Files
intel-graphics-compiler/visa/SpillCleanup.cpp
2018-01-30 10:00:45 -08:00

2116 lines
73 KiB
C++

/*===================== begin_copyright_notice ==================================
Copyright (c) 2017 Intel Corporation
Permission is hereby granted, free of charge, to any person obtaining a
copy of this software and associated documentation files (the
"Software"), to deal in the Software without restriction, including
without limitation the rights to use, copy, modify, merge, publish,
distribute, sublicense, and/or sell copies of the Software, and to
permit persons to whom the Software is furnished to do so, subject to
the following conditions:
The above copyright notice and this permission notice shall be included
in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
======================= end_copyright_notice ==================================*/
#include "FlowGraph.h"
#include "GraphColor.h"
#include "SpillManagerGMRF.h"
#include <list>
#include "SpillCleanup.h"
uint32_t computeFillMsgDesc(unsigned int payloadSize, unsigned int offset);
uint32_t computeSpillMsgDesc(unsigned int payloadSize, unsigned int offset);
namespace vISA
{
G4_SrcRegRegion* CoalesceSpillFills::generateCoalescedSpill(unsigned int scratchOffset, unsigned int payloadSize,
G4_SendMsgDescriptor* sample, bool useNoMask, G4_InstOption mask, int srcCISAOff, G4_Declare* spillDcl, unsigned int row)
{
// Generate split send instruction with specified payload size and offset
auto header = kernel.fg.builder->createSrcRegRegion(Mod_src_undef, Direct,
kernel.fg.builder->getBuiltinR0()->getRegVar(), 0, 0,
kernel.fg.builder->getRegionStride1(), Type_UD);
auto spillSrcPayload = kernel.fg.builder->createSrcRegRegion(Mod_src_undef, Direct, spillDcl->getRegVar(),
(short)row, 0, kernel.fg.builder->getRegionStride1(), Type_UD);
uint32_t spillMsgDesc = computeSpillMsgDesc(payloadSize, scratchOffset);
G4_SendMsgDescriptor* msgDesc = kernel.fg.builder->createSendMsgDesc(spillMsgDesc & 0x000FFFFFu,
0, 1, SFID_DP_DC, false, payloadSize, 0, false, true,
sample->getBti(), sample->getSti());
G4_Imm* msgDescImm = kernel.fg.builder->createImm(msgDesc->getDesc(), Type_UD);
G4_Imm* extDesc = kernel.fg.builder->createImm(msgDesc->getExtendedDesc(), Type_UD);
// Create send instruction with payloadSize starting at scratch offset min
unsigned int option = useNoMask ? InstOpt_WriteEnable : 0;
auto spillInst = kernel.fg.builder->createSplitSendInst(nullptr, G4_sends, 16,
kernel.fg.builder->createNullDst(Type_UW), header, spillSrcPayload, msgDescImm, option, msgDesc, extDesc);
if (!useNoMask)
{
// Care about mask only for non-NoMask sends
spillInst->setMaskOption(mask);
}
spillInst->setCISAOff(srcCISAOff);
#if 0
spillInst->dump();
#endif
return spillSrcPayload;
}
G4_DstRegRegion* CoalesceSpillFills::generateCoalescedFill(unsigned int scratchOffset, unsigned int payloadSize,
unsigned int dclSize, G4_SendMsgDescriptor* sample, int srcCISAOff, G4_Align alignment)
{
// Generate split send instruction with specified payload size and offset
// Construct fillDst
char* dclName = kernel.fg.builder->getNameString(kernel.fg.mem, 32,
"COAL_FILL_%d", kernel.Declares.size());
auto fillDcl = kernel.fg.builder->createDeclareNoLookup(dclName, G4_GRF,
8, (unsigned short)dclSize, Type_UD, DeclareType::CoalescedFill);
fillDcl->setAlign(alignment);
fillDcl->setDoNotSpill();
auto fillDst = kernel.fg.builder->createDstRegRegion(Direct, fillDcl->getRegVar(), 0,
0, 1, Type_UW);
auto header = kernel.fg.builder->createSrcRegRegion(Mod_src_undef, Direct,
kernel.fg.builder->getBuiltinR0()->getRegVar(), 0, 0,
kernel.fg.builder->getRegionStride1(), Type_UD);
uint32_t fillMsgDesc = computeFillMsgDesc(payloadSize, scratchOffset);
G4_SendMsgDescriptor* msgDesc = kernel.fg.builder->createSendMsgDesc(fillMsgDesc,
payloadSize, 1, sample->getFuncId(), false, 0, sample->getExtFuncCtrl(), true, false,
sample->getBti(), sample->getSti());
G4_Imm* msgDescImm = kernel.fg.builder->createImm(msgDesc->getDesc(), Type_UD);
G4_Imm* extDesc = kernel.fg.builder->createImm(sample->getExtendedDesc(), Type_UD);
// Create send instruction with payloadSize starting at scratch offset min
auto fillInst = kernel.fg.builder->createSendInst(nullptr, G4_send, 16,
fillDst, header, extDesc, msgDescImm, InstOpt_WriteEnable, true, false, nullptr);
fillInst->setCISAOff(srcCISAOff);
#if 0
fillInst->dump();
#endif
return fillDst;
}
void CoalesceSpillFills::copyToOldFills(G4_DstRegRegion* coalescedFillDst, std::list<std::pair<G4_DstRegRegion*, std::pair<unsigned int, unsigned int>>> indFills,
INST_LIST_ITER f, G4_BB* bb, int srcCISAOff)
{
// Copy data from coalesced fill in to older fills.
// This way we dont carry entire coalesced payload
// till last fill.
for (auto oldFill : indFills)
{
unsigned int numGRFs = (oldFill.first->getRightBound() - oldFill.first->getLeftBound()
+ G4_GRF_REG_NBYTES - 1) / G4_GRF_REG_NBYTES;
unsigned int rowOff = 0;
// TODO: Check for > 2 GRF dst
while (numGRFs > 0)
{
unsigned int simdSize = 8;
unsigned int off = oldFill.second.first;
unsigned int size = oldFill.second.second;
unsigned int scratchOff = coalescedFillDst->getInst()->getMsgDesc()->getScratchRWOffset();
// Scratch msg offset is always equal or lower than individual fills
unsigned int offToUse = off - scratchOff + rowOff;
if (size > 8)
simdSize = 16;
G4_DstRegRegion* movDst = kernel.fg.builder->createDstRegRegion(Direct,
oldFill.first->getBase(), (short)rowOff, 0, 1, Type_UD);
G4_SrcRegRegion* src = kernel.fg.builder->createSrcRegRegion(Mod_src_undef, Direct,
coalescedFillDst->getBase(), (short)offToUse, 0, kernel.fg.builder->getRegionStride1(), Type_UD);
G4_INST* copy = kernel.fg.builder->createInternalInst(nullptr, G4_mov, nullptr, false, (unsigned char)simdSize,
movDst, src, nullptr, InstOpt_WriteEnable);
copy->setCISAOff(srcCISAOff);
bb->instList.insert(f, copy);
numGRFs -= simdSize == 8 ? 1 : 2;
rowOff += simdSize == 8 ? 1 : 2;
}
}
}
G4_Declare* CoalesceSpillFills::createCoalescedSpillDcl(unsigned int payloadSize)
{
// Construct spill src
char* dclName = nullptr;
G4_Declare* spillDcl = nullptr;
dclName = kernel.fg.builder->getNameString(kernel.fg.mem, 32,
"COAL_SPILL_%d", kernel.Declares.size());
spillDcl = kernel.fg.builder->createDeclareNoLookup(dclName, G4_GRF,
8, (unsigned short)payloadSize, Type_UD, DeclareType::CoalescedSpill);
spillDcl->setDoNotSpill();
return spillDcl;
}
void CoalesceSpillFills::coalesceSpills(std::list<INST_LIST_ITER>& coalesceableSpills, unsigned int min,
unsigned int max, bool useNoMask, G4_InstOption mask, G4_BB* bb, int srcCISAOff)
{
// Generate fill with minimum size = max-min. This should be compatible with
// payload sizes supported by hardware.
