void TR_LoadExtensions::flagPreferredLoadExtensions(TR::Node* parent)
{
if (isSupportedType(parent) && parent->getOpCode().isConversion())
{
TR::Node* child = parent->getFirstChild();
bool canSkipConversion = false;
if (isSupportedType(child))
{
if (parent->getSize() == child->getSize())
{
TR::DebugCounter::incStaticDebugCounter(comp(), TR::DebugCounter::debugCounterName(comp(), "codegen/LoadExtensions/success/unneededConversion/%s", comp()->signature()));
parent->setUnneededConversion(true);
}
else
{
TR::ILOpCode& childOpCode = child->getOpCode();
if (childOpCode.isLoadReg()
&& !(parent->getSize() > 4 && TR::Compiler->target.is32Bit())
&& excludedNodes->count(parent) == 0)
{
TR::Node* useRegLoad = child;
TR_UseDefInfo* useDefInfo = optimizer()->getUseDefInfo();
if (useDefInfo != NULL && useDefInfo->infoIsValid() && useRegLoad->getUseDefIndex() != 0 && useDefInfo->isUseIndex(useRegLoad->getUseDefIndex() != 0))
{
TR_UseDefInfo::BitVector info(comp()->allocator());
if (useDefInfo->getUseDef(info, useRegLoad->getUseDefIndex()))
{
TR_UseDefInfo::BitVector::Cursor cursor(info);
int32_t firstDefIndex = useDefInfo->getFirstRealDefIndex();
int32_t firstUseIndex = useDefInfo->getFirstUseIndex();
canSkipConversion = true;
bool forceExtensionOnAnyLoads = false;
bool forceExtensionOnAllLoads = true;
for (cursor.SetToFirstOne(); cursor.Valid() && canSkipConversion; cursor.SetToNextOne())
{
int32_t defIndex = cursor;
// We've examined all the defs of this particular use
if (defIndex >= firstUseIndex)
{
break;
}
// Do not consider defs that correspond to method arguments as we cannot force extension on those
if (defIndex < firstDefIndex)
{
continue;
}
TR::Node* defRegLoad = useDefInfo->getNode(defIndex);
if (defRegLoad != NULL)
{
TR::Node* defRegLoadChild = defRegLoad->getFirstChild();
bool forceExtension = false;
canSkipConversion = TR_LoadExtensions::canSkipConversion(parent, defRegLoadChild, forceExtension);
forceExtensionOnAnyLoads |= forceExtension;
forceExtensionOnAllLoads &= forceExtension;
// If we have to force extension on any loads which feed a def of this use ensure we must also
// force extension on all such loads. Conversely the conversion can be skipped if none of the
// loads feeding the def of this use need to be extended. This ensures either all loads feeding
// into defs of this use should be extended or none of them.
canSkipConversion &= forceExtensionOnAllLoads == forceExtensionOnAnyLoads;
if (trace())
{
traceMsg(comp(), "\t\tPeeked through %s [%p] and found %s [%p] with child %s [%p] - conversion %s be skipped\n",
useRegLoad->getOpCode().getName(),
useRegLoad,
defRegLoad->getOpCode().getName(),
defRegLoad,
defRegLoadChild->getOpCode().getName(),
defRegLoadChild,
canSkipConversion ?
