The inliner strips dead allocas now.  Add a use to allow this test to
test the right thing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30393 91177308-0d34-0410-b5e6-96231b3b80d8

Revert my previous work on ArgumentPromotion.  Further investigation has revealed these
changes to be incorrect.  They just weren't showing up in any of our current testcases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30385 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak the JIT

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30384 91177308-0d34-0410-b5e6-96231b3b80d8

Keep track of the start of MBB's in a separate map from instructions.  This
is faster and is needed for future improvements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30383 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a unnecessary check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30382 91177308-0d34-0410-b5e6-96231b3b80d8

Relax this check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30381 91177308-0d34-0410-b5e6-96231b3b80d8

Undo previous check-in.
Reintroduce recursive assignDFSNumber().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30380 91177308-0d34-0410-b5e6-96231b3b80d8

Fold (X & C1) | (Y & C2)  -> (X|Y) & C3 when possible.
This implements CodeGen/X86/and-or-fold.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30379 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30378 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30377 91177308-0d34-0410-b5e6-96231b3b80d8

Split rotate matching code out to its own function.  Make it stronger, by
matching things like ((x >> c1) & c2) | ((x << c3) & c4) to (rot x, c5) & c6

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30376 91177308-0d34-0410-b5e6-96231b3b80d8

Adding generated files for the last commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30375 91177308-0d34-0410-b5e6-96231b3b80d8

Adding dllimport, dllexport and external weak linkage types.
DLL* linkages got full (I hope) codegeneration support in C & both x86
assembler backends.
External weak linkage added for future use, we don't provide any
codegeneration, etc. support for it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30374 91177308-0d34-0410-b5e6-96231b3b80d8

Use getOffset() instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30327 91177308-0d34-0410-b5e6-96231b3b80d8

Use MachineConstantPoolEntry getOffset() and getType() accessors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30326 91177308-0d34-0410-b5e6-96231b3b80d8

Add MachineConstantPoolEntry getOffset() accessor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30325 91177308-0d34-0410-b5e6-96231b3b80d8

The top bit is used to determine whether it's a MachineConstantPoolValue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30324 91177308-0d34-0410-b5e6-96231b3b80d8

revert accidentally committed file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30323 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead methods, add getNumBlockIDs() method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30322 91177308-0d34-0410-b5e6-96231b3b80d8

Use new config.h macro

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30321 91177308-0d34-0410-b5e6-96231b3b80d8

Add a HAVE_MACH_MACH_H #define to detect the presence of the mach/mach.h
header file on Darwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30319 91177308-0d34-0410-b5e6-96231b3b80d8

add a new (hidden) -disable-core-files option

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30318 91177308-0d34-0410-b5e6-96231b3b80d8

On Mac OS/X, make Process::PreventCoreFiles disable crash reporter for
the process in addition to disabling core file emission.  This speeds up
bugpoint on default-configured macs by several orders of magnitude.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30317 91177308-0d34-0410-b5e6-96231b3b80d8

A MachineConstantPool may have mixed Constant* and MachineConstantPoolValue* values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30316 91177308-0d34-0410-b5e6-96231b3b80d8

Add comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30315 91177308-0d34-0410-b5e6-96231b3b80d8

Mirrors ConstantPoolSDNode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30314 91177308-0d34-0410-b5e6-96231b3b80d8

Document new --enable-valgrind option.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30313 91177308-0d34-0410-b5e6-96231b3b80d8

Fix --enable-valgrind. Add room for the new arguments, and don't keep
a pointer to a temporary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30312 91177308-0d34-0410-b5e6-96231b3b80d8

Add --enable-valgrind option to run optimizations through valgrind to
pick up on memory errors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30311 91177308-0d34-0410-b5e6-96231b3b80d8

GlobalValue with InternalLinkage may have operands with ExternalLinkage
type. Do not ignore these operands while finding external references.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30310 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid recursion in assignDFSNumber(). Move def from ET-Forest.h
to Dominators.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30309 91177308-0d34-0410-b5e6-96231b3b80d8

add note about switch lowering

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30308 91177308-0d34-0410-b5e6-96231b3b80d8

Second half of the fix for Transforms/Inline/inline_cleanup.ll

This folds unconditional branches that are often produced by code
specialization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30307 91177308-0d34-0410-b5e6-96231b3b80d8

Add some more consistency checks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30305 91177308-0d34-0410-b5e6-96231b3b80d8

Fix unionSets so that it can merge correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30304 91177308-0d34-0410-b5e6-96231b3b80d8

Implement the first half of Transforms/Inline/inline_cleanup.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30303 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30302 91177308-0d34-0410-b5e6-96231b3b80d8

