compute dominator tree children in a deterministic order that does not depend
on the address of BasicBlock objects in memory.  This eliminates stuff like this:

 Inorder Dominator Tree:
   [1]  %entry
     [2]  %loopentry
-      [3]  %loopexit
       [3]  %no_exit
-        [4]  %endif
         [4]  %then
+        [4]  %endif
+      [3]  %loopexit
       [3]  %return

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14253 91177308-0d34-0410-b5e6-96231b3b80d8

Print out immediate dominators in program order, not in random order based on the address
of BasicBlock objects

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14252 91177308-0d34-0410-b5e6-96231b3b80d8

This will hopefully fix a heisenbug that Vladimir Merzliakov is running
into valiantly trying to compile stuff on freebsd.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14251 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note about GCC 3.3.2 optimization bug that causes llc to spin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14250 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a nasty bug, noticed by Reid

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14249 91177308-0d34-0410-b5e6-96231b3b80d8

Fix one source of nondeterminism in the -licm pass: the hoist pass
was processing blocks in whatever order they happened to end up in the
dominator tree data structure.  Force an ordering.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14248 91177308-0d34-0410-b5e6-96231b3b80d8

Change to use the StableBasicBlockNumbering class

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14247 91177308-0d34-0410-b5e6-96231b3b80d8

Initial checkin of the StableBasicBlockNumbering, a little helper class for computing
(strangely enough) a stable (determinstic) numbering for basic blocks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14246 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a tiny bug in the -no-aa pass, in which it did not ever get a target data.
This is a regression from 1.2, though noone uses -no-aa anyway

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14245 91177308-0d34-0410-b5e6-96231b3b80d8

Do not let the numbering of PHI nodes placed in the function depend on
non-deterministic things like the ordering of blocks in the dominance
frontier of a BB.  Unfortunately, I don't know of a better way to solve
this problem than to explicitly sort the BB's in function-order before
processing them.  This is guaranteed to slow the pass down a bit, but
is absolutely necessary to get usable diffs between two different tools
executing the mem2reg or scalarrepl pass.

Before this, bazillions of spurious diff failures occurred all over the
place due to the different order of processing PHIs:

-       %tmp.111 = getelementptr %struct.Connector_struct* %upcon.0.0, uint 0, uint 0
+       %tmp.111 = getelementptr %struct.Connector_struct* %upcon.0.1, uint 0, uint 0

Now, the diffs match.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14244 91177308-0d34-0410-b5e6-96231b3b80d8

Do not sort by the address of LLVM ConstantInt* objects.  This produces
nondeterministic results that depend on where these objects land in memory.
Instead, sort by the value of the constant, which is stable.

Before this patch, the -simplifycfg pass run from two different compilers
could cause different code to be generated, though it was semantically the
same:

@@ -12258,8 +12258,8 @@
        %s_addr.1 = phi sbyte* [ %s, %entry ], [ %inc.0, %no_exit ]             ; <sbyte*> [#uses=5]
        %tmp.1 = load sbyte* %s_addr.1          ; <sbyte> [#uses=1]
        switch sbyte %tmp.1, label %no_exit [
-                sbyte 0, label %loopexit
                 sbyte 46, label %loopexit
+                sbyte 0, label %loopexit
        ]

We need to stomp all of this stuff out.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14243 91177308-0d34-0410-b5e6-96231b3b80d8

Do not loop over uses as we delete them.  This causes iterators to be
invalidated out from under us.  This bug goes back to revision 1.1: scary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14242 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14241 91177308-0d34-0410-b5e6-96231b3b80d8

Test for PR374

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14240 91177308-0d34-0410-b5e6-96231b3b80d8

Fix relative link to the CVS mirrors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14239 91177308-0d34-0410-b5e6-96231b3b80d8

Add link to the CVS mirror hosted by Reid

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14238 91177308-0d34-0410-b5e6-96231b3b80d8

Add link documenting shared/static library differences across platforms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14237 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14236 91177308-0d34-0410-b5e6-96231b3b80d8

Capitalize Cygwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14235 91177308-0d34-0410-b5e6-96231b3b80d8

