xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp (revision a7dea1671b87c07d2d266f836bfa8b58efc7c134)
1 //===- SIMachineFunctionInfo.cpp - SI Machine Function Info ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include "SIMachineFunctionInfo.h"
10 #include "AMDGPUArgumentUsageInfo.h"
11 #include "AMDGPUSubtarget.h"
12 #include "SIRegisterInfo.h"
13 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
14 #include "Utils/AMDGPUBaseInfo.h"
15 #include "llvm/ADT/Optional.h"
16 #include "llvm/CodeGen/MachineBasicBlock.h"
17 #include "llvm/CodeGen/MachineFrameInfo.h"
18 #include "llvm/CodeGen/MachineFunction.h"
19 #include "llvm/CodeGen/MachineRegisterInfo.h"
20 #include "llvm/IR/CallingConv.h"
21 #include "llvm/IR/Function.h"
22 #include <cassert>
23 #include <vector>
24 
25 #define MAX_LANES 64
26 
27 using namespace llvm;
28 
29 SIMachineFunctionInfo::SIMachineFunctionInfo(const MachineFunction &MF)
30   : AMDGPUMachineFunction(MF),
31     Mode(MF.getFunction()),
32     PrivateSegmentBuffer(false),
33     DispatchPtr(false),
34     QueuePtr(false),
35     KernargSegmentPtr(false),
36     DispatchID(false),
37     FlatScratchInit(false),
38     WorkGroupIDX(false),
39     WorkGroupIDY(false),
40     WorkGroupIDZ(false),
41     WorkGroupInfo(false),
42     PrivateSegmentWaveByteOffset(false),
43     WorkItemIDX(false),
44     WorkItemIDY(false),
45     WorkItemIDZ(false),
46     ImplicitBufferPtr(false),
47     ImplicitArgPtr(false),
48     GITPtrHigh(0xffffffff),
49     HighBitsOf32BitAddress(0),
50     GDSSize(0) {
51   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
52   const Function &F = MF.getFunction();
53   FlatWorkGroupSizes = ST.getFlatWorkGroupSizes(F);
54   WavesPerEU = ST.getWavesPerEU(F);
55 
56   Occupancy = ST.computeOccupancy(MF, getLDSSize());
57   CallingConv::ID CC = F.getCallingConv();
58 
59   if (CC == CallingConv::AMDGPU_KERNEL || CC == CallingConv::SPIR_KERNEL) {
60     if (!F.arg_empty())
61       KernargSegmentPtr = true;
62     WorkGroupIDX = true;
63     WorkItemIDX = true;
64   } else if (CC == CallingConv::AMDGPU_PS) {
65     PSInputAddr = AMDGPU::getInitialPSInputAddr(F);
66   }
67 
68   if (!isEntryFunction()) {
69     // Non-entry functions have no special inputs for now, other registers
70     // required for scratch access.
71     ScratchRSrcReg = AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3;
72     ScratchWaveOffsetReg = AMDGPU::SGPR33;
73 
74     // TODO: Pick a high register, and shift down, similar to a kernel.
75     FrameOffsetReg = AMDGPU::SGPR34;
76     StackPtrOffsetReg = AMDGPU::SGPR32;
77 
78     ArgInfo.PrivateSegmentBuffer =
79       ArgDescriptor::createRegister(ScratchRSrcReg);
80     ArgInfo.PrivateSegmentWaveByteOffset =
81       ArgDescriptor::createRegister(ScratchWaveOffsetReg);
82 
83     if (F.hasFnAttribute("amdgpu-implicitarg-ptr"))
84       ImplicitArgPtr = true;
85   } else {
86     if (F.hasFnAttribute("amdgpu-implicitarg-ptr")) {
87       KernargSegmentPtr = true;
88       MaxKernArgAlign = std::max(ST.getAlignmentForImplicitArgPtr(),
89                                  MaxKernArgAlign);
90     }
91   }
92 
93   if (F.hasFnAttribute("amdgpu-work-group-id-x"))
94     WorkGroupIDX = true;
95 
96   if (F.hasFnAttribute("amdgpu-work-group-id-y"))
97     WorkGroupIDY = true;
98 
99   if (F.hasFnAttribute("amdgpu-work-group-id-z"))
100     WorkGroupIDZ = true;
101 
102   if (F.hasFnAttribute("amdgpu-work-item-id-x"))
103     WorkItemIDX = true;
104 
105   if (F.hasFnAttribute("amdgpu-work-item-id-y"))
106     WorkItemIDY = true;
107 
108   if (F.hasFnAttribute("amdgpu-work-item-id-z"))
109     WorkItemIDZ = true;
110 
111   const MachineFrameInfo &FrameInfo = MF.getFrameInfo();
112   bool HasStackObjects = FrameInfo.hasStackObjects();
113 
114   if (isEntryFunction()) {
115     // X, XY, and XYZ are the only supported combinations, so make sure Y is
116     // enabled if Z is.
