-0410-b5e6-96231b3b80d8

Instructions with variable operands (variable_ops) can have a number required
operands. e.g.
def CALL32r : I<0xFF, MRM2r, (ops GR32:$dst, variable_ops),
                "call {*}$dst", [(X86call GR32:$dst)]>;
TableGen should emit operand informations for the "required" operands.

Added a target instruction info flag M_VARIABLE_OPS to indicate the target
instruction may have more operands in addition to the minimum required
operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28791 91177308-0d34-0410-b5e6-96231b3b80d8

Allow more use of iPTR in patterns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28790 91177308-0d34-0410-b5e6-96231b3b80d8

X86 call instructions can take variable number of operands. Parameters of
vector types are passed via XMM registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28789 91177308-0d34-0410-b5e6-96231b3b80d8

Added support for variable_ops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28788 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28787 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase, not currently working.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28786 91177308-0d34-0410-b5e6-96231b3b80d8

Fix support for optional input flag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28784 91177308-0d34-0410-b5e6-96231b3b80d8

Add argument registers to the end of call operand list (partial fix).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28783 91177308-0d34-0410-b5e6-96231b3b80d8

Change versioning to per debug info descriptor (merged with tag.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28782 91177308-0d34-0410-b5e6-96231b3b80d8

Place dwarf headers at earliest possible point.  Well behaved when skipping
functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28781 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Transforms/LoopUnswitch/2006-06-13-SingleEntryPHI.ll, a loop unswitch
bug exposed by the recent lcssa work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28779 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase, distilled from povray

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28778 91177308-0d34-0410-b5e6-96231b3b80d8

When asked not to delete useless PHIs, really don't delete them, no matter how
redundant they are.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28777 91177308-0d34-0410-b5e6-96231b3b80d8

Use the PotDoms map to memoize 'dominating value' lookup.  With this patch,
LCSSA is still the slowest pass when gccas'ing 252.eon, but now it only takes
39s instead of 289s. :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28776 91177308-0d34-0410-b5e6-96231b3b80d8

getOperandNum(): error if specified operand number is out of range.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28775 91177308-0d34-0410-b5e6-96231b3b80d8

Fix another instance where PHI nodes need special treatment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28774 91177308-0d34-0410-b5e6-96231b3b80d8

Update isLCSSAForm to handle PHI nodes specially for live-out detection.  This
is the same as the recent patch to LCSSA.cpp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28773 91177308-0d34-0410-b5e6-96231b3b80d8

I am sure I had commited this workaround before.  Perhaps soon I should sort it all out

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28772 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug that was causing major slowdowns in povray.  This was due to LCSSA
not handling PHI nodes correctly when determining if a value was live-out.

This patch reduces the number of detected live-out variables in the testcase
from 6565 to 485.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28771 91177308-0d34-0410-b5e6-96231b3b80d8

It really helps to be returning to the correct place

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28769 91177308-0d34-0410-b5e6-96231b3b80d8

Port some bugfixes in shift handling from SimplifyDemandedBits over to
ComputeMaskedBits.  DemandedMasks and KnownZero/One masks should never have
bits set out of the range of the base datatype.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28768 91177308-0d34-0410-b5e6-96231b3b80d8

My original test case was bogus.  Reverting to crasher case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28767 91177308-0d34-0410-b5e6-96231b3b80d8

Tests to make sure that ComputeMaskedBits discards bits for shifts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28766 91177308-0d34-0410-b5e6-96231b3b80d8

TargetLowering::ComputeMaskedBits was not clearing reciprocal bits on shifts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28765 91177308-0d34-0410-b5e6-96231b3b80d8

Cygwin support: use _alloca to allocate stack if > 4k. Patch by Anton Korobeynikov.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28764 91177308-0d34-0410-b5e6-96231b3b80d8

Teach bugpoint to kill optimization passes that run over the timeout limit,
which allows it to debug optimizer infinite loops.  This patch is contributed
by Nick Lewycky, thanks!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28763 91177308-0d34-0410-b5e6-96231b3b80d8

Decribe the "implementation" keyword.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28762 91177308-0d34-0410-b5e6-96231b3b80d8

Gaar!  Don't use r11 for CR save/restore, use R0.  R11 can be register
allocated, thus live across the save/reload.  This fixes

llc-beta /MultiSource/Applications/spiff/spiff
llc-beta /MultiSource/Benchmarks/sim/sim:
llc-beta /MultiSource/Benchmarks/Ptrdist/bc/bc
llc-beta /MultiSource/Benchmarks/McCat/12-IOtest/iotest:
llc-beta /MultiSource/Benchmarks/FreeBench/fourinarow/fourinarow
llc-beta /MultiSource/Benchmarks/Fhourstones-3.1/fhourstones3.1
llc-beta /MultiSource/Benchmarks/mediabench/adpcm/rawdaudio/rawdaudio
llc-beta /MultiSource/Benchmarks/mediabench/adpcm/rawcaudio/rawcaudio
llc-beta /MultiSource/Benchmarks/mediabench/g721/g721encode/encode
llc-beta /MultiSource/Benchmarks/mediabench/jpeg/jpeg-6a/cjpeg

and probably others, with -regalloc=local.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28761 91177308-0d34-0410-b5e6-96231b3b80d8

Fix spilling and reloading of CR regs to reload the right values.  This fixes
Olden/power (and probably others) with -regalloc=local.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28760 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply my 6/9 changes.  The bug Evan saw no longer occurs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28759 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an infinite loop on Transforms/SimplifyCFG/2006-06-12-InfLoop.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28758 91177308-0d34-0410-b5e6-96231b3b80d8

TEstcase that makes simplifycfg infloop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28757 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new -compileflags options, remove old externals location specifiers.
People should just use -with-externals.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28756 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure to update the CFG correctly if a switch only has a default dest.
This fixes CodeGen/Generic/2006-06-12-LowerSwitchCrash.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28755 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase that crashes llc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28754 91177308-0d34-0410-b5e6-96231b3b80d8

Let the alpha breakage begin.  First Formals and RET.  next Calls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28753 91177308-0d34-0410-b5e6-96231b3b80d8

Start on my todo list

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28752 91177308-0d34-0410-b5e6-96231b3b80d8

lower more then 4 formal arguments. The offset is currently hard coded.
implement SelectFrameIndex

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28751 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for 2006-06-26-MultipleExitsSingleBlock.

If a single exit block has multiple predecessors within the loop, it will
appear in the exit blocks list more than once.  LCSSA needs to take that into
account so that it doesn't double process that exit block.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28750 91177308-0d34-0410-b5e6-96231b3b80d8