Fix for PR960. Improves bugpoint so that it removes global variable
initializers as well. This is only a first pass. It can be slow because
it clones the module for each pass. An obvious improvement is not to do that.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31182 91177308-0d34-0410-b5e6-96231b3b80d8

Supply alignment info to linker through LLVMSymbol.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31181 91177308-0d34-0410-b5e6-96231b3b80d8

turn off tail merging for now

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31180 91177308-0d34-0410-b5e6-96231b3b80d8

be more aggressive about matching identical instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31179 91177308-0d34-0410-b5e6-96231b3b80d8

Allow this to pass on ppc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31178 91177308-0d34-0410-b5e6-96231b3b80d8

Modernize the testcase, allow it to pass on ppc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31177 91177308-0d34-0410-b5e6-96231b3b80d8

TargetData is not subclassed. So no need to have virtual method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31173 91177308-0d34-0410-b5e6-96231b3b80d8

Targets should depend on all the intrinsics.td files also, otherwise they
will compute a locally wrong numbering for the intrinsics.  This fixes a
nasty issue where the x86 backend started miscompiling stuff in a 'cvs up'd
build after the altivec intrinsics were added.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31172 91177308-0d34-0410-b5e6-96231b3b80d8

Move getPreferredAlignmentLog from AsmPrinter to TargetData

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31171 91177308-0d34-0410-b5e6-96231b3b80d8

expand ISD::VACOPY

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31170 91177308-0d34-0410-b5e6-96231b3b80d8

Removed extraneous semi-colon; this was prevening the grammar file from
bison'ing correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31169 91177308-0d34-0410-b5e6-96231b3b80d8

Instead of hard coding global prefix, use TargetAsmInfo.
Add LTO destructor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31168 91177308-0d34-0410-b5e6-96231b3b80d8

visitSwitchCase knows how to insert conditional branches well.  Change
visitBr to just call visitSwitchCase, eliminating duplicate logic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31167 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize CaseBlock a bit more:

Rename LHSBB/RHSBB to TrueBB/FalseBB.  Allow the RHS value to be null,
in which case the LHS is treated as a bool.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31166 91177308-0d34-0410-b5e6-96231b3b80d8

update comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31165 91177308-0d34-0410-b5e6-96231b3b80d8

Fix CodeGen/IA64/ret-0.ll, which has apparently been broken since some of the
isel changes happened months ago.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31164 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31163 91177308-0d34-0410-b5e6-96231b3b80d8

fix warning about missing newline at end of file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31162 91177308-0d34-0410-b5e6-96231b3b80d8

generalize 'CaseBlock'.  It really allows any comparison to be inserted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31161 91177308-0d34-0410-b5e6-96231b3b80d8

implement uncond branch insertion, mark branches with isBranch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31160 91177308-0d34-0410-b5e6-96231b3b80d8

implement uncond branch insertion for the branch folding pass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31159 91177308-0d34-0410-b5e6-96231b3b80d8

implement uncond branch insertion so alpha works work branchfolding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31158 91177308-0d34-0410-b5e6-96231b3b80d8

implement uncond branch insertion for sparc to fix regressions from last night
due to branchfolding

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31157 91177308-0d34-0410-b5e6-96231b3b80d8

new bad case

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31156 91177308-0d34-0410-b5e6-96231b3b80d8

Don't do dead block elimination in fast mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31155 91177308-0d34-0410-b5e6-96231b3b80d8

fix assert comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31154 91177308-0d34-0410-b5e6-96231b3b80d8

LinearScanner hotspot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31153 91177308-0d34-0410-b5e6-96231b3b80d8

Tighter data structure for deleted debug labels.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31152 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Transforms/ScalarRepl/2006-10-23-PointerUnionCrash.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31151 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase that crashes scalarrepl

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31150 91177308-0d34-0410-b5e6-96231b3b80d8

move single basic blocks that are neither fallen into nor fall out of into
a place more useful.  In particular, if we can put them in a place where code
will be able to fall into it, do so.  Otherwise, put it in a place it can fall
through into a successor.  Otherwise, if preventing a fallthrough, move to the
end of the function, out of the way.