unsigned int payloadSize = (max - min) + 1;
MUST_BE_TRUE(payloadSize == 1 || payloadSize == 2 || payloadSize == 4 || payloadSize == 8,
"Unsupported payload size");
std::set<G4_Declare*> declares;
unsigned int minRow = UINT_MAX;
for (auto d : coalesceableSpills)
{
auto src1Opnd = (*d)->getSrc(1)->asSrcRegRegion();
auto curRow = src1Opnd->getLeftBound() / G4_GRF_REG_NBYTES;
declares.insert(src1Opnd->getTopDcl());
minRow = minRow > curRow ? curRow : minRow;
}
G4_Declare* dcl = nullptr;
if (declares.size() == 1)
{
dcl = (*declares.begin());
}
else
{
dcl = createCoalescedSpillDcl(payloadSize);
minRow = 0;
}
auto coalescedSpillSrc = generateCoalescedSpill(min,
payloadSize, (*coalesceableSpills.front())->getMsgDesc(),
useNoMask, mask, srcCISAOff, dcl, minRow);
if (declares.size() != 1)
{
for (auto c : coalesceableSpills)
{
unsigned int scratchOffset, scratchSize;
getScratchMsgInfo((*c), scratchOffset, scratchSize);
MUST_BE_TRUE((*c)->getSrc(0)->asSrcRegRegion()->getTopDcl() == kernel.fg.builder->getBuiltinR0(),
"Unexpected src0");
unsigned int rowOff = scratchOffset - min;
replaceMap.insert(std::make_pair((*c)->getSrc(1)->getTopDcl(),
std::make_pair(coalescedSpillSrc->getTopDcl(), rowOff)));
}
}
auto f = coalesceableSpills.back();
f++;
for (auto spill : coalesceableSpills)
{
bb->instList.erase(spill);
}
coalesceableSpills.clear();
auto copyIt = bb->instList.insert(f, coalescedSpillSrc->getInst());
}
void CoalesceSpillFills::coalesceFills(std::list<INST_LIST_ITER>& coalesceableFills, unsigned int min,
unsigned int max, G4_BB* bb, int srcCISAOff)
{
// Generate fill with minimum size = max-min. This should be compatible with
// payload sizes supported by hardware.
unsigned int payloadSize = (max - min) + 1;
if (payloadSize == 3)
payloadSize = 4;
else if (payloadSize > 4)
payloadSize = 8;
else if (payloadSize == 0)
payloadSize = 1;
MUST_BE_TRUE(payloadSize == 1 || payloadSize == 2 || payloadSize == 4 || payloadSize == 8,
"Unsupported payload size");
// dclSize could be larger than payload size when
// 2 variables across scratch writes are coalesced.
unsigned int dclSize = payloadSize;
for (auto c : coalesceableFills)
{
unsigned int scratchOffset, scratchSize;
auto fill = (*c);
getScratchMsgInfo((*c), scratchOffset, scratchSize);
auto fillDst = fill->getDst();
auto fillDstRegOff = fillDst->getRegOff();
unsigned int dstDclRows = fillDst->getTopDcl()->getNumRows();
unsigned int maxRow = dstDclRows + scratchOffset - fillDstRegOff - min;
if (maxRow > dclSize)
dclSize = maxRow;
}
auto leadInst = *coalesceableFills.front();
auto coalescedFillDst = generateCoalescedFill(min, payloadSize, dclSize,
leadInst->getMsgDesc(), srcCISAOff, leadInst->getDst()->getTopDcl()->getAlign());
for (auto c : coalesceableFills)
{
unsigned int scratchOffset, scratchSize;
getScratchMsgInfo((*c), scratchOffset, scratchSize);
unsigned int rowOff = scratchOffset - min;
replaceMap.insert(std::make_pair((*c)->getDst()->getTopDcl(),
std::make_pair(coalescedFillDst->getTopDcl(), rowOff)));
}
auto f = coalesceableFills.front();
f++;
for (auto fill : coalesceableFills)
{
if (fill == f)
{
f++;
}
bb->instList.erase(fill);
}
coalesceableFills.clear();
auto copyIt = bb->instList.insert(f, coalescedFillDst->getInst());
// Insert pseudo kill for coalesced range
auto pseudoKill = kernel.fg.builder->createInternalInst(nullptr, G4_pseudo_kill, nullptr,
false, 1, kernel.fg.builder->createDstRegRegion(*coalescedFillDst), nullptr, nullptr, 0);
bb->instList.insert(copyIt, pseudoKill);
// copyToOldFills(coalescedFillDst, indFills, f, bb, srcCISAOff);
}
// Return true if heuristic agrees to coalescing.
bool CoalesceSpillFills::fillHeuristic(std::list<INST_LIST_ITER>& coalesceableFills,
std::list<INST_LIST_ITER>& instList, const std::list<INST_LIST_ITER>& origInstList,
unsigned int& min, unsigned int& max)
{
#if 0
std::bitset<8> bits(0);
MUST_BE_TRUE(cMaxFillPayloadSize == 8, "Handle other max fill payload size");
#else
std::bitset<4> bits(0);
MUST_BE_TRUE(cMaxFillPayloadSize == 4, "Handle other max fill payload size");
#endif
if (coalesceableFills.size() <= 1)
{
return false;
}
min = 0xffffffff, max = 0;
for (auto f : coalesceableFills)
{
unsigned int scratchOffset, scratchSize;
getScratchMsgInfo(*f, scratchOffset, scratchSize);
if (scratchSize == 8)
{
return false;
}
if (addrTakenSpillFillDcl.find((*f)->getDst()->getTopDcl()) !=
addrTakenSpillFillDcl.end())
{
return false;
}
for (auto i = scratchOffset; i < (scratchOffset + scratchSize); i++)
bits.set(i - scratchOffset);
if (min > scratchOffset)
min = scratchOffset;
if (max < (scratchOffset + scratchSize - 1))
max = (scratchOffset + scratchSize - 1);
}
// Iterate over coalescable fills and ensure all rows of a variable
// are fill candidates. If not, then dont fill. This helps cases like,
// #1 FILL_V10(0,0) <-- load 0x10 ... (4 GRFs)
// #2 FILL_V10(4,0) <-- load 0x14 ... (1 GRF)
// #3 send ... FILL_V10(0,0) ... (use 3 GRFs of FILL_V10)
// #4 FILL_V11(0,0) <-- load 0x15 ... (1 GRF)
//
// Loads at #2 and #4 can be coalesced. But this requires a new coalesced
// variable of size = 6 GRFs. This size can quickly increase for Cm where
// payloads of 8 GRF are also present. So instead of allowing these cases
// at the risk of spilling more, we require that all rows of fill range
// are candidates in coalesceableFills list. So when only #2 and #4 are
// fill candidates, we will not coalesce them. This makes us miss some
// cases where subset rows of fills have been converted to movs.
const int maxDclSize = 128;
std::map<G4_Declare*, std::bitset<maxDclSize>> allRows;
for (auto c : coalesceableFills)
{
auto topdcl = (*c)->getDst()->getTopDcl();
unsigned int scratchOffset, scratchSize;
getScratchMsgInfo(*c, scratchOffset, scratchSize);
auto it = allRows.find(topdcl);
if (it == allRows.end())
{
allRows.insert(std::make_pair(topdcl, std::bitset<maxDclSize>()));
it = allRows.find(topdcl);
}
// Now mark bits corresponding to rows
unsigned int regOff = (*c)->getDst()->getRegOff();
for (unsigned int r = regOff;
r < (regOff+scratchSize); r++)
{
it->second.set(r);
}
}
// Check whether all dcls in map have all rows filled
for (auto&& r : allRows)
{
unsigned int numRows = r.first->getNumRows();
for (unsigned int i = 0; i < numRows; i++)
{
if (r.second.test(i) == false)
{
// Found a row of variable that isnt captured in
// list of candidate fills.
return false;
}
}
}
#if 0
for (auto f : coalesceableFills)
{
unsigned int scratchOffset, scratchSize;
getScratchMsgInfo(*f, scratchOffset, scratchSize);
for (auto i = scratchOffset; i < (scratchOffset + scratchSize); i++)
bits.set(i - min);
}
#endif
if (max - min <= 3)
{
// Will emit at most 4GRF read
if (bits[0] != bits[1] &&
bits[2] != bits[3])
{
// Dont coalesce patterns like
// 1010, 0101
return false;
}
if ((bits[0] & bits[3]) &&
!(bits[1] | bits[2]))
{
// 1001
return false;
}
}
return true;
}
// instList contains all instructions (fills or spills) within window size.