Skip over first operand when determining REX prefix for two-address code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30300 91177308-0d34-0410-b5e6-96231b3b80d8

Erase dead instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30298 91177308-0d34-0410-b5e6-96231b3b80d8

The sense of this branch was backwards

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30296 91177308-0d34-0410-b5e6-96231b3b80d8

Turn X < 0 -> TEST X,X js

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30294 91177308-0d34-0410-b5e6-96231b3b80d8

The sense of this branch was inverted :(

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30293 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a ton of jit failures

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30292 91177308-0d34-0410-b5e6-96231b3b80d8

add shifts to addressing mode 1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30291 91177308-0d34-0410-b5e6-96231b3b80d8

If LSR went through a lot of trouble to put constants (e.g. the addr of a global
in a specific BB, don't undo this!).  This allows us to compile
CodeGen/X86/loop-hoist.ll into:

_foo:
        xorl %eax, %eax
***     movl L_Arr$non_lazy_ptr, %ecx
        movl 4(%esp), %edx
LBB1_1: #cond_true
        movl %eax, (%ecx,%eax,4)
        incl %eax
        cmpl %edx, %eax
        jne LBB1_1      #cond_true
LBB1_2: #return
        ret

instead of:

_foo:
        xorl %eax, %eax
        movl 4(%esp), %ecx
LBB1_1: #cond_true
***     movl L_Arr$non_lazy_ptr, %edx
        movl %eax, (%edx,%eax,4)
        incl %eax
        cmpl %ecx, %eax
        jne LBB1_1      #cond_true
LBB1_2: #return
        ret

This was noticed in 464.h264ref.  This doesn't usually affect PPC,
but strikes X86 all the time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30290 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30289 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a regression in the 32-bit port from the 64-bit port landing.
We now compile CodeGen/X86/lea-2.ll into:

_test:
        movl 4(%esp), %eax
        movl 8(%esp), %ecx
        leal -5(%ecx,%eax,4), %eax
        ret

instead of:

_test:
        movl 4(%esp), %eax
        leal (,%eax,4), %eax
        addl 8(%esp), %eax
        addl $4294967291, %eax
        ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30288 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30287 91177308-0d34-0410-b5e6-96231b3b80d8

new note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30286 91177308-0d34-0410-b5e6-96231b3b80d8

new note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30285 91177308-0d34-0410-b5e6-96231b3b80d8

Compile X << 1 (where X is a long-long) to:

        addl %ecx, %ecx
        adcl %eax, %eax

instead of:

        movl %ecx, %edx
        addl %edx, %edx
        shrl $31, %ecx
        addl %eax, %eax
        orl %ecx, %eax

and to:

        addc r5, r5, r5
        adde r4, r4, r4

instead of:

        slwi r2,r9,1
        srwi r0,r11,31
        slwi r3,r11,1
        or r2,r0,r2

on PPC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30284 91177308-0d34-0410-b5e6-96231b3b80d8

Compile X > -1   -> text X,X; js dest
This implements CodeGen/X86/jump_sign.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30283 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30282 91177308-0d34-0410-b5e6-96231b3b80d8

Initialize DontInternalize.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30281 91177308-0d34-0410-b5e6-96231b3b80d8

Reflects MachineConstantPoolEntry changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30279 91177308-0d34-0410-b5e6-96231b3b80d8

Added support for machine specific constantpool values. These are useful for
representing expressions that can only be resolved at link time, etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30278 91177308-0d34-0410-b5e6-96231b3b80d8

Reflect MachineConstantPoolEntry changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30277 91177308-0d34-0410-b5e6-96231b3b80d8

Correct the URL to the attrib.incl file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30276 91177308-0d34-0410-b5e6-96231b3b80d8

An sinkable instruction may exist with uses, if those uses are in dead blocks.
Handle this.  This fixes PR908 and Transforms/LICM/2006-09-12-DeadUserOfSunkInstr.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30275 91177308-0d34-0410-b5e6-96231b3b80d8

testcase for PR908

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30274 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase that crashes the C++ FE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30273 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30271 91177308-0d34-0410-b5e6-96231b3b80d8

Add ability to remove nodes from DominatorTree, for when a BasicBlock
is being removed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30270 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase noticed from PR906

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30269 91177308-0d34-0410-b5e6-96231b3b80d8

add compilable testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30268 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for PR906

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30267 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR905 and InstCombine/2006-09-11-EmptyStructCrash.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30266 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for PR905