117     if (WorkItemIDZ)
118       WorkItemIDY = true;
119 
120     PrivateSegmentWaveByteOffset = true;
121 
122     // HS and GS always have the scratch wave offset in SGPR5 on GFX9.
123     if (ST.getGeneration() >= AMDGPUSubtarget::GFX9 &&
124         (CC == CallingConv::AMDGPU_HS || CC == CallingConv::AMDGPU_GS))
125       ArgInfo.PrivateSegmentWaveByteOffset =
126           ArgDescriptor::createRegister(AMDGPU::SGPR5);
127   }
128 
129   bool isAmdHsaOrMesa = ST.isAmdHsaOrMesa(F);
130   if (isAmdHsaOrMesa) {
131     PrivateSegmentBuffer = true;
132 
133     if (F.hasFnAttribute("amdgpu-dispatch-ptr"))
134       DispatchPtr = true;
135 
136     if (F.hasFnAttribute("amdgpu-queue-ptr"))
137       QueuePtr = true;
138 
139     if (F.hasFnAttribute("amdgpu-dispatch-id"))
140       DispatchID = true;
141   } else if (ST.isMesaGfxShader(F)) {
142     ImplicitBufferPtr = true;
143   }
144 
145   if (F.hasFnAttribute("amdgpu-kernarg-segment-ptr"))
146     KernargSegmentPtr = true;
147 
148   if (ST.hasFlatAddressSpace() && isEntryFunction() && isAmdHsaOrMesa) {
149     auto hasNonSpillStackObjects = [&]() {
150       // Avoid expensive checking if there's no stack objects.
151       if (!HasStackObjects)
152         return false;
153       for (auto OI = FrameInfo.getObjectIndexBegin(),
154                 OE = FrameInfo.getObjectIndexEnd(); OI != OE; ++OI)
155         if (!FrameInfo.isSpillSlotObjectIndex(OI))
156           return true;
157       // All stack objects are spill slots.
158       return false;
159     };
160     // TODO: This could be refined a lot. The attribute is a poor way of
161     // detecting calls that may require it before argument lowering.
