remove skeleton target

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26239 91177308-0d34-0410-b5e6-96231b3b80d8

Rework the SelectionDAG-based implementations of SimplifyDemandedBits
and ComputeMaskedBits to match the new improved versions in instcombine.
Tested against all of multisource/benchmarks on ppc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26238 91177308-0d34-0410-b5e6-96231b3b80d8

don't build the skeleton target

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26237 91177308-0d34-0410-b5e6-96231b3b80d8

remove support for the skeleton target

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26236 91177308-0d34-0410-b5e6-96231b3b80d8

Change SplitBlock to increment a BasicBlock::iterator, not an Instruction*.  Apparently they do different things :)

This fixes a testcase that nate reduced from spass.

Also included are a couple minor code changes that don't affect the generated
code at all.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26235 91177308-0d34-0410-b5e6-96231b3b80d8

MOVAPSrr and MOVAPDrr instruction format should be MRMSrcReg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26234 91177308-0d34-0410-b5e6-96231b3b80d8

distinguish between objects and register names, now we can have stuff
with names like "f84", "in6" etc etc.

this should fix one or two tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26232 91177308-0d34-0410-b5e6-96231b3b80d8

If the false case is the current basic block, then this is a self loop.
We do not want to emit "Loop: ... brcond Out; br Loop", as it adds an extra
instruction in the loop.  Instead, invert the condition and emit
"Loop: ... br!cond Loop; br Out.

Generalize the fix by moving it from PPCDAGToDAGISel to SelectionDAGLowering.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26231 91177308-0d34-0410-b5e6-96231b3b80d8

Lowering of sdiv X, pow2 was broken, this fixes it.  This patch is written
by Nate, I'm just committing it for him.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26230 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a minor makefile bug with lex/yacc handling that nate noticed.  We don't
want to copy the files when the .cpp file changes, we want to copy them
to the .cvs versions when the .l/.y file change (like the comments even say).
This avoids having bogus changes show up in diffs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26229 91177308-0d34-0410-b5e6-96231b3b80d8

Fix VC++ warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26228 91177308-0d34-0410-b5e6-96231b3b80d8

Visual Studio enters the future of bisoning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26227 91177308-0d34-0410-b5e6-96231b3b80d8

Use movaps / movapd (instead of movss / movsd) to do FR32 / FR64 reg to reg
transfer.

According to the Intel P4 Optimization Manual:

Moves that write a portion of a register can introduce unwanted
dependences. The movsd reg, reg instruction writes only the bottom
64 bits of a register, not to all 128 bits. This introduces a dependence on
the preceding instruction that produces the upper 64 bits (even if those
bits are not longer wanted). The dependence inhibits register renaming,
and thereby reduces parallelism.

Not to mention movaps is shorter than movss.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26226 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bug where we unswitched the wrong way

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26225 91177308-0d34-0410-b5e6-96231b3b80d8

A bit more memset / memcpy optimization.
Turns them into calls to memset / memcpy if 1) buffer(s) are not DWORD aligned,
2) size is not known to be greater or equal to some minimum value (currently 128).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26224 91177308-0d34-0410-b5e6-96231b3b80d8

Implement trivial unswitching for switch stmts.  This allows us to trivial
unswitch this loop on 2 before sweating to unswitch on 1/3.

void test4(int N, int i, int C, int*P, int*Q) {
  int j;
  for (j = 0; j < N; ++j) {
    switch (C) {                // general unswitching.
    default: P[i+j] = 0; break;
    case 1: Q[i+j] = 0; break;
    case 3: P[i+j] = Q[i+j]; break;
    case 2: break;              //  TRIVIAL UNSWITCH on C==2
    }
  }
}

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26223 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26222 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an unused function parameter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26221 91177308-0d34-0410-b5e6-96231b3b80d8

make "trivial" unswitching significantly more general.  It can now handle
this for example:

  for (j = 0; j < N; ++j) {     // trivial unswitch
    if (C)
      P[i+j] = 0;
  }

turning it into the obvious code without bothering to duplicate an empty loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26220 91177308-0d34-0410-b5e6-96231b3b80d8

Turn a memcpy from string constant into a series of stores of constant values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26219 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bunch of alpha regressions.  see bug 709

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26218 91177308-0d34-0410-b5e6-96231b3b80d8

new test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26217 91177308-0d34-0410-b5e6-96231b3b80d8

Should not combine ISD::LOCATIONs until we have scheme to remove from
MachineDebugInfo tables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26216 91177308-0d34-0410-b5e6-96231b3b80d8

Checking the wrong value.  This caused us to emit silly code like
Y = seteq bool X, true
instead of just using X :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26215 91177308-0d34-0410-b5e6-96231b3b80d8

Code sufficiently protected against this test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26213 91177308-0d34-0410-b5e6-96231b3b80d8

reverting previous change, will add support for other compilers later

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26211 91177308-0d34-0410-b5e6-96231b3b80d8