* Fix file header and name
* Order #includes alphabetically

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14234 91177308-0d34-0410-b5e6-96231b3b80d8

Use the machine-independent method of querying the page size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14233 91177308-0d34-0410-b5e6-96231b3b80d8

Add a target-independent way to query page size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14232 91177308-0d34-0410-b5e6-96231b3b80d8

.zero doesn't work in the Solaris assembler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14231 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of selects the easy way

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14230 91177308-0d34-0410-b5e6-96231b3b80d8

Make visitAllocaInst() look more like its X86 counterpart.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14229 91177308-0d34-0410-b5e6-96231b3b80d8

Mess around with allocation order. In particular, I think we ought to be
using the local & in regs first because they are not clobbered by calls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14228 91177308-0d34-0410-b5e6-96231b3b80d8

JMPL has a delay slot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14227 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up the commented-out F3_3 stuff.
Replace it with a working class for FP instrs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14226 91177308-0d34-0410-b5e6-96231b3b80d8

Fix jmpl.
Add some FP moves.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14225 91177308-0d34-0410-b5e6-96231b3b80d8

Support printing base+offset pairs where the offset is a register.
Use this for printing the jmpl indirect-call instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14224 91177308-0d34-0410-b5e6-96231b3b80d8

Support intrinsic calls (although no particular intrinsics are supported yet).
Support indirect calls.
Support returning a float value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14223 91177308-0d34-0410-b5e6-96231b3b80d8

Calls to __builtin_isunordered and friends no longer cause llvmgcc to ICE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14222 91177308-0d34-0410-b5e6-96231b3b80d8

Implement Transforms/InstCombine/and.ll:test17, a common case that
occurs due to unordered comparison macros in math.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14221 91177308-0d34-0410-b5e6-96231b3b80d8

New test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14220 91177308-0d34-0410-b5e6-96231b3b80d8

Do not function resolve intrinsics.  This prevents warnings and possible bad
things from happening due to

declare bool %llvm.isunordered(double, double)
declare bool %llvm.isunordered(float, float)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14219 91177308-0d34-0410-b5e6-96231b3b80d8

Test that C99 unordered comparison functions are not being turned into libc calls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14218 91177308-0d34-0410-b5e6-96231b3b80d8

Add load instructions for floating-point registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14217 91177308-0d34-0410-b5e6-96231b3b80d8

Support alloca instructions.
Support copying floating-point constants to registers.
Add assertion to visitCallInst to abort if we hit a NULL calledFunction, for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14216 91177308-0d34-0410-b5e6-96231b3b80d8

Fix printing of Argument objects, problem found by Patrick Meredith

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14215 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase for GCC unordered comparison builtins

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14214 91177308-0d34-0410-b5e6-96231b3b80d8

Codegen sub C, X a little bit better for register pressure.  Instead of

mov REG, C
sub REG, X

generate:

neg X
add X, C

which uses one less reg

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14213 91177308-0d34-0410-b5e6-96231b3b80d8

Fold setcc instructions into select and branches that are not in the same BB as
the setcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14212 91177308-0d34-0410-b5e6-96231b3b80d8

Make storeRegToStackSlot slightly shorter.
Make copyRegToReg return 1 instead of -1.
Edit a comment in emitPrologue().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14211 91177308-0d34-0410-b5e6-96231b3b80d8

Set the isBranch and isTerminator flags on branch instructions correctly.
Add a FIXME about the (currently unused) JMPL instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14210 91177308-0d34-0410-b5e6-96231b3b80d8

Emit stores correctly; don't fail an assertion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14209 91177308-0d34-0410-b5e6-96231b3b80d8

Support generating machine instructions for Phi nodes (based on x86, but with
modifications for 1 LLVM BB --> many MBBs).
Fix store operand order: make it always be Base, Offset, SrcReg (think
"[ Base + Offset ] = SrcReg").
Rewrite visitBranchInst() to be even dumber (but working) -- give up on
the branch fallthrough trick, for the time being.
Make visitSetCondInst() work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14208 91177308-0d34-0410-b5e6-96231b3b80d8

Recognize more branches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14207 91177308-0d34-0410-b5e6-96231b3b80d8