"can" :
"cannot");
}
}
}
if (canSkipConversion && performTransformation(comp(), "%sSkipping conversion %s [%p] after RegLoad\n", optDetailString(), parent->getOpCode().getName(), parent))
{
TR::DebugCounter::incStaticDebugCounter(comp(), TR::DebugCounter::debugCounterName(comp(), "codegen/LoadExtensions/success/unneededConversion/GRA/%s", comp()->signature()));
parent->setUnneededConversion(true);
if (forceExtensionOnAllLoads)
{
TR_UseDefInfo::BitVector info(comp()->allocator());
if (useDefInfo->getUseDef(info, useRegLoad->getUseDefIndex()))
{
TR_UseDefInfo::BitVector::Cursor cursor(info);
for (cursor.SetToFirstOne(); cursor.Valid(); cursor.SetToNextOne())
{
int32_t defIndex = cursor;
// We've examined all the defs of this particular use
if (defIndex >= firstUseIndex)
{
break;
}
// Do not consider defs that correspond to method arguments as we cannot force extension on those
if (defIndex < firstDefIndex)
{
continue;
}
TR::Node *defRegLoad = useDefInfo->getNode(defIndex);
if (defRegLoad != NULL)
{
TR::Node* defRegLoadChild = defRegLoad->getFirstChild();
const int32_t preference = getExtensionPreference(defRegLoadChild);
if (preference > 0)
{
if (trace())
{
traceMsg(comp(), "\t\t\tForcing sign extension on %s [%p]\n",
defRegLoadChild->getOpCode().getName(),
defRegLoadChild);
}
if (parent->getSize() == 8 || parent->useSignExtensionMode())
{
defRegLoadChild->setSignExtendTo64BitAtSource(true);
}
else
{
defRegLoadChild->setSignExtendTo32BitAtSource(true);
}
}
if (preference < 0)
{
if (trace())
{
traceMsg(comp(), "\t\t\tForcing zero extension on %s [%p]\n",
defRegLoadChild->getOpCode().getName(),
defRegLoadChild);
}
if (parent->getSize() == 8 || parent->useSignExtensionMode())
{
defRegLoadChild->setZeroExtendTo64BitAtSource(true);
}
else
{
defRegLoadChild->setZeroExtendTo32BitAtSource(true);
}
}
}
}
}
}
if (parent->getType().isInt64() && parent->getSize() > child->getSize())
{
if (trace())
{
traceMsg(comp(), "\t\t\tSet global register %s in getExtendedToInt64GlobalRegisters for child %s [%p] with parent node %s [%p]\n",
comp()->getDebug()->getGlobalRegisterName(child->getGlobalRegisterNumber()),
child->getOpCode().getName(),
child,
parent->getOpCode().getName(),
parent);
}
// getExtendedToInt64GlobalRegisters is used by the evaluators to force a larger virtual register to be used when
// evaluating the regload so any instructions generated by local RA are the correct size to preserve the upper bits
cg()->getExtendedToInt64GlobalRegisters()[child->getGlobalRegisterNumber()] = true;
}
}
}
}
}
}
}
if (!canSkipConversion)
{
bool forceExtension = false;
canSkipConversion = TR_LoadExtensions::canSkipConversion(parent, child, forceExtension);
if (canSkipConversion && performTransformation(comp(), "%sSkipping conversion %s [%p]\n", optDetailString(), parent->getOpCode().getName(), parent))
{
TR::DebugCounter::incStaticDebugCounter(comp(), TR::DebugCounter::debugCounterName(comp(), "codegen/LoadExtensions/success/unneededConversion/%s", comp()->signature()));
parent->setUnneededConversion(true);
if (forceExtension)
{
const int32_t preference = getExtensionPreference(child);
if (preference > 0)
{
if (trace())
{
traceMsg(comp(), "\t\t\tForcing sign extension on %s [%p]\n",
child->getOpCode().getName(),
child);
}
if (parent->getSize() == 8 || parent->useSignExtensionMode())
{
child->setSignExtendTo64BitAtSource(true);
}
else
{
child->setSignExtendTo32BitAtSource(true);
}
}
if (preference < 0)
{
if (trace())
{
traceMsg(comp(), "\t\t\tForcing zero extension on %s [%p]\n",
child->getOpCode().getName(),
child);
}
if (parent->getSize() == 8 || parent->useSignExtensionMode())
{
child->setZeroExtendTo64BitAtSource(true);
}
else
{
child->setZeroExtendTo32BitAtSource(true);
}
}
}
}
}
}
}
const bool TR_LoadExtensions::canSkipConversion(TR::Node* conversion, TR::Node* child, bool& forceExtension)
{
bool result = false;
// Assume we are not forcing the load (if the child is really a load) to be zero/sign extended
forceExtension = false;
if (trace())
{
traceMsg(comp(), "\t\tExamining conversion %s [%p]\n",
conversion->getOpCode().getName(),
conversion);
}
if (isSupportedType(child) && excludedNodes->count(child) == 0)
{
const int32_t preference = getExtensionPreference(child);
const bool loadPrefersSignExtension = preference > 0;
const bool loadPrefersZeroExtension = preference < 0;
TR::ILOpCode& conversionOpCode = conversion->getOpCode();
if (isSupportedLoad(child) &&
// Only consider widening conversions
conversion->getSize() > child->getSize() &&
// Ensure we do not use register pairs for 64-bit loads on 32-bit platforms
(TR::Compiler->target.is64Bit() || comp()->cg()->use64BitRegsOn32Bit() || conversion->getSize() != 8) &&
// Ensure the conversion matches our preferred extension on the load
(loadPrefersSignExtension && loadPrefersSignExtension == conversionOpCode.isSignExtension() ||
loadPrefersZeroExtension && loadPrefersZeroExtension == conversion->isZeroExtension()))
{
if (trace())
{
traceMsg(comp(), "\t\tDetected sign extension pattern on widening conversion %s [%p] and load %s [%p]\n",
conversion->getOpCode().getName(),
conversion,
child->getOpCode().getName(),
child);
}
forceExtension = true;
result = true;
}
if (conversion->getSize() < child->getSize())
{
// TODO (Issue #2213): Determine whether this case is ever needed and why? Shouldn't the simplifier have eliminated such IL?