Convert over to the new way of handling lex/bison checked into cvs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26209 91177308-0d34-0410-b5e6-96231b3b80d8

Check the new form for bison output into CVS

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26208 91177308-0d34-0410-b5e6-96231b3b80d8

bugfixes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26207 91177308-0d34-0410-b5e6-96231b3b80d8

Convert this over to work with the new makefiles

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26206 91177308-0d34-0410-b5e6-96231b3b80d8

Convert the bison-output-checked-into-cvs makefile handling stuff to work
like the flex stuff, which actually works when people do cvs updates and
get conflicts in the updated checked in file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26205 91177308-0d34-0410-b5e6-96231b3b80d8

Sparc actually *DOES* have a directive for emitting zeros.  In fact, it requires
it, because this:

.bss
X:
.byte 0

results in the assembler warning: "initialization in bss segment".  Annoying.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26204 91177308-0d34-0410-b5e6-96231b3b80d8

random lexer change to test the makefile updating stuff

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26203 91177308-0d34-0410-b5e6-96231b3b80d8

Fix SingleSource/Regression/C/2004-08-12-InlinerAndAllocas.c on Sparc.

The ABI specifies that there is a register save area at the bottom of the
stack, which means the actual used pointer needs to be an offset from
the subtracted value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26202 91177308-0d34-0410-b5e6-96231b3b80d8

HP aCC (and a bunch of other compilers, no doubt) don't share
GCC's syntax for auto-dependency generation stuff. This should
be changed to be disabling dependency stuff unless GCC/ICC is
found.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26201 91177308-0d34-0410-b5e6-96231b3b80d8

oops, I meant this

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26200 91177308-0d34-0410-b5e6-96231b3b80d8

zap

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26199 91177308-0d34-0410-b5e6-96231b3b80d8

previously, configure would die if GCC or ICC was not found. Now it'll
go through, but we do want to know if we're using GCC/ICC since they
share certain funky command line options (for dependency generation
stuff)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26198 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26197 91177308-0d34-0410-b5e6-96231b3b80d8

Use .zerofill on x86/darwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26196 91177308-0d34-0410-b5e6-96231b3b80d8

Lower memcpy with small constant size operand into a series of load / store
ops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26195 91177308-0d34-0410-b5e6-96231b3b80d8

more refactoring, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26194 91177308-0d34-0410-b5e6-96231b3b80d8

cvtsd2ss / cvtss2sd encoding bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26193 91177308-0d34-0410-b5e6-96231b3b80d8

movaps, movapd encoding bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26192 91177308-0d34-0410-b5e6-96231b3b80d8

pull some code out into a function

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26191 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase that broke unswitch due to loopsimplify not doing the right thing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26190 91177308-0d34-0410-b5e6-96231b3b80d8

Canonicalize inner loops before outer loops.  Inner loop canonicalization
can provide work for the outer loop to canonicalize.

This fixes a case that breaks unswitching.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26189 91177308-0d34-0410-b5e6-96231b3b80d8

Doh again!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26188 91177308-0d34-0410-b5e6-96231b3b80d8

When splitting exit edges to canonicalize loops, make sure to put the new
block in the appropriate loop nest.

Third time is the charm, right?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26187 91177308-0d34-0410-b5e6-96231b3b80d8

new note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26186 91177308-0d34-0410-b5e6-96231b3b80d8

If we have zero initialized data with external linkage, use .zerofill to
emit it (instead of .space), saving a bit of space in the .o file.

For example:
int foo[100];
int bar[100] = {};

when compiled with C++ or -fno-common results in shrinkage from 1160 to 360
bytes of space.  The X86 backend can also do this on darwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26185 91177308-0d34-0410-b5e6-96231b3b80d8

Using wrong DW_FORM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26184 91177308-0d34-0410-b5e6-96231b3b80d8

Don't special case XS, XD prefixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26183 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fix: XS, XD prefixes were being emitted twice.
XMM registers were not being handled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26182 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure that weak functions are aligned properly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26181 91177308-0d34-0410-b5e6-96231b3b80d8

Duh

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26180 91177308-0d34-0410-b5e6-96231b3b80d8

Remove -disable-x86-sse

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26179 91177308-0d34-0410-b5e6-96231b3b80d8

add an assert

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26178 91177308-0d34-0410-b5e6-96231b3b80d8

Keep to < 80 cols

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26177 91177308-0d34-0410-b5e6-96231b3b80d8

Missed a break so memcpy cases fell through to memset. Doh.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26176 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed a build breakage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26175 91177308-0d34-0410-b5e6-96231b3b80d8

Rename maxStoresPerMemSet to maxStoresPerMemset, etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26174 91177308-0d34-0410-b5e6-96231b3b80d8

Add a entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26173 91177308-0d34-0410-b5e6-96231b3b80d8

Set maxStoresPerMemSet to 16. Ditto for maxStoresPerMemCpy and
maxStoresPerMemMove. Although the last one is not used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26172 91177308-0d34-0410-b5e6-96231b3b80d8