// At exit, instList will contain instructions that will not be coalesced.
// coalescable list will contain instructions within min-max offset range.
// First instruction's offset in instList is set to be min. max is
// min + maxPayloadSize - 1.
void CoalesceSpillFills::sendsInRange(std::list<INST_LIST_ITER>& instList,
std::list<INST_LIST_ITER>& coalescable,
unsigned int maxPayloadSize, unsigned int& min, unsigned int& max)
{
min = 0xffffffff;
max = 0;
bool isFirstNoMask = false;
unsigned int mask = 0;
for (auto iter = instList.begin();
iter != instList.end();
)
{
unsigned scratchOffset, sizeInGrfUnit, lastScratchOffset;
auto inst = *(*iter);
scratchOffset = inst->getMsgDesc()->getScratchRWOffset();
sizeInGrfUnit = inst->getMsgDesc()->getScratchRWSize();
lastScratchOffset = scratchOffset + sizeInGrfUnit - 1;
if (min == 0xffffffff && max == 0)
{
// First spill is definitely a candidate
min = scratchOffset;
max = lastScratchOffset;
coalescable.push_back(*iter);
iter = instList.erase(iter);
isFirstNoMask = inst->isWriteEnableInst();
mask = inst->getMaskOption();
if (addrTakenSpillFillDcl.find(inst->getDst()->getTopDcl()) !=
addrTakenSpillFillDcl.end())
{
return;
}
continue;
}
if (min != 0xffffffff || max != 0)
{
bool maskMatch = (isFirstNoMask && inst->isWriteEnableInst()) ||
(mask == inst->getMaskOption());
// don't coalesce if non-leading fill inst has alignment requirements,
// as we may not be able to satisfy it
bool fillDstisAligned = inst->getDst()->getTopDcl()->getAlign() != Either;
if (!maskMatch || fillDstisAligned)
{
iter++;
continue;
}
// Check whether min/max can be extended
if (scratchOffset <= min &&
(min - scratchOffset) <= (cMaxFillPayloadSize - 1) &&
(max - scratchOffset) <= (cMaxFillPayloadSize - 1))
{
// This instruction can be coalesced
min = scratchOffset;
if (max < lastScratchOffset)
max = lastScratchOffset;
//MUST_BE_TRUE(max - min <= (cMaxFillPayloadSize - 1), "Unexpected fills coalesced. (max - min) is out of bounds - 1");
coalescable.push_back(*iter);
iter = instList.erase(iter);
}
else if (scratchOffset >= max &&
(lastScratchOffset - min) <= (cMaxFillPayloadSize - 1) &&
(lastScratchOffset - max) <= (cMaxFillPayloadSize - 1))
{
max = lastScratchOffset;
//MUST_BE_TRUE(max - min <= cMaxFillPayloadSize, "Unexpected spills coalesced. (max - min) is out of bounds - 2");
coalescable.push_back(*iter);
iter = instList.erase(iter);
}
else if (scratchOffset >= min &&
lastScratchOffset <= max)
{
coalescable.push_back(*iter);
iter = instList.erase(iter);
}
else
{
iter++;
}
}
}
}
// instList contains all spills seen in window.
// coalescable is empty and should contain consecutive spills.
// This funtion will prune spills so they write consecutive
// memory slots. First spill is first candidate to start window.
void CoalesceSpillFills::keepConsecutiveSpills(std::list<INST_LIST_ITER>& instList,
std::list<INST_LIST_ITER>& coalescable,
unsigned int maxPayloadSize, unsigned int& minOffset, unsigned int& maxOffset,
bool& useNoMask, G4_InstOption& mask)
{
// allowed list contains instructions to be coalesced in
// ascending order of their spill slots.
std::list<INST_LIST_ITER> allowed;
auto origInstList = instList;
allowed.push_back(instList.front());
instList.pop_front();
unsigned int maskOffset = (*allowed.front())->getMaskOption();
mask = (G4_InstOption)(maskOffset & InstOpt_QuarterMasks);
useNoMask = (maskOffset & InstOpt_WriteEnable) ? true : false;
unsigned int size;
getScratchMsgInfo(*allowed.front(), minOffset, size);
maxOffset = minOffset + size - 1;
bool firstSpillFromSend = false;
G4_Declare* sendDstTopDcl = (*allowed.front())->getSrc(1)->getTopDcl();
if (sendDstDcl.find(sendDstTopDcl) != sendDstDcl.end())
firstSpillFromSend = true;
for (auto instIt : instList)
{
auto inst = (*instIt);
useNoMask &= inst->isWriteEnableInst();
if (!useNoMask)
break;
}
if (useNoMask)
{
// Spill coalescing doesnt work as expected without NoMask
bool redo;
do
{
redo = false;
for (auto spillIt = instList.begin();
spillIt != instList.end();
spillIt++)
{
unsigned int scratchOffset, scratchSize;
getScratchMsgInfo(*(*spillIt), scratchOffset, scratchSize);
auto src1 = (*(*spillIt))->getSrc(1);
if (src1 &&
addrTakenSpillFillDcl.find(src1->getTopDcl()) !=
addrTakenSpillFillDcl.end())
{
// Address taken dcls should not be coalesed with others.
// This is dangerous because nothing ties indirect opnd
// with fill/spill instructions for it. Only after RA do
// we update offset of address register holding the
// indirect operand, based on RA assigment to spill/fill
// address taken variable.
continue;
}
if (// Consecutive scratch offsets
scratchOffset == maxOffset + 1 &&
// Scratch offset + size is within max payload size
(scratchOffset + scratchSize - 1) <= (minOffset + maxPayloadSize - 1) &&
// Either both masks are same or both are NoMask
(((*(*spillIt))->getMaskOption() == (maskOffset & InstOpt_QuarterMasks)) ||
(useNoMask && (*(*spillIt))->isWriteEnableInst())))
{
auto curInstDstTopDcl = (*(*spillIt))->getSrc(1)->getTopDcl();
// Check whether current inst's topdcl was spilled in a send.
// If it was and first instruction in instList wasnt then
// dont consider current instruction as coalescing candidate.
if (!firstSpillFromSend &&
sendDstDcl.find(curInstDstTopDcl) != sendDstDcl.end())
{
continue;
}
// This condition allows send coalescing iff
// a. Either none of the vars are defined in a send
// b. All vars defined in same send
if (!firstSpillFromSend ||
curInstDstTopDcl == sendDstTopDcl)
{
if (curInstDstTopDcl == sendDstTopDcl)
{
// Make sure src1 operands are consecutive
auto curSrc1Row = (*(*spillIt))->getSrc(1)->asSrcRegRegion()->getRegOff();
bool success = true;
for (auto candidate : allowed)
{
unsigned int candOffset, candSize;
getScratchMsgInfo(*candidate, candOffset, candSize);
auto prevSrc1Row = (*candidate)->getSrc(1)->asSrcRegRegion()->getRegOff();
unsigned int scratchOffDelta = scratchOffset - candOffset;
if ((prevSrc1Row + scratchOffDelta) != curSrc1Row)
{
// Following is disallowed
// send (8) V10(1,0) ... <-- resLen = 4
// sends (8) null r0 V10(1,0) 0x100 <-- extLen = 1
// mov (8) T2 V10(2,0)
// sends (8) null r0 r10(3,0) 0x101 <-- extLen = 1
// mov (8) T4 V10(4,0)
// Two scratch writes cannot be coalesced here
// because their src1 regions arent consecutive.