This deletes several hundred unconditional branches from spass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31149 91177308-0d34-0410-b5e6-96231b3b80d8

Add intrinsics for the rest of the DCB* instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31148 91177308-0d34-0410-b5e6-96231b3b80d8

add missing dcb* intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31147 91177308-0d34-0410-b5e6-96231b3b80d8

Revert back to r1.21, which was the last revision of predsimplify that
passes llvm-gcc bootstrap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31146 91177308-0d34-0410-b5e6-96231b3b80d8

add moveBefore/moveAfter helper methods

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31145 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo. Add more comment. Avoid extra hash_map search.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31144 91177308-0d34-0410-b5e6-96231b3b80d8

add two helper methods

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31143 91177308-0d34-0410-b5e6-96231b3b80d8

Add removeModule().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31142 91177308-0d34-0410-b5e6-96231b3b80d8

Enable tail merging by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31140 91177308-0d34-0410-b5e6-96231b3b80d8

Update test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31139 91177308-0d34-0410-b5e6-96231b3b80d8

implement STRB and STRH

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31138 91177308-0d34-0410-b5e6-96231b3b80d8

expand ISD::MEMSET

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31137 91177308-0d34-0410-b5e6-96231b3b80d8

Make these hack flags hidden, like other dsa hack flags

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31136 91177308-0d34-0410-b5e6-96231b3b80d8

Revert the aggressive inlining for now, it doesn't handle external correctly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31135 91177308-0d34-0410-b5e6-96231b3b80d8

change this back

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31134 91177308-0d34-0410-b5e6-96231b3b80d8

restore these tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31133 91177308-0d34-0410-b5e6-96231b3b80d8

Handle fallout from the recent branch-on-undef changes.  This fixes
Prolangs-C/agrep and SCCP/2006-10-23-IPSCCP-Crash.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31132 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31131 91177308-0d34-0410-b5e6-96231b3b80d8

Minor tweak.  Instead of generating:

        movl 32(%esp), %eax
        cmpl $1, %eax
        je LBB1_1       #bb
LBB1_4: #entry
        cmpl $2, %eax
        je LBB1_2       #bb2
        jmp LBB1_3      #UnifiedReturnBlock
LBB1_1: #bb

notice that we would miss the fall through and emit this instead:

        movl 32(%esp), %eax
        cmpl $2, %eax
        je LBB1_2       #bb2
LBB1_4: #entry
        cmpl $1, %eax
        jne LBB1_3      #UnifiedReturnBlock
LBB1_1: #bb

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31130 91177308-0d34-0410-b5e6-96231b3b80d8

More complete solution to deleting blocks and debug info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31129 91177308-0d34-0410-b5e6-96231b3b80d8

[SU]int update - inactive code may get activated someday

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31128 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31127 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the Backwards operation. Resolving now works at the time when a
property is added by running through the list of uses of the value and
adding resolved properties to the property set.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31126 91177308-0d34-0410-b5e6-96231b3b80d8

Fix phi node updating for switches lowered to linear sequences of branches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31125 91177308-0d34-0410-b5e6-96231b3b80d8

disable this code for now, it's not yet safely updating phi nodes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31124 91177308-0d34-0410-b5e6-96231b3b80d8

Fix similar missing optimization opportunity in XOR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31123 91177308-0d34-0410-b5e6-96231b3b80d8

this part implemented.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31122 91177308-0d34-0410-b5e6-96231b3b80d8

Whoops! Add missing NULL check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31121 91177308-0d34-0410-b5e6-96231b3b80d8

Do section switching right way for Mingw\Cygwin and Linux.
This will allow (after llvm-gcc linkage fix) more dummy tests using
libstdc++ to pass on mingw :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31120 91177308-0d34-0410-b5e6-96231b3b80d8

Implement PR964 and Regression/CodeGen/Generic/SwitchLowering.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31119 91177308-0d34-0410-b5e6-96231b3b80d8

Handle "if ((x|y) != 0)" for ints like we do for bools. Fixes missed
optimization opportunity pointed out by Chris Lattner.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31118 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for PR964

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31117 91177308-0d34-0410-b5e6-96231b3b80d8

AllocaInst can't return a null pointer. Fixes missed optimization
opportunity pointed out by Andrew Lewycky.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31115 91177308-0d34-0410-b5e6-96231b3b80d8

Add a workaround for PR962, disabling the more aggressive form of this
transformation.  This speeds up a C++ app 2.25x.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31113 91177308-0d34-0410-b5e6-96231b3b80d8

Don't generate a prototype for _setjmp. At least on Linux, this function
has a different prototype than the one #included from <setjmp.h>. This
patch fixes siod and a number of other test cases on Linux that were
failing the CBE because of this _setjmp issue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31112 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31111 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unneeded CHECK_FOR_ERROR invocations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31110 91177308-0d34-0410-b5e6-96231b3b80d8

regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31109 91177308-0d34-0410-b5e6-96231b3b80d8

change keyword to datalayout

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31108 91177308-0d34-0410-b5e6-96231b3b80d8

Change keyword to datalayout

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31107 91177308-0d34-0410-b5e6-96231b3b80d8

3 Changes:
1. Better document what is going on here.
2. Only hack on one branch per iteration, making the results less conservative.
3. Handle the problematic case by marking edges executable instead of by
   playing with value lattice states.  This is far less pessimistic, and fixes
   SCCP/ipsccp-gvar.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31106 91177308-0d34-0410-b5e6-96231b3b80d8

