LLVM 20.0.0git
AMDGPUMachineFunction.cpp
Go to the documentation of this file.
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
10#include "AMDGPU.h"
12#include "AMDGPUSubtarget.h"
16#include "llvm/IR/Constants.h"
17#include "llvm/IR/Metadata.h"
19
20using namespace llvm;
21
22static const GlobalVariable *
24 const Module *M = F.getParent();
25 SmallString<64> KernelDynLDSName("llvm.amdgcn.");
26 KernelDynLDSName += F.getName();
27 KernelDynLDSName += ".dynlds";
28 return M->getNamedGlobal(KernelDynLDSName);
29}
30
31static bool hasLDSKernelArgument(const Function &F) {
32 for (const Argument &Arg : F.args()) {
33 Type *ArgTy = Arg.getType();
34 if (auto PtrTy = dyn_cast<PointerType>(ArgTy)) {
35 if (PtrTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS)
36 return true;
37 }
38 }
39 return false;
40}
41
43 const AMDGPUSubtarget &ST)
44 : IsEntryFunction(AMDGPU::isEntryFunctionCC(F.getCallingConv())),
45 IsModuleEntryFunction(
46 AMDGPU::isModuleEntryFunctionCC(F.getCallingConv())),
47 IsChainFunction(AMDGPU::isChainCC(F.getCallingConv())) {
48
49 // FIXME: Should initialize KernArgSize based on ExplicitKernelArgOffset,
50 // except reserved size is not correctly aligned.
51
52 Attribute MemBoundAttr = F.getFnAttribute("amdgpu-memory-bound");
53 MemoryBound = MemBoundAttr.getValueAsBool();
54
55 Attribute WaveLimitAttr = F.getFnAttribute("amdgpu-wave-limiter");
56 WaveLimiter = WaveLimitAttr.getValueAsBool();
57
58 // FIXME: How is this attribute supposed to interact with statically known
59 // global sizes?
60 StringRef S = F.getFnAttribute("amdgpu-gds-size").getValueAsString();
61 if (!S.empty())
63
64 // Assume the attribute allocates before any known GDS globals.
66
67 // Second value, if present, is the maximum value that can be assigned.
68 // Useful in PromoteAlloca or for LDS spills. Could be used for diagnostics
69 // during codegen.
70 std::pair<unsigned, unsigned> LDSSizeRange = AMDGPU::getIntegerPairAttribute(
71 F, "amdgpu-lds-size", {0, UINT32_MAX}, true);
72
73 // The two separate variables are only profitable when the LDS module lowering
74 // pass is disabled. If graphics does not use dynamic LDS, this is never
75 // profitable. Leaving cleanup for a later change.
76 LDSSize = LDSSizeRange.first;
78
79 CallingConv::ID CC = F.getCallingConv();
81 ExplicitKernArgSize = ST.getExplicitKernArgSize(F, MaxKernArgAlign);
82
83 // FIXME: Shouldn't be target specific
84 Attribute NSZAttr = F.getFnAttribute("no-signed-zeros-fp-math");
86 NSZAttr.isStringAttribute() && NSZAttr.getValueAsString() == "true";
87
89 if (DynLdsGlobal || hasLDSKernelArgument(F))
90 UsesDynamicLDS = true;
91}
92
94 const GlobalVariable &GV,
95 Align Trailing) {
96 auto Entry = LocalMemoryObjects.insert(std::pair(&GV, 0));
97 if (!Entry.second)
98 return Entry.first->second;
99
100 Align Alignment =
101 DL.getValueOrABITypeAlignment(GV.getAlign(), GV.getValueType());
102
103 unsigned Offset;
105
106 std::optional<uint32_t> MaybeAbs = getLDSAbsoluteAddress(GV);
107 if (MaybeAbs) {
108 // Absolute address LDS variables that exist prior to the LDS lowering
109 // pass raise a fatal error in that pass. These failure modes are only
110 // reachable if that lowering pass is disabled or broken. If/when adding
111 // support for absolute addresses on user specified variables, the
112 // alignment check moves to the lowering pass and the frame calculation
113 // needs to take the user variables into consideration.
114
115 uint32_t ObjectStart = *MaybeAbs;
116
117 if (ObjectStart != alignTo(ObjectStart, Alignment)) {
118 report_fatal_error("Absolute address LDS variable inconsistent with "
119 "variable alignment");
120 }
121
122 if (isModuleEntryFunction()) {
123 // If this is a module entry function, we can also sanity check against
124 // the static frame. Strictly it would be better to check against the
125 // attribute, i.e. that the variable is within the always-allocated
126 // section, and not within some other non-absolute-address object
127 // allocated here, but the extra error detection is minimal and we would
128 // have to pass the Function around or cache the attribute value.