162     if (hasNonSpillStackObjects() || F.hasFnAttribute("amdgpu-flat-scratch"))
163       FlatScratchInit = true;
164   }
165 
166   Attribute A = F.getFnAttribute("amdgpu-git-ptr-high");
167   StringRef S = A.getValueAsString();
168   if (!S.empty())
169     S.consumeInteger(0, GITPtrHigh);
170 
171   A = F.getFnAttribute("amdgpu-32bit-address-high-bits");
172   S = A.getValueAsString();
173   if (!S.empty())
174     S.consumeInteger(0, HighBitsOf32BitAddress);
175 
176   S = F.getFnAttribute("amdgpu-gds-size").getValueAsString();
177   if (!S.empty())
178     S.consumeInteger(0, GDSSize);
179 }
180 
181 void SIMachineFunctionInfo::limitOccupancy(const MachineFunction &MF) {
182   limitOccupancy(getMaxWavesPerEU());
183   const GCNSubtarget& ST = MF.getSubtarget<GCNSubtarget>();
184   limitOccupancy(ST.getOccupancyWithLocalMemSize(getLDSSize(),
185                  MF.getFunction()));
186 }
187 
188 unsigned SIMachineFunctionInfo::addPrivateSegmentBuffer(
189   const SIRegisterInfo &TRI) {
190   ArgInfo.PrivateSegmentBuffer =
191     ArgDescriptor::createRegister(TRI.getMatchingSuperReg(
192     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SGPR_128RegClass));
193   NumUserSGPRs += 4;
194   return ArgInfo.PrivateSegmentBuffer.getRegister();
195 }
196 
197 unsigned SIMachineFunctionInfo::addDispatchPtr(const SIRegisterInfo &TRI) {
198   ArgInfo.DispatchPtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg(
199     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass));
200   NumUserSGPRs += 2;
201   return ArgInfo.DispatchPtr.getRegister();
202 }
203 
204 unsigned SIMachineFunctionInfo::addQueuePtr(const SIRegisterInfo &TRI) {
205   ArgInfo.QueuePtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg(
206     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass));
207   NumUserSGPRs += 2;
208   return ArgInfo.QueuePtr.getRegister();
209 }
210 
211 unsigned SIMachineFunctionInfo::addKernargSegmentPtr(const SIRegisterInfo &TRI) {
212   ArgInfo.KernargSegmentPtr
213     = ArgDescriptor::createRegister(TRI.getMatchingSuperReg(
214     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass));
215   NumUserSGPRs += 2;
216   return ArgInfo.KernargSegmentPtr.getRegister();
217 }
218 
219 unsigned SIMachineFunctionInfo::addDispatchID(const SIRegisterInfo &TRI) {
220   ArgInfo.DispatchID = ArgDescriptor::createRegister(TRI.getMatchingSuperReg(
221     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass));
222   NumUserSGPRs += 2;
223   return ArgInfo.DispatchID.getRegister();
224 }
225 
226 unsigned SIMachineFunctionInfo::addFlatScratchInit(const SIRegisterInfo &TRI) {
227   ArgInfo.FlatScratchInit = ArgDescriptor::createRegister(TRI.getMatchingSuperReg(
228     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass));
229   NumUserSGPRs += 2;
230   return ArgInfo.FlatScratchInit.getRegister();
231 }
232 
233 unsigned SIMachineFunctionInfo::addImplicitBufferPtr(const SIRegisterInfo &TRI) {
234   ArgInfo.ImplicitBufferPtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg(
235     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass));
236   NumUserSGPRs += 2;
237   return ArgInfo.ImplicitBufferPtr.getRegister();
238 }
239 
240 static bool isCalleeSavedReg(const MCPhysReg *CSRegs, MCPhysReg Reg) {
241   for (unsigned I = 0; CSRegs[I]; ++I) {
242     if (CSRegs[I] == Reg)
243       return true;
244   }
245 
246   return false;
247 }
248 
249 /// \p returns true if \p NumLanes slots are available in VGPRs already used for
250 /// SGPR spilling.
251 //
252 // FIXME: This only works after processFunctionBeforeFrameFinalized
253 bool SIMachineFunctionInfo::haveFreeLanesForSGPRSpill(const MachineFunction &MF,
254                                                       unsigned NumNeed) const {
255   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
256   unsigned WaveSize = ST.getWavefrontSize();
257   return NumVGPRSpillLanes + NumNeed <= WaveSize * SpillVGPRs.size();
258 }
259 
260 /// Reserve a slice of a VGPR to support spilling for FrameIndex \p FI.
261 bool SIMachineFunctionInfo::allocateSGPRSpillToVGPR(MachineFunction &MF,
262                                                     int FI) {
263   std::vector<SpilledReg> &SpillLanes = SGPRToVGPRSpills[FI];
264 
265   // This has already been allocated.