New, wonderful isSuccessor method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31103 91177308-0d34-0410-b5e6-96231b3b80d8

don't break infinite loops

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31102 91177308-0d34-0410-b5e6-96231b3b80d8

Implement support for branch reversal, fix a bug in branch analysis.
This provides stuff like:

        cmpw cr0, r15, r29
        mr r14, r15
-       bge cr0, LBB3_111       ;bb656
-       b LBB3_90       ;bb501
+       blt cr0, LBB3_90        ;bb501
 LBB3_111:      ;bb656
        lwz r18, 68(r1)

which is particularly good for dispatch group formation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31101 91177308-0d34-0410-b5e6-96231b3b80d8

Use branch reversal to do stuff like this:

        call L_strcmp$stub
        testl %eax, %eax
-       jne LBB26_208   #cond_true6020
-       jmp LBB26_227   #bb7119
+       je LBB26_227    #bb7119
 LBB26_208:     #cond_true6020
        movl $l31_str14, 4(%esp)

        testl %eax, %eax
-       jne LBB26_704   #cond_true13042
-       jmp LBB26_713   #bb13151
+       je LBB26_713    #bb13151
 LBB26_704:     #cond_true13042
        movl $_str52, 4(%esp)

        cmpl 76(%ecx), %eax
-       jge LBB26_1628  #cond_false63.i.i
-       jmp LBB26_1769  #_Z8makeGridP13mrSurfaceListidiidd.exit.i
+       jl LBB26_1769   #_Z8makeGridP13mrSurfaceListidiidd.exit.i
 LBB26_1628:    #cond_false63.i.i
        movl $0, 48964(%esp)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31100 91177308-0d34-0410-b5e6-96231b3b80d8

Implement support for branch condition reversal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31099 91177308-0d34-0410-b5e6-96231b3b80d8

Transform code like:

  jle FOO
  jmp BAR
BAR:

into:

  jle FOO
BAR:

... whoa!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31098 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify code, no functionality change

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31097 91177308-0d34-0410-b5e6-96231b3b80d8

implement support for inserting a cond branch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31096 91177308-0d34-0410-b5e6-96231b3b80d8

allow insertion of a conditional branch with fall-through

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31095 91177308-0d34-0410-b5e6-96231b3b80d8

Three changes:
1. Remove a bunch of ifdef'd code.
2. When a block just contains an uncond branch, change all blocks branching
   to it to jump to the destination instead.
3. If branch analysis tells us some edges in the machinecfg are not actually
   possible, remove them.

#2 triggers a suprisingly large number of times.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31094 91177308-0d34-0410-b5e6-96231b3b80d8

update assert message

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31093 91177308-0d34-0410-b5e6-96231b3b80d8

Add an experimental cross-jumping implementation.
This is currently disabled by default and limited in several ways, but does
have a positive effect.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31090 91177308-0d34-0410-b5e6-96231b3b80d8

typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31089 91177308-0d34-0410-b5e6-96231b3b80d8

implement MachineOperand::isIdenticalTo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31088 91177308-0d34-0410-b5e6-96231b3b80d8

add isIdenticalTo method to machineinstr/operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31087 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-extract should remove module-level asm

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31086 91177308-0d34-0410-b5e6-96231b3b80d8

bugfix

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31074 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an ugly problem in SCCP.  This fixes Benchmarks/Misc-C++/mandel-text.cpp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31073 91177308-0d34-0410-b5e6-96231b3b80d8

Fix miscompilation of MallocBench/espresso which code review pointed out
but apparently didn't make it into the final patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31070 91177308-0d34-0410-b5e6-96231b3b80d8

Changed the pattern matching so that it will hopefully work on Linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31069 91177308-0d34-0410-b5e6-96231b3b80d8

Test case for
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20061016/038780.html

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31068 91177308-0d34-0410-b5e6-96231b3b80d8

Make flag and chain edges visually distinguishable from value edges in DOT
output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31067 91177308-0d34-0410-b5e6-96231b3b80d8

add method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31066 91177308-0d34-0410-b5e6-96231b3b80d8

Implement branch analysis/xform hooks required by the branch folding pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31065 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a comment to be accurate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31064 91177308-0d34-0410-b5e6-96231b3b80d8

For PR950:
This patch implements the first increment for the Signless Types feature.
All changes pertain to removing the ConstantSInt and ConstantUInt classes
in favor of just using ConstantInt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31063 91177308-0d34-0410-b5e6-96231b3b80d8

While creating mask, use 1ULL instead of 1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31062 91177308-0d34-0410-b5e6-96231b3b80d8

Fix SimplifyCFG/2006-10-19-UncondDiv.ll by disabling a bad xform.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31061 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for PR957

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31060 91177308-0d34-0410-b5e6-96231b3b80d8