I love the smell of a freshly broken PowerPC build in the morning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14206 91177308-0d34-0410-b5e6-96231b3b80d8

Make debugging dumps w/ multiple MachineBBs for a given LLVM BB readable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14205 91177308-0d34-0410-b5e6-96231b3b80d8

Do not fold loads into instructions if it is used more than once.  In particular
we do not want to fold the load in cases like this:

  X = load
    = add A, X
    = add B, X

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14204 91177308-0d34-0410-b5e6-96231b3b80d8

Fix compilation problem on freebsd.  Problem noted by Vladimir Merzliakov in
PR371

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14203 91177308-0d34-0410-b5e6-96231b3b80d8

Use addGlobalAddress and addMBB for call & branch targets instead of addPCDisp.
Abort if we see a PCRelativeDisp MachineOperand, to be safe. This matches
the X86 backend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14202 91177308-0d34-0410-b5e6-96231b3b80d8

Rename Type::PrimitiveID to TypeId and ::getPrimitiveID() to ::getTypeID()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14201 91177308-0d34-0410-b5e6-96231b3b80d8

Rename Type::PrimitiveID to TypeId and ::getPrimitiveID() to ::getTypeID()
Delete two functions that are now methods on the Type class

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14200 91177308-0d34-0410-b5e6-96231b3b80d8

Minor cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14199 91177308-0d34-0410-b5e6-96231b3b80d8

Make header comment fit within 80 cols.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14198 91177308-0d34-0410-b5e6-96231b3b80d8

Since the Great LLVM Tool Renaming(tm), we no longer have collisions between our
assembler/linker and the system equivalents.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14197 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo in DEBUG printout.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14196 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a recent regression in Applications/sgefa that Alkis pointed out to me.
The vector may actually be empty if the register that we are marking as
recently used is not actually allocatable.  This happens for physical registers
that are not allocatable, like the ST(x) registers on X86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14195 91177308-0d34-0410-b5e6-96231b3b80d8

Add int ferror(FILE *)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14194 91177308-0d34-0410-b5e6-96231b3b80d8

I'm afraid this doesn't exist.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14193 91177308-0d34-0410-b5e6-96231b3b80d8

Um, did someone make a typo or something?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14192 91177308-0d34-0410-b5e6-96231b3b80d8

isnan is dead

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14191 91177308-0d34-0410-b5e6-96231b3b80d8

llvm.isnan is dead, long live llvm.isunordered!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14190 91177308-0d34-0410-b5e6-96231b3b80d8

Remove support for llvm.isnan.  Alkis wins :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14189 91177308-0d34-0410-b5e6-96231b3b80d8

This testcase is a bit silly now, but oh well :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14188 91177308-0d34-0410-b5e6-96231b3b80d8

Remove isnan support, correct isunordered support

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14187 91177308-0d34-0410-b5e6-96231b3b80d8

Remove support for the isnan intrinsic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14186 91177308-0d34-0410-b5e6-96231b3b80d8

Add basic support for the isunordered intrinsic.  The isnan stuff still needs to go

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14185 91177308-0d34-0410-b5e6-96231b3b80d8

Right, stop being silly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14184 91177308-0d34-0410-b5e6-96231b3b80d8

Add a nan, stop testing isnan

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14183 91177308-0d34-0410-b5e6-96231b3b80d8

Fix thinko in visitor... ShiftInsts should currently be delegated
to visitBinaryOperator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14182 91177308-0d34-0410-b5e6-96231b3b80d8

Do not dereference end iterators.  It's really bad for the asmwriter's health.
This possibly fixes PR370

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14181 91177308-0d34-0410-b5e6-96231b3b80d8

I think we'll use the standard lowering passes for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14179 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bug generating code for void call instructions: don't call
getReg() on void value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14178 91177308-0d34-0410-b5e6-96231b3b80d8

Squash a warning from the Solaris assembler by aligning the stack
on a double-word boundary instead of a single-word boundary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14177 91177308-0d34-0410-b5e6-96231b3b80d8

Allow special-casing of operand printing based on opcode.  Print
non-register, non-immed. arguments to SETHI and OR using %hi() and
%lo() respectively.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14176 91177308-0d34-0410-b5e6-96231b3b80d8