success = false;
break;
}
}
if (!success)
continue;
}
allowed.push_back(*spillIt);
instList.erase(spillIt);
redo = true;
maxOffset += scratchSize;
break;
}
}
}
} while (redo);
}
while (allowed.size() > 1)
{
unsigned int slots = maxOffset - minOffset + 1;
if (slots == 2 || slots == 4)
{
// Insert coalescable spills in order of appearance
for (auto origInst : origInstList)
{
for (auto allowedSpills : allowed)
{
if (*origInst == *allowedSpills)
{
coalescable.push_back(origInst);
break;
}
}
}
MUST_BE_TRUE(coalescable.size() == allowed.size(),
"Coalesced spills list missing entries");
break;
}
else
{
allowed.pop_back();
unsigned int scratchOffset, scratchSize;
getScratchMsgInfo(*allowed.back(), scratchOffset, scratchSize);
maxOffset = scratchOffset + scratchSize - 1;
}
}
instList = origInstList;
for (auto coalIt = coalescable.begin(),
instIt = instList.begin();
coalIt != coalescable.end();
coalIt++)
{
if (*instIt == *coalIt)
instIt = instList.erase(instIt);
else
{
while (*instIt != *coalIt)
{
instIt++;
}
instIt = instList.erase(instIt);
}
}
}
INST_LIST_ITER CoalesceSpillFills::analyzeSpillCoalescing(std::list<INST_LIST_ITER>& instList,
INST_LIST_ITER start, INST_LIST_ITER end, G4_BB* bb)
{
// Check and perform coalescing, if possible, amongst spills in instList.
// Return inst iter points to either last inst+1 in instList if all spills
// were coalesced. Otherwise, it points to first spill that wasnt coalesced.
// Spill coalescing is possible only when all slots in coalesced range
// have a write.
INST_LIST_ITER last = end;
last++;
#if 0
unsigned int startCISAOff = (*instList.front())->getCISAOff();
#endif
if (instList.size() < 2)
{
return last;
}
std::list<INST_LIST_ITER> coalesceableSpills;
auto origInstList = instList;
unsigned int min, max;
G4_InstOption mask;
bool useNoMask;
keepConsecutiveSpills(instList, coalesceableSpills, cMaxSpillPayloadSize, min, max, useNoMask, mask);
#if 0
printf("Start -- \n");
if (coalesceableSpills.size() > 0)
{
printf("Will coalesce following spill (offset, size) pairs:\n");
for (auto k : coalesceableSpills)
{
printf("(%d, %d) @ $%d,\t", (*k)->getMsgDesc()->getScratchRWOffset(), (*k)->getMsgDesc()->getScratchRWSize(), (*k)->getCISAOff());
}
printf("\n\n");
}
if (instList.size() > 0)
{
printf("Will NOT coalesce following spill (offset, size) pairs:\n");
for (auto k : instList)
{
printf("(%d, %d) @ $%d,\t", (*k)->getMsgDesc()->getScratchRWOffset(), (*k)->getMsgDesc()->getScratchRWSize(), (*k)->getCISAOff());
}
printf("\n\n");
}
printf("End --\n");
#endif
if (coalesceableSpills.size() > 1)
{
coalesceSpills(coalesceableSpills, min, max, useNoMask, mask, bb, (*coalesceableSpills.front())->getCISAOff());
}
else
{
// When coalescing is not done, we want to
// move to second instruction in instList in
// next loop iteration.
instList.pop_front();
}
if (instList.size() == 0)
{
return last;
}
else
{
return instList.front();
}
}
INST_LIST_ITER CoalesceSpillFills::analyzeFillCoalescing(std::list<INST_LIST_ITER>& instList,
INST_LIST_ITER start, INST_LIST_ITER end, G4_BB* bb)
{
// Check and perform coalescing, if possible, amongst fills in instList.
// Return inst iter points to either last inst+1 in instList if all fills
// were coalesced. Otherwise, it points to first fill that wasnt coalesced.
INST_LIST_ITER last = end;
last++;
#if 0
G4_INST* lastInst = nullptr;
if (last != bb->instList.end())
lastInst = (*last);
#endif
if (instList.size() < 2)
{
return last;
}
std::list<INST_LIST_ITER> coalesceableFills;
auto origInstList = instList;
unsigned int min, max;
sendsInRange(instList, coalesceableFills, cMaxFillPayloadSize, min, max);
bool heuristic = fillHeuristic(coalesceableFills, instList, origInstList, min, max);
if (!heuristic)
{
coalesceableFills.clear();
instList = origInstList;
instList.pop_front();
#if 0
printf("Fill heuristic didnt agree to coalescing\n");
#endif
}
#if 0
printf("Start -- \n");
if (coalesceableFills.size() > 0)
{
printf("Will coalesce following fill (offset, size) pairs:\n");
for (auto k : coalesceableFills)
{
printf("(%d, %d) @ $%d,\t", (*k)->getMsgDesc()->getScratchRWOffset(), (*k)->getMsgDesc()->getScratchRWSize(), (*k)->getCISAOff());
}
printf("\n\n");
}
if (instList.size() > 0)
{
printf("Will NOT coalesce following fill (offset, size) pairs:\n");
for (auto k : instList)
{
printf("(%d, %d) @ $%d,\t", (*k)->getMsgDesc()->getScratchRWOffset(), (*k)->getMsgDesc()->getScratchRWSize(), (*k)->getCISAOff());
}
printf("\n\n");
}
printf("End --\n");
#endif
if (coalesceableFills.size() > 1)
{
coalesceFills(coalesceableFills, min, max, bb, (*coalesceableFills.front())->getCISAOff());
}
if (instList.size() == 0)
{
return last;
}
else
{
return instList.front();
}
}
bool CoalesceSpillFills::overlap(G4_INST* inst1, G4_INST* inst2, bool& isFullOverlap)
{
unsigned int scratchOffset1, scratchSize1, scratchOffset2, scratchSize2;
unsigned int scratchEnd1, scratchEnd2;
getScratchMsgInfo(inst1, scratchOffset1, scratchSize1);
getScratchMsgInfo(inst2, scratchOffset2, scratchSize2);
// isFullOverlap is true only if inst1 full covers inst2
isFullOverlap = false;
scratchEnd1 = scratchOffset1 + scratchSize1 - 1;
scratchEnd2 = scratchOffset2 + scratchSize2 - 1;
if (scratchOffset1 <= scratchOffset2)
{
// inst1 |---------| or |----------|
// inst2 |------| |---|
if (scratchEnd1 >= scratchOffset2)
{
if (scratchOffset1 <= scratchOffset2 &&
(scratchOffset1 + scratchSize1) >= (scratchOffset2 + scratchSize2))
{
isFullOverlap = true;
}
return true;
}
}
else
{
// inst1 |------| or |-----|
// inst2 |-----| |-----------|
if (scratchEnd2 >= scratchOffset1)
{
if (scratchOffset1 <= scratchOffset2 &&
(scratchOffset1 + scratchSize1) >= (scratchOffset2 + scratchSize2))
{
isFullOverlap = true;
}
return true;
}
}
return false;
}
bool CoalesceSpillFills::overlap(G4_INST* inst, std::list<INST_LIST_ITER>& allInsts)
{
for (auto sp : allInsts)
{
bool t;
auto spillInst = (*sp);
if (overlap(inst, spillInst, t))
return true;
}
return false;
}
void CoalesceSpillFills::removeWARFills(std::list<INST_LIST_ITER>& fills, std::list<INST_LIST_ITER>& spills)
{
for (auto flIt = fills.begin();
flIt != fills.end();
)
{
if (overlap((*(*flIt)), spills))
{
flIt = fills.erase(flIt);
continue;
}
flIt++;
}
}
void CoalesceSpillFills::replaceCoalescedOperands(G4_INST* inst)
{
auto dst = inst->getDst();
if (dst &&
dst->getTopDcl())
{
auto dcl = dst->getTopDcl();
auto it = replaceMap.find(dcl);
if (it != replaceMap.end())
{
auto dstRgn = dst->asDstRegRegion();
auto newDstRgn = kernel.fg.builder->createDstRegRegion(Direct, it->second.first->getRegVar(),
it->second.second + dstRgn->getRegOff(), dstRgn->getSubRegOff(), dstRgn->getHorzStride(), dstRgn->getType());
inst->setDest(newDstRgn);
}
}
for (unsigned int i = 0; i < G4_MAX_SRCS; i++)
{
auto opnd = inst->getSrc(i);
if (opnd &&
opnd->getTopDcl())
{
auto dcl = opnd->getTopDcl();
auto it = replaceMap.find(dcl);
if (it == replaceMap.end())
continue;
if (opnd->isSrcRegRegion())
{
auto srcRgn = opnd->asSrcRegRegion();
auto oldRgnDesc = srcRgn->getRegion();
auto newSrcRgn = kernel.fg.builder->createSrcRegRegion(srcRgn->getModifier(), Direct,
it->second.first->getRegVar(), it->second.second + srcRgn->getRegOff(),
srcRgn->getSubRegOff(), oldRgnDesc,
opnd->getType());
inst->setSrc(newSrcRgn, i);
}
}
}
}
void CoalesceSpillFills::insertKill(G4_BB* bb, INST_LIST_ITER instIt, std::set<G4_Declare*>& coalescedRangeKills)
{
// Check whether current instruction is first reference of coalesced
// spill range. If it is then insert pseudo_kill for coalesced var
// before instIt.
