xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUMachineFunction.cpp (revision 0d8fe2373503aeac48492f28073049a8bfa4feb5)
1 //===-- AMDGPUMachineFunctionInfo.cpp ---------------------------------------=//
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 "AMDGPUMachineFunction.h"
10 #include "AMDGPUPerfHintAnalysis.h"
11 #include "AMDGPUSubtarget.h"
12 #include "llvm/CodeGen/MachineModuleInfo.h"
13 #include "llvm/Target/TargetMachine.h"
14 
15 using namespace llvm;
16 
17 AMDGPUMachineFunction::AMDGPUMachineFunction(const MachineFunction &MF)
18     : MachineFunctionInfo(), Mode(MF.getFunction()),
19       IsEntryFunction(
20           AMDGPU::isEntryFunctionCC(MF.getFunction().getCallingConv())),
21       IsModuleEntryFunction(
22           AMDGPU::isModuleEntryFunctionCC(MF.getFunction().getCallingConv())),
23       NoSignedZerosFPMath(MF.getTarget().Options.NoSignedZerosFPMath) {
24   const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(MF);
25 
26   // FIXME: Should initialize KernArgSize based on ExplicitKernelArgOffset,
27   // except reserved size is not correctly aligned.
28   const Function &F = MF.getFunction();
29 
30   Attribute MemBoundAttr = F.getFnAttribute("amdgpu-memory-bound");
31   MemoryBound = MemBoundAttr.isStringAttribute() &&
32                 MemBoundAttr.getValueAsString() == "true";
33 
34   Attribute WaveLimitAttr = F.getFnAttribute("amdgpu-wave-limiter");
35   WaveLimiter = WaveLimitAttr.isStringAttribute() &&
36                 WaveLimitAttr.getValueAsString() == "true";
37 
38   CallingConv::ID CC = F.getCallingConv();
39   if (CC == CallingConv::AMDGPU_KERNEL || CC == CallingConv::SPIR_KERNEL)
40     ExplicitKernArgSize = ST.getExplicitKernArgSize(F, MaxKernArgAlign);
41 }
42 
43 unsigned AMDGPUMachineFunction::allocateLDSGlobal(const DataLayout &DL,
44                                                   const GlobalVariable &GV) {
45   auto Entry = LocalMemoryObjects.insert(std::make_pair(&GV, 0));
46   if (!Entry.second)
47     return Entry.first->second;
48 
49   Align Alignment =
50       DL.getValueOrABITypeAlignment(GV.getAlign(), GV.getValueType());
51 
52   /// TODO: We should sort these to minimize wasted space due to alignment
53   /// padding. Currently the padding is decided by the first encountered use
54   /// during lowering.
55   unsigned Offset = StaticLDSSize = alignTo(StaticLDSSize, Alignment);
56 
57   Entry.first->second = Offset;
58   StaticLDSSize += DL.getTypeAllocSize(GV.getValueType());
59 
60   // Update the LDS size considering the padding to align the dynamic shared
61   // memory.
62   LDSSize = alignTo(StaticLDSSize, DynLDSAlign);
63 
64   return Offset;
65 }
66 
67 void AMDGPUMachineFunction::setDynLDSAlign(const DataLayout &DL,
68                                            const GlobalVariable &GV) {
69   assert(DL.getTypeAllocSize(GV.getValueType()).isZero());
70 
71   Align Alignment =
72       DL.getValueOrABITypeAlignment(GV.getAlign(), GV.getValueType());
73   if (Alignment <= DynLDSAlign)
74     return;
75 
76   LDSSize = alignTo(StaticLDSSize, Alignment);
77   DynLDSAlign = Alignment;
78 }
79