266   if (!SpillLanes.empty())
267     return true;
268 
269   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
270   const SIRegisterInfo *TRI = ST.getRegisterInfo();
271   MachineFrameInfo &FrameInfo = MF.getFrameInfo();
272   MachineRegisterInfo &MRI = MF.getRegInfo();
273   unsigned WaveSize = ST.getWavefrontSize();
274 
275   unsigned Size = FrameInfo.getObjectSize(FI);
276   assert(Size >= 4 && Size <= 64 && "invalid sgpr spill size");
277   assert(TRI->spillSGPRToVGPR() && "not spilling SGPRs to VGPRs");
278 
279   int NumLanes = Size / 4;
280 
281   const MCPhysReg *CSRegs = MRI.getCalleeSavedRegs();
282 
283   // Make sure to handle the case where a wide SGPR spill may span between two
284   // VGPRs.
285   for (int I = 0; I < NumLanes; ++I, ++NumVGPRSpillLanes) {
286     unsigned LaneVGPR;
287     unsigned VGPRIndex = (NumVGPRSpillLanes % WaveSize);
288 
289     if (VGPRIndex == 0) {
290       LaneVGPR = TRI->findUnusedRegister(MRI, &AMDGPU::VGPR_32RegClass, MF);
291       if (LaneVGPR == AMDGPU::NoRegister) {
292         // We have no VGPRs left for spilling SGPRs. Reset because we will not
293         // partially spill the SGPR to VGPRs.
294         SGPRToVGPRSpills.erase(FI);
295         NumVGPRSpillLanes -= I;
296         return false;
297       }
298 
299       Optional<int> CSRSpillFI;
300       if ((FrameInfo.hasCalls() || !isEntryFunction()) && CSRegs &&
301           isCalleeSavedReg(CSRegs, LaneVGPR)) {
302         CSRSpillFI = FrameInfo.CreateSpillStackObject(4, 4);
303       }
304 
305       SpillVGPRs.push_back(SGPRSpillVGPRCSR(LaneVGPR, CSRSpillFI));
306 
307       // Add this register as live-in to all blocks to avoid machine verifer
308       // complaining about use of an undefined physical register.
309       for (MachineBasicBlock &BB : MF)
310         BB.addLiveIn(LaneVGPR);
311     } else {
312       LaneVGPR = SpillVGPRs.back().VGPR;
313     }
314 
315     SpillLanes.push_back(SpilledReg(LaneVGPR, VGPRIndex));
316   }
317 
318   return true;
319 }
320 
321 /// Reserve AGPRs or VGPRs to support spilling for FrameIndex \p FI.
322 /// Either AGPR is spilled to VGPR to vice versa.
323 /// Returns true if a \p FI can be eliminated completely.
324 bool SIMachineFunctionInfo::allocateVGPRSpillToAGPR(MachineFunction &MF,
325                                                     int FI,
326                                                     bool isAGPRtoVGPR) {
327   MachineRegisterInfo &MRI = MF.getRegInfo();
328   MachineFrameInfo &FrameInfo = MF.getFrameInfo();
329   const GCNSubtarget &ST =  MF.getSubtarget<GCNSubtarget>();
330 
331   assert(ST.hasMAIInsts() && FrameInfo.isSpillSlotObjectIndex(FI));
332 
333   auto &Spill = VGPRToAGPRSpills[FI];
334 
335   // This has already been allocated.
336   if (!Spill.Lanes.empty())
337     return Spill.FullyAllocated;
338 
339   unsigned Size = FrameInfo.getObjectSize(FI);
340   unsigned NumLanes = Size / 4;
341   Spill.Lanes.resize(NumLanes, AMDGPU::NoRegister);
342 
343   const TargetRegisterClass &RC =
344       isAGPRtoVGPR ? AMDGPU::VGPR_32RegClass : AMDGPU::AGPR_32RegClass;
345   auto Regs = RC.getRegisters();
346 
347   auto &SpillRegs = isAGPRtoVGPR ? SpillAGPR : SpillVGPR;
348   const SIRegisterInfo *TRI = ST.getRegisterInfo();
349   Spill.FullyAllocated = true;
350 
351   // FIXME: Move allocation logic out of MachineFunctionInfo and initialize
352   // once.
353   BitVector OtherUsedRegs;
354   OtherUsedRegs.resize(TRI->getNumRegs());
355 
356   const uint32_t *CSRMask =
357       TRI->getCallPreservedMask(MF, MF.getFunction().getCallingConv());
358   if (CSRMask)
359     OtherUsedRegs.setBitsInMask(CSRMask);
360 
361   // TODO: Should include register tuples, but doesn't matter with current
362   // usage.