auto inst = (*instIt);
auto dst = inst->getDst();
if (dst)
{
auto topdcl = dst->getTopDcl();
if (topdcl)
{
auto entry = replaceMap.find(topdcl);
if (entry != replaceMap.end() &&
coalescedRangeKills.find(entry->second.first) == coalescedRangeKills.end())
{
// This means topdcl is first reference of to-be-replaced
// dcl. Insert pseudokill for coalesced dcl here.
auto killDst = kernel.fg.builder->createDstRegRegion(Direct,
entry->second.first->getRegVar(), 0, 0, 1, Type_UD);
auto kill = kernel.fg.builder->createInternalInst(nullptr,
G4_pseudo_kill, nullptr, false, 1, killDst, nullptr, nullptr, 0);
bb->instList.insert(instIt, kill);
coalescedRangeKills.insert(entry->second.first);
}
}
}
}
bool CoalesceSpillFills::allSpillsSameVar(std::list<INST_LIST_ITER>& spills)
{
// Return true if all vars in spills list have same dcl
G4_Declare* dcl = nullptr;
for (auto s : spills)
{
auto topdcl = (*s)->getSrc(1)->getTopDcl();
if (!dcl)
dcl = topdcl;
if (topdcl != dcl)
{
return false;
}
}
// Allow only if all dcls are defined by same send
if (sendDstDcl.find(dcl) != sendDstDcl.end())
return true;
return false;
}
void CoalesceSpillFills::fills()
{
// Iterate over all BBs, find fills that are closeby and coalesce
// a bunch of them. Insert movs as required.
for (auto bb : kernel.fg.BBs)
{
auto endIter = bb->instList.end();
std::list<INST_LIST_ITER> fillsToCoalesce;
std::list<INST_LIST_ITER> spills;
INST_LIST_ITER startIter = bb->instList.begin();
unsigned int w = 0;
for (auto instIter = startIter;
instIter != endIter;)
{
auto inst = (*instIter);
if (inst->isPseudoKill() ||
inst->isLabel())
{
instIter++;
continue;
}
if (inst->isSend())
{
if (inst->getMsgDesc()->isScratchWrite())
{
spills.push_back(instIter);
}
else if (inst->getMsgDesc()->isScratchRead())
{
// Check if coalescing is possible
if (fillsToCoalesce.size() == 0)
{
w = 0;
startIter = instIter;
spills.clear();
}
if (!overlap(*instIter, spills))
{
fillsToCoalesce.push_back(instIter);
}
}
}
if (fillsToCoalesce.size() > 0 &&
rpe.getRegisterPressure(inst) > 180)
{
// High register pressure region so reduce window size to 3
w = (cWindowSize - w > 3) ? cWindowSize - 3 : w;
}
if (w == cWindowSize || inst == bb->instList.back())
{
if (fillsToCoalesce.size() > 1)
{
instIter = analyzeFillCoalescing(fillsToCoalesce, startIter, instIter, bb);
}
else if (w == cWindowSize)
{
startIter = instIter;
}
else if (inst == bb->instList.back())
{
break;
}
w = 0;
fillsToCoalesce.clear();
spills.clear();
continue;
}
if (fillsToCoalesce.size() > 0)
{
w++;
}
instIter++;
}
// One pass to replace old fills with coalesced dcl
for (auto instIt = bb->instList.begin();
instIt != bb->instList.end();
)
{
auto inst = (*instIt);
if (inst->isPseudoKill() &&
replaceMap.find(inst->getDst()->getTopDcl()) != replaceMap.end())
{
instIt = bb->instList.erase(instIt);
continue;
}
replaceCoalescedOperands(inst);
instIt++;
}
}
}
void CoalesceSpillFills::populateSendDstDcl()
{
// Find and store all G4_Declares that are dest in sends
// and are spilled. This is required when coalescing
// scratch writes for such spills. We cannot mix coalescing
// for G4_Declares from one and and other instructions.
// Otherwise register pressure increases significantly.
for (auto bb : kernel.fg.BBs)
{
for (auto inst : bb->instList)
{
if (inst->isSend() &&
inst->getDst())
{
if (!inst->getDst()->isNullReg())
{
if (!inst->getMsgDesc()->isScratchRW())
{
auto topdcl = inst->getDst()->getTopDcl();
sendDstDcl.insert(topdcl);
}
}
else if (inst->getMsgDesc()->isScratchWrite() &&
inst->getSrc(1)->getBase()->asRegVar()->isRegVarCoalesced())
{
auto topdcl = inst->getSrc(1)->getTopDcl();
sendDstDcl.insert(topdcl);
}
}
}
}
}
void CoalesceSpillFills::spills()
{
populateSendDstDcl();
// Iterate over all BBs, find fills that are closeby and coalesce
// a bunch of them. Insert movs as required.
for (auto bb : kernel.fg.BBs)
{
auto endIter = bb->instList.end();
std::list<INST_LIST_ITER> spillsToCoalesce;
INST_LIST_ITER startIter = bb->instList.begin();
unsigned int w = 0;
for (auto instIter = startIter;
instIter != endIter;)
{
auto inst = (*instIter);
if (inst->isPseudoKill() ||
inst->isLabel())
{
instIter++;
continue;
}
bool earlyCoalesce = false;
if (inst->isSend())
{
if (inst->getMsgDesc()->isScratchWrite())
{
// Check if coalescing is possible
if (spillsToCoalesce.size() == 0)
{
w = 0;
startIter = instIter;
spillsToCoalesce.clear();
}
for (auto coalIt = spillsToCoalesce.begin();
coalIt != spillsToCoalesce.end();
)
{
bool fullOverlap = false;
if (overlap(*instIter, *(*coalIt), fullOverlap))
{
if (fullOverlap)
{
#if 0
printf("Deleting spill at $%d due to %d\n", (*(*coalIt))->getCISAOff(), (*instIter)->getCISAOff());
#endif
// Delete earlier spill since its made redundant
// by current spill.
bb->instList.erase(*coalIt);
}
coalIt = spillsToCoalesce.erase(coalIt);
continue;
}
coalIt++;
}
spillsToCoalesce.push_back(instIter);
}
else if (inst->getMsgDesc()->isScratchRead())
{
for (auto coalIt = spillsToCoalesce.begin();
coalIt != spillsToCoalesce.end();
)
{
bool temp = false;
if (overlap(*instIter, *(*coalIt), temp))
{
#if 1
// Instead of deleting scratch writes try coalescing
// at this point. This way maybe the fill can also
// be cleaned up in later phase.