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30265 91177308-0d34-0410-b5e6-96231b3b80d8

implement SRL and MUL

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30262 91177308-0d34-0410-b5e6-96231b3b80d8

add the correct fallback for ARMDAGToDAGISel::SelectAddrMode1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30261 91177308-0d34-0410-b5e6-96231b3b80d8

burg is gone

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30253 91177308-0d34-0410-b5e6-96231b3b80d8

partial implementation of the ARM Addressing Mode 1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30252 91177308-0d34-0410-b5e6-96231b3b80d8

Skip the linear search if the answer is already known.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30251 91177308-0d34-0410-b5e6-96231b3b80d8

No longer used, and confuses maintainers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30248 91177308-0d34-0410-b5e6-96231b3b80d8

Include the UIUC CS department attribution on the main doxygen page.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30247 91177308-0d34-0410-b5e6-96231b3b80d8

call AsmPrinter::doInitialization in ARMAsmPrinter::doInitialization

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30246 91177308-0d34-0410-b5e6-96231b3b80d8

Updates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30245 91177308-0d34-0410-b5e6-96231b3b80d8

Update README file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30244 91177308-0d34-0410-b5e6-96231b3b80d8

Rename a test-not-cmp.ll to cmp-test.ll; add a test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30243 91177308-0d34-0410-b5e6-96231b3b80d8

1) With X86 lowering change, the following can no longer happen since
   the branch's chain is also produced by cmp.
          [ch, r : ld]
             ^ ^
             | |
      [XX]--/   \- [flag : cmp]
       ^             ^
       |             |
       \---[br flag]-

Remove an isel check which prevents loads from being folded into cmp / test
instructions.

2) Whenever possible, delete a selected node to allow more load folding
opportunities. Note not all nodes can be deleted after it has been
selected. Some may have simply morphed; some have not changed at all (e.g.
EntryToken).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30242 91177308-0d34-0410-b5e6-96231b3b80d8

X86ISD::CMP now produces a chain as well as a flag. Make that the chain
operand of a conditional branch to allow load folding into CMP / TEST
instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30241 91177308-0d34-0410-b5e6-96231b3b80d8

Behold, more work on relocations.  Things are looking pretty good now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30240 91177308-0d34-0410-b5e6-96231b3b80d8

Removed unnecessary Mangler creation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30239 91177308-0d34-0410-b5e6-96231b3b80d8

add a testcase for this commit: lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20050606/026673.html

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30238 91177308-0d34-0410-b5e6-96231b3b80d8

Allow tail duplication in more cases, relaxing the previous restriction a
bit.  This fixes Regression/Transforms/TailDup/MergeTest.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30237 91177308-0d34-0410-b5e6-96231b3b80d8

This code was trying too hard.  By eliminating redundant edges in the CFG
due to switch cases going to the same place, it make #pred != #phi entries,
breaking live interval analysis.

This fixes 458.sjeng on x86 with llc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30236 91177308-0d34-0410-b5e6-96231b3b80d8

Replace EquivalenceClasses with a custom-built data structure. Many common
operations (like findProperties) should be faster, at the expense of
unionSets being slower in cases that are rare in practise.

Don't erase a dead Instruction. This fixes a memory corruption issue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30235 91177308-0d34-0410-b5e6-96231b3b80d8

Implement Transforms/InstCombine/hoist_instr.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30234 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30233 91177308-0d34-0410-b5e6-96231b3b80d8

Move ipsccp pass earlier to clean up obvious bogosities

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30232 91177308-0d34-0410-b5e6-96231b3b80d8

Make inlining costs more accurate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30231 91177308-0d34-0410-b5e6-96231b3b80d8

Turn div X, (Cond ? Y : 0) -> div X, Y
This implements select.ll::test18.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30230 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30229 91177308-0d34-0410-b5e6-96231b3b80d8

Missing anchor

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30227 91177308-0d34-0410-b5e6-96231b3b80d8

Add cbe support for powi

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30226 91177308-0d34-0410-b5e6-96231b3b80d8

Implement the fpowi now by lowering to a libcall

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30225 91177308-0d34-0410-b5e6-96231b3b80d8

(indirect) test for the __builtin_powi builtin which we now support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30224 91177308-0d34-0410-b5e6-96231b3b80d8

Implement new fpowi node

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30223 91177308-0d34-0410-b5e6-96231b3b80d8

Allow targets to custom lower expanded BIT_CONVERT's

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30217 91177308-0d34-0410-b5e6-96231b3b80d8

First pass at supporting relocations.  Relocations are written correctly to
the file now, however the relocated address is currently wrong.  Fixing
that will require some deep pondering.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30207 91177308-0d34-0410-b5e6-96231b3b80d8