Expand memset dst, c, size to a series of stores if size falls below the
target specific theshold, e.g. 16 for x86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26171 91177308-0d34-0410-b5e6-96231b3b80d8

Enable SSE (for the right subtargets)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26169 91177308-0d34-0410-b5e6-96231b3b80d8

Match changes to unix build system.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26167 91177308-0d34-0410-b5e6-96231b3b80d8

now that libcalls don't suck, we can remove this hack

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26164 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust to new style "generated files in CVS" mechanism for lex output

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26163 91177308-0d34-0410-b5e6-96231b3b80d8

adjust to new style of handling lexer changes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26162 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust to new form of handling lexer dependencies, this way shouldn't have
the problems the old way did.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26161 91177308-0d34-0410-b5e6-96231b3b80d8

Implement an alternative way of handling generated lex files in CVS.  This
should solve the "updating cvs when .l files change give me conflict markers
that break my build" issue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26160 91177308-0d34-0410-b5e6-96231b3b80d8

Wrap a couple more long lines

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26159 91177308-0d34-0410-b5e6-96231b3b80d8

wrap long lines

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26158 91177308-0d34-0410-b5e6-96231b3b80d8

Use statistics to keep track of what flavors of loops we are unswitching

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26157 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a latent bug in the call sequence handling stuff.  Some targets (e.g. x86)
create these nodes with flag results.  Remember that we legalized them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26156 91177308-0d34-0410-b5e6-96231b3b80d8

Implement Instcombine/and.ll:test34

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26155 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26154 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo that caused build failures for things trying to use m_Or.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26153 91177308-0d34-0410-b5e6-96231b3b80d8

If any of the sign extended bits are demanded, the input sign bit is demanded
for a sign extension.

This fixes InstCombine/2006-02-13-DemandedMiscompile.ll and Ptrdist/bc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26152 91177308-0d34-0410-b5e6-96231b3b80d8

testcase reduced from Ptrdist/bc that is miscompiled by recent instcombine changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26151 91177308-0d34-0410-b5e6-96231b3b80d8

Another hack due to allowing multiple symbols with the same name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26150 91177308-0d34-0410-b5e6-96231b3b80d8

Another work around for the 'symbols with different types can have the same
name' issue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26149 91177308-0d34-0410-b5e6-96231b3b80d8

improved zap discovery

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26148 91177308-0d34-0410-b5e6-96231b3b80d8

test for improved zap discovery

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26147 91177308-0d34-0410-b5e6-96231b3b80d8

Sync up the tag numbers with gcc4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26146 91177308-0d34-0410-b5e6-96231b3b80d8

Rename to better reflect usage (current and planned.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26145 91177308-0d34-0410-b5e6-96231b3b80d8

Completely rewrite libcall insertion by the legalizer, providing the
following handy-dandy properties:

1. it is always correct now
2. it is much faster than before
3. it is easier to understand

This implementation builds off of the recent simplifications of the
legalizer that made it single-pass instead of iterative.

This fixes JM/lencod, JM/ldecod, and
CodeGen/Generic/2006-02-12-InsertLibcall.ll (at least on PPC).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26144 91177308-0d34-0410-b5e6-96231b3b80d8

Switch targets over to using SelectionDAG::getCALLSEQ_START to create
CALLSEQ_START nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26143 91177308-0d34-0410-b5e6-96231b3b80d8

Switch to using getCALLSEQ_START instead of using our own creation calls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26142 91177308-0d34-0410-b5e6-96231b3b80d8

Add a method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26141 91177308-0d34-0410-b5e6-96231b3b80d8

this passes now, due to Nate's recent efforts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26140 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce this testcase a bit more, with the help of llvm-extract and some hand tweaks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26139 91177308-0d34-0410-b5e6-96231b3b80d8

Be careful not to request or look at bits shifted in from outside the size
of the input.  This fixes the mediabench/gsm/toast failure last night.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26138 91177308-0d34-0410-b5e6-96231b3b80d8

Added a test case for a libcall insertion bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26137 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing patterns for andi. and andis., fixing test/Regression/CodeGen/
PowerPC/and-imm.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26136 91177308-0d34-0410-b5e6-96231b3b80d8

remove some more dead special case code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26135 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate special case hacks that are superceded by general purpose hacks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26134 91177308-0d34-0410-b5e6-96231b3b80d8

tweaks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26133 91177308-0d34-0410-b5e6-96231b3b80d8

Three changes:
1. Teach GetConstantInType to handle boolean constants.
2. Teach instcombine to fold (compare X, CST) when X has known 0/1 bits.
   Testcase here: set.ll:test22
3. Improve the "(X >> c1) & C2 == 0" folding code to allow a noop cast
   between the shift and and.  More aggressive bitfolding for other reasons
   was turning signed shr's into unsigned shr's, leaving the noop cast in
   the way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26131 91177308-0d34-0410-b5e6-96231b3b80d8