if (child->getOpCode().isConversion())
{
TR::Node* grandChild = child->getFirstChild();
if (isSupportedLoad(grandChild) &&
// Conversion is narrowing down to the original width (i.e. stacked conversion which is a NOP)
conversion->getSize() == grandChild->getSize())
{
if (trace())
{
traceMsg(comp(), "\t\tDetected sign extension pattern on narrowing conversion %s [%p] and load %s [%p]\n",
conversion->getOpCode().getName(),
conversion,
child->getOpCode().getName(),
child);
}
result = true;
}
}
}
}
return result;
}
void TR_LoadExtensions::findPreferredLoadExtensions(TR::Node* parent)
{
TR::ILOpCode& parentOpCode = parent->getOpCode();
// count how a load is being used. As a signed or unsigned number?
if (isSupportedType(parent) && parentOpCode.isConversion())
{
TR::Node* child = parent->getFirstChild();
// Only examine non-trivial conversions
if (isSupportedType(child) && parent->getSize() != child->getSize())
{
if (isSupportedLoad(child))
{
setExtensionPreference(child, parent);
}
else if (child->getOpCode().isLoadReg())
{
TR::Node* useRegLoad = child;
TR_UseDefInfo* useDefInfo = optimizer()->getUseDefInfo();
// If we have usedef info we can traverse all defs of this particular use and if all the defs are stores
// of supported counted loads then we can count such loads as well. If this criteria is not met then there
// exists at least one def (store) of this particular use which feeds from a non-load operation (an
// addition for example). These are not candidates for skipping extension because we cannot easily extend
// a non-load operation.
if (useDefInfo != NULL && useDefInfo->infoIsValid() && useRegLoad->getUseDefIndex() != 0 && useDefInfo->isUseIndex(useRegLoad->getUseDefIndex() != 0))
{
TR_UseDefInfo::BitVector info(comp()->allocator());
if (useDefInfo->getUseDef(info, useRegLoad->getUseDefIndex()))
{
if (trace())
{
traceMsg(comp(), "\t\tPeeking through RegLoad %p for conversion %s [%p]\n",
useRegLoad,
parentOpCode.getName(),
parent);
}
TR_UseDefInfo::BitVector::Cursor cursor(info);
int32_t firstDefIndex = useDefInfo->getFirstRealDefIndex();
int32_t firstUseIndex = useDefInfo->getFirstUseIndex();
for (cursor.SetToFirstOne(); cursor.Valid(); cursor.SetToNextOne())
{
int32_t defIndex = cursor;
// We've examined all the defs of this particular use
if (defIndex >= firstUseIndex)
{
break;
}
// Do not consider defs that correspond to method arguments as we cannot force extension on those
if (defIndex < firstDefIndex)
{
(*excludedNodes)[parent] = true;
break;
}
TR::Node* defRegLoad = useDefInfo->getNode(defIndex);
if (defRegLoad != NULL)
{
TR::Node* defRegLoadChild = defRegLoad->getFirstChild();
if (defRegLoad->getOpCode().isStoreReg() && isSupportedType(defRegLoadChild) && isSupportedLoad(defRegLoadChild))
{
if (trace())
{
traceMsg(comp(), "\t\tPeeked through use %s [%p] and found def %s [%p] with child %s [%p] - Counting [%p]\n",
useRegLoad->getOpCode().getName(),
useRegLoad,
defRegLoad->getOpCode().getName(),
defRegLoad,
defRegLoadChild->getOpCode().getName(),
defRegLoadChild,
defRegLoadChild);
}
setExtensionPreference(defRegLoadChild, parent);
}
else
{
if (trace())
{
traceMsg(comp(), "\t\tPeeked through use %s [%p] and found def %s [%p] with child %s [%p] - Excluding [%p]\n",
useRegLoad->getOpCode().getName(),
useRegLoad,
defRegLoad->getOpCode().getName(),
defRegLoad,
defRegLoadChild != NULL ?
