pr970-1.patch
Performing this audit found several minor issues.

This is the first of a small series of patches. It includes changes which
should be uncontroversial, including several typo fixes and stripping some code
which was found to be dead. There's one diff which doesn't read well; it's just
removing an if block and correcting interior indentation.

pr970-2.patch
This patch modifies several assertions to be correct. Comparison instructions
do not apply to vectors, and SCEV will introduce canonical induction variables
only of an integer type.

pr970-3.patch
This patch includes two changes I'm not entirely certain of.

First, in PatternMatch.h, the m_Neg matcher is incorrect for VP vectors. It
matches sub <+0.0, +0.0, ...>, %x but the correct pattern is sub <-0.0, -0.0,
...>, %x. Since this matcher is dead, I simply commented it out with a note
that it was broken. It would be simple to delete or fix it instead.

Second, in GlobalOpt.cpp, the ShrinkGlobalToBoolean transformation excludes FP,
but allows vectors. It doesn't explain why. FP exclusion was added last
November to fix a crash, it looks like. I made it exclude vectors, too, but
there's no overwhelming reason why.

pr970-4.patch
This patch fixes several routines which deal with negation but failed to use
the
proper floating point negation formula, f(x) = -0.0 - x, for vectors. I simply
factored out the logic to get the properly-signed "zero" value and extended it
to get the correct floating-point vector.

In most cases, these local logic errors were harmless because callers operated
exclusively on scalar integers, but in four cases there look to be potential
problems or missed optimizations.

This patch fixes the dead m_Neg matcher that patch 3 commented out.

at lib/Transforms/Scalar/InstructionCombining.cpp:6788:

      NegVal = InsertNewInstBefore(
	    BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"),

SI);

This can pass vectors to createNeg and looks like it might miscompile code by
inserting a sub <+0.0, +0.0, etc.>, %x as it simplifies an expression. This
patch should fix this case.

at lib/Target/CBackend/CBackend.cpp:2129

This is now llvm/llvm-bugzilla-archive#1126 .

at lib/Transforms/Scalar/InstructionCombining.cpp:325:

Please review this call site in getComplexity. It is clearly reachable with FP
vector operands, and isNeg will now return true for vector negation, which is
not a single machine instruction as fneg or integer negation are. Therefore,
existing behavior might be preferable. If so, exclude vectors:

  - if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
  + if ((!isa<PackedType>(V->getType()) && BinaryOperator::isNeg(V)) ||
  +     BinaryOperator::isNot(V))

at lib/Transforms/Scalar/InstructionCombining.cpp:473:

static inline Value *dyn_castNegVal(Value *V) {
if (BinaryOperator::isNeg(V))
return BinaryOperator::getNegArgument(V);

This could clearly be invoked with FP vector operands. With this patch, the
following transformations would be extended to FP vectors:

-A + B   -->  B - A
A + -B   -->  A - B
A - (-B) -->  A + B
-X * -Y  -->  X * Y

These would already have applied to scalar FP, so I have to assume that they're

IEEE-approved.

Just one more!

pr970-5.patch
This final patch simplifies the logic in SelectionDAGISel.cpp. No functional
change. This construct was in use:

if (I.getType()->isFloatingPoint())
  visitFPBinary(I, ISD::FADD, ISD::VADD); 
else
  visitIntBinary(I, ISD::ADD, ISD::VADD); 

Which confusingly sent even floating-point vector operations through the
"visitIntBinary" routine. However, this was not a bug because visitFPBinary and
visitIntBinary were identical (as is the vector opcode).

This patch streamlines the logic.

Oh, and one more thing...

In LowerSelect.cpp:

bool LowerSelect::runOnFunction(Function &F) {
bool Changed = false;
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
if (SelectInst *SI = dyn_cast(I))
if (!OnlyFP || SI->getType()->isFloatingPoint()) {
// Split this basic block in half right before the select instruction.

This pass is dead except for 'opt -lowerselect', so there's no client to say which (isFloatingPoint or
isFPOrFPVector) is the correct behavior for the OnlyFP option. The pass could be deleted, the OnlyFP
option disabled, the test switched to !OnlyFP || SI->getType()->isFPOrFPVector(), or the pass extended
to lower any combination of FP and integer scalars and vectors.

Gordon,

With a few minor cleanups, these patches have all been committed.

Thanks for the great patches!

Reid.

Mine.

Thanks, Reid.