Support constant GEP expressions.
Support copying long constants to register pairs.
Support copying ConstantPointerNulls and ConstantPointerRefs to registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14175 91177308-0d34-0410-b5e6-96231b3b80d8

The correct prototype is `int atoi(CONST char*)'.  Unbroke AIX build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14173 91177308-0d34-0410-b5e6-96231b3b80d8

Add file comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14172 91177308-0d34-0410-b5e6-96231b3b80d8

Quick hack to get this file compiling again on Mac OS X.  The right thing to do
is write an autoconf macro that checks whether __isnan or isnan actually works
**using the C++ compiler after #include <cmath>**, instead of doing it the easy
way with AC_CHECK_FUNCS().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14171 91177308-0d34-0410-b5e6-96231b3b80d8

Make -print-machineinstrs even stronger. You get to see the final code after
peepholing, and make it work the same way in the JIT as in LLC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14170 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of "might be uninitialized" warnings when compiling with GCC 3.3.2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14169 91177308-0d34-0410-b5e6-96231b3b80d8

Add constant folding capabilities to the isunordered intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14168 91177308-0d34-0410-b5e6-96231b3b80d8

Really add the docs this time :-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14167 91177308-0d34-0410-b5e6-96231b3b80d8

Make assertions more consistent with the rest of the intrinsic
function verification and make it a requirement that both arguments to
llvm.isunordered are of the same type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14165 91177308-0d34-0410-b5e6-96231b3b80d8

Add the isunordered intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14159 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce the number of columns in the alpha index to 4 so that it fits
on a 1280x1024 screen!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14158 91177308-0d34-0410-b5e6-96231b3b80d8

It is no longer 2001

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14157 91177308-0d34-0410-b5e6-96231b3b80d8

Don't trim @PrevDays (causing an error) if it isn't long enough.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14156 91177308-0d34-0410-b5e6-96231b3b80d8

Fix grammar: 's is for possessive only.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14155 91177308-0d34-0410-b5e6-96231b3b80d8

Implement tracking of bytecode instruction size and the number of long
instructions generated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14154 91177308-0d34-0410-b5e6-96231b3b80d8

Add two options to the program:
-release causes ENABLE_OPTIMIZED=1 to be added to the make line so that
         the test runs against optimized code.
-pedantic causes the CompileOptimizeOpts variable to be overridden to
          add several additional warnings not picked up by -Wall. This
  should help catch additional programming faux pas. See the diff
  for the specific details.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14153 91177308-0d34-0410-b5e6-96231b3b80d8

Fix fallout from getOffsetOfLocalArea() being negated.  Debugging dumps were being
printed incorrectly, and we were reserving 8 extra bytes of stack space for functions
on X86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14152 91177308-0d34-0410-b5e6-96231b3b80d8

llvm.isnan doesn't access memory

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14151 91177308-0d34-0410-b5e6-96231b3b80d8

Constant fold the isnan intrinsic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14150 91177308-0d34-0410-b5e6-96231b3b80d8

Add tests for isnan

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14149 91177308-0d34-0410-b5e6-96231b3b80d8

By far, one of the most common uses of isnan is to make 'isunordered'
comparisons.  In an 'isunordered' predicate, which looks like this at
the LLVM level:

        %a = call bool %llvm.isnan(double %X)
        %b = call bool %llvm.isnan(double %Y)
        %COM = or bool %a, %b

We used to generate this code:

        fxch %ST(1)
        fucomip %ST(0), %ST(0)
        setp %AL
        fucomip %ST(0), %ST(0)
        setp %AH
        or %AL, %AH

With this patch, we generate this code:

        fucomip %ST(0), %ST(1)
        fstp %ST(0)
        setp %AL

Which should make alkis happy.  Tested as X86/compare_folding.llx:test1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14148 91177308-0d34-0410-b5e6-96231b3b80d8

Test that the X86 backend is only emitting one fucom instruction
for each 'COM =' line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14147 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bug in previous checkin

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14146 91177308-0d34-0410-b5e6-96231b3b80d8