earlyCoalesce = true;
break;
#else
coalIt = spillsToCoalesce.erase(coalIt);
continue;
#endif
}
coalIt++;
}
}
}
if (spillsToCoalesce.size() > 0 &&
rpe.getRegisterPressure(inst) > 120)
{
if (!allSpillsSameVar(spillsToCoalesce))
{
// High register pressure region so reduce window size to 3
w = (cWindowSize - w > 3) ? cWindowSize - 3 : w;
}
else
{
#if 0
printf("Found register pressure = %d at %d. Still coalescing spills because all spills are from same var.\n",
rpe.getRegisterPressure(inst), inst->getCISAOff());
#endif
}
}
if (w == cWindowSize || inst == bb->instList.back() ||
earlyCoalesce)
{
if (spillsToCoalesce.size() > 1)
{
instIter = analyzeSpillCoalescing(spillsToCoalesce, startIter, instIter, bb);
}
else if (w == cWindowSize)
{
startIter = instIter;
}
else if (inst == bb->instList.back())
{
break;
}
w = 0;
spillsToCoalesce.clear();
continue;
}
if (spillsToCoalesce.size() > 0)
{
w++;
}
instIter++;
}
std::set<G4_Declare*> coalescedRangeKills;
// One pass to replace old fills with coalesced dcl
for (auto instIt = bb->instList.begin();
instIt != bb->instList.end();
)
{
auto inst = (*instIt);
if (inst->isPseudoKill() &&
replaceMap.find(inst->getDst()->getTopDcl()) != replaceMap.end())
{
instIt = bb->instList.erase(instIt);
continue;
}
insertKill(bb, instIt, coalescedRangeKills);
replaceCoalescedOperands(inst);
instIt++;
}
}
}
void CoalesceSpillFills::fixSendsSrcOverlap()
{
// Overlap for sends src operands is not allowed.
//
// Fix for following code pattern after spill/fill coalescing:
// send (16) COAL_FILL_373(0,0)<1>:ud r0 0xa 0x24c2001:ud{Align1, NoMask} // #??:$365:%657:&-1 // scratch read, resLen=4, msgLen=1
// sends(1) null:ud COAL_FILL_373(0, 0) COAL_FILL_373(1, 0) 0x4c : ud 0x40680ff : ud{ Align1, Q1, NoMask } // #??:$365:%365:&-1 // a64 scatt
// ered write, resLen = 0, msgLen = 2, extMsgLen = 1
//
// for CISA:
// svm_scatter.1.1 (M1_NM, 1) V441.0 V449.0 /// $365
//
// where V441 and V449 are both scalars of type :uq and :ud respectively
//
for (auto bb : kernel.fg.BBs)
{
for (auto instIt = bb->instList.begin();
instIt != bb->instList.end();
instIt++)
{
auto inst = (*instIt);
if (!inst->isSplitSend())
{
continue;
}
auto src0 = inst->getSrc(0);
auto src1 = inst->getSrc(1);
if (src0->getTopDcl() == src1->getTopDcl())
{
auto lb0 = src0->getLeftBound();
auto rb0 = src0->getRightBound();
auto lb1 = src1->getLeftBound();
auto rb1 = src1->getRightBound();
if ((lb0 < lb1 && rb0 > lb1) ||
(lb1 < lb0 && rb1 > lb0))
{
// Ideally we should create copy of
// operand with less number of GRFs,
// but this is a really corner case
// and probably shows up only for
// force spills. So we simply choose
// src1 of sends.
char* dclName = kernel.fg.builder->getNameString(kernel.fg.mem, 32,
"COPY_%d", kernel.Declares.size());
G4_Declare* copyDcl = kernel.fg.builder->createDeclareNoLookup(dclName, G4_GRF,
8, src1->getTopDcl()->getNumRows(),
Type_UD);
unsigned int elems = copyDcl->getNumElems();
short row = 0;
while (elems > 0)
{
G4_SrcRegRegion* srcRgn = kernel.fg.builder->createSrcRegRegion(
Mod_src_undef, Direct, src1->getTopDcl()->getRegVar(), row, 0,
kernel.fg.builder->getRegionStride1(), Type_UD);
G4_DstRegRegion* dstRgn = kernel.fg.builder->createDstRegRegion(
Direct, copyDcl->getRegVar(), row, 0, 1, Type_UD);
G4_INST* copyInst = kernel.fg.builder->createInternalInst(nullptr,
G4_mov, nullptr, false, 8, dstRgn, srcRgn, nullptr, InstOpt_WriteEnable);
copyInst->setCISAOff(inst->getCISAOff());
bb->instList.insert(instIt, copyInst);
elems -= 8;
row++;
}
G4_SrcRegRegion* sendSrc1 = kernel.fg.builder->createSrcRegRegion(Mod_src_undef,
Direct, copyDcl->getRegVar(), 0, 0, kernel.fg.builder->getRegionStride1(),
Type_UD);
inst->setSrc(sendSrc1, 1);
}
}
}
}
}
void CoalesceSpillFills::removeRedundantSplitMovs()
{
// send (8) V_SAMPLER -- resLen = 3
// COAL_0(0,0) = V_SAMPLE(0,0)
// COAL_0(1,0) = V_SAMPLE(1,0)
// send (8) <null> COAL_0(0,0) <-- len = 2
// TV0(0,0) = V_SAMPLE(2,0)
// ===>
// send (8) V_SAMPLER -- resLen = 3
// send (8) <null> V_SAMPLE(0,0) <-- len = 2
// TV0(0,0) = V_SAMPLE(2,0)
// Look for scratch writes. Src1 is data to write to memory.
// Iterate in bottom-order to check whether raw movs exist
// that define src1 of scratch write and whether their
// source operands are consecutive.
// Store numUses for dcls replaced and location of defs.
// This structure is used to eliminate redundant movs
// later.
typedef std::pair<G4_BB*, INST_LIST_ITER> MovLoc;
typedef unsigned int NumRefs;
std::map<G4_Declare*, std::pair<NumRefs, std::list<MovLoc>>> movs;
for (auto bb : kernel.fg.BBs)
{
// Store all dcls defined by non scratch sends
// as only they are candidates for this pass.
// Without this, we might end up identifying
// other raw movs coming from partial write like:
// add (8) r8.0<1>:q r20.0<4;4,1>:q r4.0<0;1,0>:ud {Align1, Q1}
// send(16) r27.0<1>:uw r26 0xa 0x22c1000 : ud{ Align1, NoMask } // scratch read, fill, offset = 0, resLen=2, msgLen=1
// mov(8) r27.0<1> : q r8.0<4; 4, 1> : q{ Align1, Q1 }
// sends(16) null : uw r26 r27 0x8a : ud 0x20f1000 : ud{ Align1, NoMask } // scratch write, spill, offset = 0, resLen=0, msgLen=1, extMsgLen=2
//
// Although there is a raw mov before scratch write,
// it has to be preserved for correctness.