363   for (MCPhysReg Reg : SpillAGPR)
364     OtherUsedRegs.set(Reg);
365   for (MCPhysReg Reg : SpillVGPR)
366     OtherUsedRegs.set(Reg);
367 
368   SmallVectorImpl<MCPhysReg>::const_iterator NextSpillReg = Regs.begin();
369   for (unsigned I = 0; I < NumLanes; ++I) {
370     NextSpillReg = std::find_if(
371         NextSpillReg, Regs.end(), [&MRI, &OtherUsedRegs](MCPhysReg Reg) {
372           return MRI.isAllocatable(Reg) && !MRI.isPhysRegUsed(Reg) &&
373                  !OtherUsedRegs[Reg];
374         });
375 
376     if (NextSpillReg == Regs.end()) { // Registers exhausted
377       Spill.FullyAllocated = false;
378       break;
379     }
380 
381     OtherUsedRegs.set(*NextSpillReg);
382     SpillRegs.push_back(*NextSpillReg);
383     Spill.Lanes[I] = *NextSpillReg++;
384   }
385 
386   return Spill.FullyAllocated;
387 }
388 
389 void SIMachineFunctionInfo::removeDeadFrameIndices(MachineFrameInfo &MFI) {
390   // The FP spill hasn't been inserted yet, so keep it around.
391   for (auto &R : SGPRToVGPRSpills) {
392     if (R.first != FramePointerSaveIndex)
393       MFI.RemoveStackObject(R.first);
394   }
395 
396   // All other SPGRs must be allocated on the default stack, so reset the stack
397   // ID.
398   for (int i = MFI.getObjectIndexBegin(), e = MFI.getObjectIndexEnd(); i != e;
399        ++i)
400     if (i != FramePointerSaveIndex)
401       MFI.setStackID(i, TargetStackID::Default);
402 
403   for (auto &R : VGPRToAGPRSpills) {
404     if (R.second.FullyAllocated)
405       MFI.RemoveStackObject(R.first);
406   }
407 }
408 
409 MCPhysReg SIMachineFunctionInfo::getNextUserSGPR() const {
410   assert(NumSystemSGPRs == 0 && "System SGPRs must be added after user SGPRs");
411   return AMDGPU::SGPR0 + NumUserSGPRs;
412 }
413 
414 MCPhysReg SIMachineFunctionInfo::getNextSystemSGPR() const {
415   return AMDGPU::SGPR0 + NumUserSGPRs + NumSystemSGPRs;
416 }
417 
418 static yaml::StringValue regToString(unsigned Reg,
419                                      const TargetRegisterInfo &TRI) {
420   yaml::StringValue Dest;
421   {
422     raw_string_ostream OS(Dest.Value);
423     OS << printReg(Reg, &TRI);
424   }
425   return Dest;
426 }
427 
428 static Optional<yaml::SIArgumentInfo>
429 convertArgumentInfo(const AMDGPUFunctionArgInfo &ArgInfo,
430                     const TargetRegisterInfo &TRI) {
431   yaml::SIArgumentInfo AI;
432 
433   auto convertArg = [&](Optional<yaml::SIArgument> &A,
434                         const ArgDescriptor &Arg) {
435     if (!Arg)
436       return false;
437 
438     // Create a register or stack argument.
439     yaml::SIArgument SA = yaml::SIArgument::createArgument(Arg.isRegister());
440     if (Arg.isRegister()) {
441       raw_string_ostream OS(SA.RegisterName.Value);
442       OS << printReg(Arg.getRegister(), &TRI);
443     } else
444       SA.StackOffset = Arg.getStackOffset();
445     // Check and update the optional mask.