defRegLoadChild->getOpCode().getName() :
"NULL",
defRegLoadChild,
parent);
}
(*excludedNodes)[parent] = true;
}
}
}
}
}
else
{
(*excludedNodes)[parent] = true;
}
}
}
}
// Exclude all loads which feed into global register stores which require sign extensions. This must be done
// because Load Extensions is a local optimization and it must respect global sign extension decisions made
// by GRA. Excluding such loads prevents a situation where GRA decided that a particular global register
// should be sign extended at its definitions however Load Extensions has determined that the same load
// should be zero extended. If local RA were to pick the same register for the global register as well as
// the load then we have a conflicting decision which will result in a conversion to be skipped when it is
// not supposed to be.
if (parentOpCode.isStoreReg() && parent->needsSignExtension() && parent->getFirstChild()->getOpCode().isLoadVar())
{
(*excludedNodes)[parent->getFirstChild()] = true;
}
}
TR::Node *
OMR::Simplifier::unaryCancelOutWithChild(TR::Node * node, TR::Node * firstChild, TR::TreeTop *anchorTree, TR::ILOpCodes opcode, bool anchorChildren)
{
if (!isLegalToUnaryCancel(node, firstChild, opcode))
return NULL;
if (firstChild->getOpCodeValue() == opcode &&
(node->getType().isAggregate() || firstChild->getType().isAggregate()) &&
(node->getSize() > firstChild->getSize() || node->getSize() != firstChild->getFirstChild()->getSize()))
{
// ensure a truncation side-effect of a conversion is not lost
// o2a size=3
// a2o size=3 // conversion truncates in addition to type cast so cannot be removed
// loadaddr size=4
// This restriction could be loosened to only disallow intermediate truncations (see BCD case above) but then would require a node
// op that would just correct for size (e.g. addrSizeMod size=3 to replace the o2a/a2o pair)
//
// Do allow cases when all three sizes are the same and when the middle node widens but the top and bottom node have the same size, e.g.
//
// i2o size=3
// o2i size=4
// oload size=3
//
// Also allow the special case where the grandchild is not really truncated as the 'truncated' bytes are known to be zero
// (i.e. there really isn't an intermediate truncation of 4->3 even though it appears that way from looking at the sizes alone)
// o2i
// i2o size=3
// iushr
// x
// iconst 8
bool disallow = true;
TR::Node *grandChild = firstChild->getFirstChild();
size_t nodeSize = node->getSize();
if (node->getType().isIntegral() &&
nodeSize == grandChild->getSize() &&
nodeSize > firstChild->getSize())
{
size_t truncatedBits = (nodeSize - firstChild->getSize()) * 8;
if (grandChild->getOpCode().isRightShift() && grandChild->getOpCode().isShiftLogical() &&
grandChild->getSecondChild()->getOpCode().isLoadConst() &&
(grandChild->getSecondChild()->get64bitIntegralValue() == truncatedBits))
{
disallow = false;
if (trace())
traceMsg(comp(),"do allow unaryCancel of node %s (%p) and firstChild %s (%p) as grandChild %s (%p) zeros the %d truncated bytes\n",
node->getOpCode().getName(),node,firstChild->getOpCode().getName(),firstChild,
grandChild->getOpCode().getName(),grandChild,truncatedBits/8);
}
}
if (disallow)
{
if (trace())
traceMsg(comp(),"disallow unaryCancel of node %s (%p) and firstChild %s (%p) due to unequal sizes (nodeSize %d, firstChildSize %d, firstChild->childSize %d)\n",
node->getOpCode().getName(),node,firstChild->getOpCode().getName(),firstChild,
node->getSize(),firstChild->getSize(),firstChild->getFirstChild()->getSize());
return NULL;
}
}
if (firstChild->getOpCodeValue() == opcode &&
performTransformation(comp(), "%sRemoving node [" POINTER_PRINTF_FORMAT "] %s and its child [" POINTER_PRINTF_FORMAT "] %s\n",
optDetailString(), node, node->getOpCode().getName(), firstChild, firstChild->getOpCode().getName()))
{
TR::Node *grandChild = firstChild->getFirstChild();
grandChild->incReferenceCount();
bool anchorChildrenNeeded = anchorChildren &&
(node->getNumChildren() > 1 ||
firstChild->getNumChildren() > 1 ||
node->getOpCode().hasSymbolReference() ||
firstChild->getOpCode().hasSymbolReference());
prepareToStopUsingNode(node, anchorTree, anchorChildrenNeeded);
node->recursivelyDecReferenceCount();
node->setVisitCount(0);
return grandChild;
}
return NULL;
}