std::set<G4_Declare*> sendDst;
for (auto inst : bb->instList)
{
if (inst->isSend() &&
inst->getDst() &&
!inst->getDst()->isNullReg() &&
!inst->getMsgDesc()->isScratchRead() &&
!inst->getMsgDesc()->isScratchWrite())
{
sendDst.insert(inst->getDst()->getTopDcl());
}
}
for (auto instIt = bb->instList.begin(), endIt = bb->instList.end();
instIt != endIt;
instIt++)
{
auto inst = (*instIt);
if (inst->isSplitSend() &&
inst->getMsgDesc()->isScratchWrite())
{
// Spill sends
auto src1Dcl = inst->getSrc(1)->getTopDcl();
unsigned int lb = inst->getSrc(1)->getLeftBound();
unsigned int rb = inst->getSrc(1)->getRightBound();
std::set<unsigned int> rows;
for (unsigned int k = lb / G4_GRF_REG_NBYTES; k != (rb + G4_GRF_REG_NBYTES - 1) / G4_GRF_REG_NBYTES; k++)
{
rows.insert(k);
}
auto tmpIt = instIt;
tmpIt--;
G4_Declare* srcDcl = nullptr;
std::map<unsigned int, unsigned int> dstSrcRowMapping;
std::list<MovLoc> copies;
while (tmpIt != bb->instList.begin())
{
auto pInst = (*tmpIt);
// Each copy should be a raw mov
if (!pInst->isRawMov())
break;
// Ensure src0 topdcl comes from a send dst in this BB
if (sendDst.find(pInst->getSrc(0)->getTopDcl()) ==
sendDst.end())
break;
// Check whether dcls match
auto pDstDcl = pInst->getDst()->getTopDcl();
if (pDstDcl != src1Dcl)
break;
unsigned int plb = pInst->getDst()->getLeftBound();
unsigned int prb = pInst->getDst()->getRightBound();
// Check whether complete row(s) defined
if ((prb - plb + 1) % G4_GRF_REG_NBYTES != 0)
break;
unsigned int rowStart = plb / G4_GRF_REG_NBYTES;
unsigned int numRows = (prb - plb + 1) / G4_GRF_REG_NBYTES;
bool punt = false;
for (unsigned int k = rowStart; k != (rowStart + numRows); k++)
{
if (rows.find(k) == rows.end())
{
punt = true;
break;
}
dstSrcRowMapping.insert(std::make_pair(k, INT_MAX));
}
if (punt)
break;
auto pSrc0 = pInst->getSrc(0);
if (!pSrc0->isSrcRegRegion())
break;
auto pSrcDcl = pSrc0->getTopDcl();
if (!srcDcl)
srcDcl = pSrcDcl;
else if (srcDcl != pSrcDcl)
break;
// mov src should be GRF aligned
if (pSrc0->getLeftBound() % G4_GRF_REG_NBYTES != 0)
break;
unsigned int src0lb = pSrc0->getLeftBound();
unsigned int src0rb = pSrc0->getRightBound();
// (rb - lb) should match dst (rb - lb)
if ((src0rb - src0lb) != (prb - plb))
break;
unsigned int pStartRow = pSrc0->getLeftBound() / G4_GRF_REG_NBYTES;
for (unsigned int k = rowStart; k != (rowStart + numRows); k++)
{
auto it = dstSrcRowMapping.find(k);
if (it == dstSrcRowMapping.end())
{
punt = true;
break;
}
it->second = pStartRow + (k - rowStart);
}
if (punt)
break;
copies.push_back(std::make_pair(bb, tmpIt));
tmpIt--;
}
if (dstSrcRowMapping.size() > 0)
{
// Now check whether each entry of src1 has a corresponding src offset
unsigned int dstRowStart = lb / G4_GRF_REG_NBYTES;
bool success = true;
auto baseIt = dstSrcRowMapping.find(dstRowStart);
if (baseIt != dstSrcRowMapping.end())
{
auto base = dstSrcRowMapping.find(dstRowStart)->second;
for (auto m : dstSrcRowMapping)
{
unsigned int curRow = m.first - dstRowStart;
if (m.second == INT_MAX)
{
success = false;
break;
}
if (m.second != (base + curRow))
{
success = false;
break;
}
}
if (success && srcDcl)
{
// Replace src1 of send with srcDcl
G4_SrcRegRegion* sendSrc1 = kernel.fg.builder->createSrcRegRegion(Mod_src_undef, Direct, srcDcl->getRegVar(),
(short)base, 0, kernel.fg.builder->getRegionStride1(), inst->getSrc(1)->getType());
inst->setSrc(sendSrc1, 1);
for (auto c : copies)
{
auto defDcl = (*c.second)->getDst()->getTopDcl();
auto it = movs.find(defDcl);
if (it == movs.end())
{
std::list<MovLoc> t;
t.push_back(c);
movs.insert(std::make_pair(defDcl, std::make_pair(0, t)));
}
else
{
it->second.second.push_back(c);
}
}
}
}
}
}
}
}
// Update number of uses of each dcl
for (auto bb : kernel.fg.BBs)
{
for (auto instIt = bb->instList.begin(), endIt = bb->instList.end();
instIt != endIt; instIt++)
{
auto inst = (*instIt);
if (inst->isPseudoKill())
{
auto dcl = inst->getDst()->getTopDcl();
auto it = movs.find(dcl);
if (it != movs.end())
{
it->second.second.push_back(std::make_pair(bb, instIt));
}
}
for (unsigned int i = 0; i < G4_MAX_SRCS; i++)
{
G4_Operand* opnd = inst->getSrc(i);
if (opnd &&
opnd->getTopDcl())
{
auto it = movs.find(opnd->getTopDcl());
if (it != movs.end())
it->second.first++;
}
}
}
}
for (auto mov : movs)
{
auto dcl = mov.first;
auto numRefs = mov.second.first;
auto& allMovs = mov.second.second;
if (numRefs == 0 && !dcl->getAddressed())
{
#if 0
printf("Removing movs/pseudoKill for dcl %s\n", dcl->getName());
#endif
for (auto m : allMovs)
{
auto bb = m.first;
auto iter = m.second;
#if 0
printf("\tFound %s occurence at $%d\n", (*iter)->opcode() == G4_mov ? "mov" : "pseudokill", (*iter)->getCISAOff());
#endif
bb->instList.erase(iter);
}
}
}
}
void CoalesceSpillFills::spillFillCleanup()
{
// Eliminate redundant fills when a write
// is close by:
//
// spill TV1 at offset = 1
// ..
// ..
// ..
// fill FP1 from offset = 1
// = FP1
// ===>
// Remove fill and replace occurence of FP1 with TV1
//
for (auto bb : kernel.fg.BBs)
{
auto startIt = bb->instList.begin();
auto endIt = bb->instList.end();
for (auto instIt = startIt;
instIt != endIt;
instIt++)
{
auto inst = (*instIt);
std::map<unsigned int, G4_INST*> writesPerOffset;
std::set<G4_Declare*> defs;
if (inst->isSend() &&
inst->getMsgDesc()->isScratchRead())
{
// Store offset, spill inst pair
unsigned int rowStart, numRows;
getScratchMsgInfo(inst, rowStart, numRows);
unsigned int lastRow = rowStart + numRows - 1;
// Scan window of instruction above current inst
// to check whether all rows read by current inst
// have been written.
auto pInstIt = instIt;
pInstIt--;
unsigned int w = cSpillFillCleanupWindowSize;
while (pInstIt != startIt &&
w > 0)
{
auto pInst = (*pInstIt);
if (pInst->isSplitSend() &&
pInst->getMsgDesc()->isScratchWrite())
{
unsigned int pRowStart, pNumRows;
getScratchMsgInfo(pInst, pRowStart, pNumRows);
// If any def of src1 dcl is found then dont
// consider this write for optimization. Its
// value in memory could be different than
// one held in variable.
auto pSrc1Dcl = pInst->getSrc(1)->getTopDcl();
if (defs.find(pSrc1Dcl) != defs.end())
{
pInstIt--;
continue;
}
for (unsigned int pRow = pRowStart;
pRow != (pRowStart + pNumRows);
pRow++)
{
auto writeIt = writesPerOffset.find(pRow);
// Check whether a more recent write was found for this row
if (writeIt != writesPerOffset.end())
continue;
writesPerOffset.insert(std::make_pair(pRow, pInst));
}
}
if (pInst->getDst() &&
pInst->getDst()->getTopDcl())
{
// Store any defs seen to handle WAR
defs.insert(pInst->getDst()->getTopDcl());
}
w--;
pInstIt--;
}
// Check whether writes for all rows were found
bool found = true;
for (auto row = rowStart; row <= lastRow; row++)
{
if (writesPerOffset.find(row) == writesPerOffset.end())
{
found = false;
break;
}
}
if (!found)
{
continue;
}
// Writes for all rows found
unsigned int execSize;
execSize = kernel.getSimdSize() > 16 ? 16 : kernel.getSimdSize();
for (auto row = rowStart; row <= lastRow;)
{
if (execSize == 16 &&
row == lastRow)
{
// In case of odd rows in SIMD16
execSize = 8;
}
else if (execSize == 16)
{
// In SIMD16 kernel 2 consecutive rows should come from same spill
if (writesPerOffset.find(row)->second != writesPerOffset.find(row + 1)->second)
{
execSize = 8;
}
}
// Insert SIMD8 mov per row
G4_DstRegRegion* nDst = kernel.fg.builder->createDstRegRegion(Direct,
inst->getDst()->getBase(), row + inst->getDst()->asDstRegRegion()->getRegOff() - rowStart,
0, 1, Type_UD);
auto write = writesPerOffset.find(row)->second;
G4_SrcRegRegion* src1Write = write->getSrc(1)->asSrcRegRegion();
unsigned int writeRowStart = write->getMsgDesc()->getScratchRWOffset();
unsigned int diff = row - writeRowStart;
G4_SrcRegRegion* nSrc = kernel.fg.builder->createSrcRegRegion(Mod_src_undef, Direct,
src1Write->getBase(), diff + src1Write->getRegOff(), 0,
kernel.fg.builder->getRegionStride1(), Type_UD);
G4_INST* mov = kernel.fg.builder->createInternalInst(nullptr, G4_mov, nullptr, false, (unsigned char)execSize,
nDst, nSrc, nullptr, InstOpt_WriteEnable);
bb->instList.insert(instIt, mov);
mov->setCISAOff(inst->getCISAOff());
row += execSize / 8;
}
auto tempIt = instIt;
tempIt--;
bb->instList.erase(instIt);
instIt = tempIt;
}
}
}
}
void CoalesceSpillFills::removeRedundantWrites()
{
typedef std::list<std::pair<G4_BB*, INST_LIST_ITER>> SPILLS;
typedef std::list<std::pair<G4_BB*, INST_LIST_ITER>> FILLS;
std::map<unsigned int, std::pair<SPILLS, FILLS>> scratchOffsetAccess;
// Traverse bottom-up to detect and remove redundant writes.