446     if (Arg.isMasked())
447       SA.Mask = Arg.getMask();
448 
449     A = SA;
450     return true;
451   };
452 
453   bool Any = false;
454   Any |= convertArg(AI.PrivateSegmentBuffer, ArgInfo.PrivateSegmentBuffer);
455   Any |= convertArg(AI.DispatchPtr, ArgInfo.DispatchPtr);
456   Any |= convertArg(AI.QueuePtr, ArgInfo.QueuePtr);
457   Any |= convertArg(AI.KernargSegmentPtr, ArgInfo.KernargSegmentPtr);
458   Any |= convertArg(AI.DispatchID, ArgInfo.DispatchID);
459   Any |= convertArg(AI.FlatScratchInit, ArgInfo.FlatScratchInit);
460   Any |= convertArg(AI.PrivateSegmentSize, ArgInfo.PrivateSegmentSize);
461   Any |= convertArg(AI.WorkGroupIDX, ArgInfo.WorkGroupIDX);
462   Any |= convertArg(AI.WorkGroupIDY, ArgInfo.WorkGroupIDY);
463   Any |= convertArg(AI.WorkGroupIDZ, ArgInfo.WorkGroupIDZ);
464   Any |= convertArg(AI.WorkGroupInfo, ArgInfo.WorkGroupInfo);
465   Any |= convertArg(AI.PrivateSegmentWaveByteOffset,
466                     ArgInfo.PrivateSegmentWaveByteOffset);
467   Any |= convertArg(AI.ImplicitArgPtr, ArgInfo.ImplicitArgPtr);
468   Any |= convertArg(AI.ImplicitBufferPtr, ArgInfo.ImplicitBufferPtr);
469   Any |= convertArg(AI.WorkItemIDX, ArgInfo.WorkItemIDX);
470   Any |= convertArg(AI.WorkItemIDY, ArgInfo.WorkItemIDY);
471   Any |= convertArg(AI.WorkItemIDZ, ArgInfo.WorkItemIDZ);
472 
473   if (Any)
474     return AI;
475 
476   return None;
477 }
478 
479 yaml::SIMachineFunctionInfo::SIMachineFunctionInfo(
480   const llvm::SIMachineFunctionInfo& MFI,
481   const TargetRegisterInfo &TRI)
482   : ExplicitKernArgSize(MFI.getExplicitKernArgSize()),
483     MaxKernArgAlign(MFI.getMaxKernArgAlign()),
484     LDSSize(MFI.getLDSSize()),
485     IsEntryFunction(MFI.isEntryFunction()),
486     NoSignedZerosFPMath(MFI.hasNoSignedZerosFPMath()),
487     MemoryBound(MFI.isMemoryBound()),
488     WaveLimiter(MFI.needsWaveLimiter()),
489     HighBitsOf32BitAddress(MFI.get32BitAddressHighBits()),
490     ScratchRSrcReg(regToString(MFI.getScratchRSrcReg(), TRI)),
491     ScratchWaveOffsetReg(regToString(MFI.getScratchWaveOffsetReg(), TRI)),
492     FrameOffsetReg(regToString(MFI.getFrameOffsetReg(), TRI)),
493     StackPtrOffsetReg(regToString(MFI.getStackPtrOffsetReg(), TRI)),
494     ArgInfo(convertArgumentInfo(MFI.getArgInfo(), TRI)),
495     Mode(MFI.getMode()) {}
496 
497 void yaml::SIMachineFunctionInfo::mappingImpl(yaml::IO &YamlIO) {
498   MappingTraits<SIMachineFunctionInfo>::mapping(YamlIO, *this);
499 }
500 
501 bool SIMachineFunctionInfo::initializeBaseYamlFields(
502   const yaml::SIMachineFunctionInfo &YamlMFI) {
503   ExplicitKernArgSize = YamlMFI.ExplicitKernArgSize;
504   MaxKernArgAlign = assumeAligned(YamlMFI.MaxKernArgAlign);
505   LDSSize = YamlMFI.LDSSize;
506   HighBitsOf32BitAddress = YamlMFI.HighBitsOf32BitAddress;
507   IsEntryFunction = YamlMFI.IsEntryFunction;
508   NoSignedZerosFPMath = YamlMFI.NoSignedZerosFPMath;
509   MemoryBound = YamlMFI.MemoryBound;
510   WaveLimiter = YamlMFI.WaveLimiter;
511   return false;
512 }
513