// Redundant writes include:
// 1. Successive writes to same offset without a fill in between,
// 2. Writes in program without any fill from that slot throughout
for (auto bb : kernel.fg.BBs)
{
auto endIt = bb->instList.end();
endIt--;
// Store spill slots that are written in to alongwith emask used
std::map<unsigned int, unsigned int> scratchOffToMask;
for (auto instIt = endIt;
instIt != bb->instList.begin();
instIt--)
{
auto inst = (*instIt);
if (inst->isSend())
{
unsigned int offset = 0, size = 0;
if (inst->getMsgDesc()->isScratchRead())
{
getScratchMsgInfo(inst, offset, size);
for (unsigned int k = offset; k != (offset + size); k++)
{
auto it = scratchOffToMask.find(k);
if (it != scratchOffToMask.end())
{
scratchOffToMask.erase(it);
}
}
}
else if (inst->getMsgDesc()->isScratchWrite())
{
getScratchMsgInfo(inst, offset, size);
bool allRowsFound = true;
unsigned int emask = inst->getMaskOption();
for (unsigned int k = offset; k != (offset + size); k++)
{
auto it = scratchOffToMask.find(k);
if (it != scratchOffToMask.end())
{
if (emask != it->second &&
(it->second & InstOpt_WriteEnable) == 0)
{
allRowsFound = false;
break;
}
}
else
{
allRowsFound = false;
break;
}
}
if (allRowsFound)
{
#if 0
printf("Removing redundant successive write at $%d\n", inst->getCISAOff());
#endif
instIt = bb->instList.erase(instIt);
}
else
{
unsigned int emask = inst->getOption();
for (unsigned int k = offset; k != (offset + size); k++)
{
scratchOffToMask.insert(std::make_pair(k, emask));
}
}
}
}
}
}
for (auto bb : kernel.fg.BBs)
{
auto endIt = bb->instList.end();
for (auto instIt = bb->instList.begin();
instIt != endIt;
instIt++)
{
auto inst = (*instIt);
if (!inst->isSend())
{
continue;
}
if (inst->getMsgDesc()->isScratchRead() ||
inst->getMsgDesc()->isScratchWrite())
{
unsigned int offset, size;
getScratchMsgInfo(inst, offset, size);
bool isRead = inst->getMsgDesc()->isScratchRead();
for (unsigned int i = offset; i != (offset + size); i++)
{
auto it = scratchOffsetAccess.find(i);
if (it != scratchOffsetAccess.end())
{
if (isRead)
{
auto& fill = it->second.second;
fill.push_back(std::make_pair(bb, instIt));
}
else
{
auto& spill = it->second.first;
spill.push_back(std::make_pair(bb, instIt));
}
}
else
{
SPILLS s;
FILLS f;
if (isRead)
f.push_back(std::make_pair(bb, instIt));
else
s.push_back(std::make_pair(bb, instIt));
scratchOffsetAccess.insert(std::make_pair(i, std::make_pair(s, f)));
}
}
}
}
}
std::map<G4_INST*, std::pair<INST_LIST_ITER, G4_BB*>> spillToRemove;
for (auto scratchAccess : scratchOffsetAccess)
{
if (scratchAccess.second.second.size() == 0 &&
scratchAccess.second.first.size() > 0)
{
// 0 fills for scratch slot
// Check whether all spill slots have 0 fills
// in case spills are coalesced.
for (auto spill : scratchAccess.second.first)
{
bool spillRequired = false;
unsigned int offset, size;
getScratchMsgInfo(*spill.second, offset, size);
// Verify that all slots from offset->(offset+size) have 0 fills
for (unsigned int slot = offset; slot != (offset + size); slot++)
{
auto it = scratchOffsetAccess.find(slot);
if (it->second.second.size() != 0)
{
spillRequired = true;
break;
}
}
if (!spillRequired)
{
spillToRemove.insert(std::make_pair(*spill.second, std::make_pair(spill.second, spill.first)));
}
}
}
else if (scratchAccess.second.first.size() == 0 &&
scratchAccess.second.second.size() > 0)
{
// 0 spills for scratch slot, non-zero fills
// Check whether all fill slots have 0 spills
// in case fills are coalesced.
for (auto fill : scratchAccess.second.second)
{
bool fillRequired = false;
unsigned int offset, size;
getScratchMsgInfo(*fill.second, offset, size);
// Verify that all slots from offset->(offset+size) have 0 spills
for (unsigned int slot = offset; slot != (offset + size); slot++)
{
auto it = scratchOffsetAccess.find(slot);
if (it->second.first.size() != 0)
{
fillRequired = true;
break;
}
}
if (!fillRequired)
{
spillToRemove.insert(std::make_pair(*fill.second, std::make_pair(fill.second, fill.first)));
}
}
}
}
for (auto removeSp : spillToRemove)
{
G4_BB* bb = removeSp.second.second;
#if 0
printf("Removing redundant scratch access at CISA $%d\n", removeSp.first->getCISAOff());
#endif
bb->instList.erase(removeSp.second.first);
}
}
void CoalesceSpillFills::run()
{
removeRedundantSplitMovs();
fills();
replaceMap.clear();
spills();
replaceMap.clear();
spillFillCleanup();
removeRedundantWrites();
fixSendsSrcOverlap();
}
void CoalesceSpillFills::dumpKernel()
{
for (auto bb : kernel.fg.BBs)
{
for (auto inst : bb->instList)
{
inst->emit(std::cerr);
std::cerr << "\t$" << inst->getCISAOff() << ", #" << rpe.getRegisterPressure(inst) << "\n";
}
}
}
void CoalesceSpillFills::dumpKernel(unsigned int v1, unsigned int v2)
{
bool start = false, end = false, canEnd = false;
for (auto bb : kernel.fg.BBs)
{
if (end)
break;
for (auto inst : bb->instList)
{
if (canEnd &&
inst->getCISAOff() > (int)v2)
{
end = true;
break;
}
if (inst->getCISAOff() == v2)
{
// This ensures invalid offsets
// are dumped till v2 is hit.
canEnd = true;
}
if (inst->getCISAOff() == v1)
start = true;
if (start && !end)
{
inst->dump();
printf(" // $%d, #%d\n", inst->getCISAOff(), rpe.getRegisterPressure(inst));
}
}
}
}
void CoalesceSpillFills::computeAddressTakenDcls()
{
for (auto dcl : kernel.Declares)
{
auto addrSpillFill = dcl->getAddrTakenSpillFill();
if (addrSpillFill)
addrTakenSpillFillDcl.insert(addrSpillFill);
}
}
}