129 uint32_t ObjectEnd =
130 ObjectStart + DL.getTypeAllocSize(GV.getValueType());
131 if (ObjectEnd > StaticLDSSize) {
133 "Absolute address LDS variable outside of static frame");
134 }
135 }
136
137 Entry.first->second = ObjectStart;
138 return ObjectStart;
139 }
140
141 /// TODO: We should sort these to minimize wasted space due to alignment
142 /// padding. Currently the padding is decided by the first encountered use
143 /// during lowering.
145
146 StaticLDSSize += DL.getTypeAllocSize(GV.getValueType());
147
148 // Align LDS size to trailing, e.g. for aligning dynamic shared memory
149 LDSSize = alignTo(StaticLDSSize, Trailing);
150 } else {
152 "expected region address space");
153
155 StaticGDSSize += DL.getTypeAllocSize(GV.getValueType());
156
157 // FIXME: Apply alignment of dynamic GDS
159 }
160
161 Entry.first->second = Offset;
162 return Offset;
163}
164
165std::optional<uint32_t>
167 // TODO: Would be more consistent with the abs symbols to use a range
168 MDNode *MD = F.getMetadata("llvm.amdgcn.lds.kernel.id");
169 if (MD && MD->getNumOperands() == 1) {
170 if (ConstantInt *KnownSize =
171 mdconst::extract<ConstantInt>(MD->getOperand(0))) {
172 uint64_t ZExt = KnownSize->getZExtValue();
173 if (ZExt <= UINT32_MAX) {
174 return ZExt;
175 }
176 }
177 }
178 return {};
179}
180
181std::optional<uint32_t>
184 return {};
185
186 std::optional<ConstantRange> AbsSymRange = GV.getAbsoluteSymbolRange();
187 if (!AbsSymRange)
188 return {};
189
190 if (const APInt *V = AbsSymRange->getSingleElement()) {
191 std::optional<uint64_t> ZExt = V->tryZExtValue();
192 if (ZExt && (*ZExt <= UINT32_MAX)) {
193 return *ZExt;
194 }
195 }
196
197 return {};
198}
199
201 const GlobalVariable &GV) {
202 const Module *M = F.getParent();
203 const DataLayout &DL = M->getDataLayout();
204 assert(DL.getTypeAllocSize(GV.getValueType()).isZero());
205
206 Align Alignment =
207 DL.getValueOrABITypeAlignment(GV.getAlign(), GV.getValueType());
208 if (Alignment <= DynLDSAlign)
209 return;
210
211 LDSSize = alignTo(StaticLDSSize, Alignment);
212 DynLDSAlign = Alignment;
213
214 // If there is a dynamic LDS variable associated with this function F, every
215 // further dynamic LDS instance (allocated by calling setDynLDSAlign) must
216 // map to the same address. This holds because no LDS is allocated after the
217 // lowering pass if there are dynamic LDS variables present.
219 if (Dyn) {
220 unsigned Offset = LDSSize; // return this?
221 std::optional<uint32_t> Expect = getLDSAbsoluteAddress(*Dyn);
222 if (!Expect || (Offset != *Expect)) {
223 report_fatal_error("Inconsistent metadata on dynamic LDS variable");
224 }
225 }
226}
227
229 UsesDynamicLDS = DynLDS;
230}
231
static bool hasLDSKernelArgument(const Function &F)
static const GlobalVariable * getKernelDynLDSGlobalFromFunction(const Function &F)
Analyzes if a function potentially memory bound and if a kernel kernel may benefit from limiting numb...
Base class for AMDGPU specific classes of TargetSubtarget.
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
This file contains the declarations for the subclasses of Constant, which represent the different fla...
#define F(x, y, z)
Definition: MD5.cpp:55
This file contains the declarations for metadata subclasses.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
AMDGPUMachineFunction(const Function &F, const AMDGPUSubtarget &ST)
static std::optional< uint32_t > getLDSKernelIdMetadata(const Function &F)
Align DynLDSAlign
Align for dynamic shared memory if any.
uint32_t LDSSize
Number of bytes in the LDS that are being used.
void setDynLDSAlign(const Function &F, const GlobalVariable &GV)
static std::optional< uint32_t > getLDSAbsoluteAddress(const GlobalValue &GV)
unsigned allocateLDSGlobal(const DataLayout &DL, const GlobalVariable &GV)
uint32_t StaticLDSSize
Number of bytes in the LDS allocated statically.
Class for arbitrary precision integers.
Definition: APInt.h:77
This class represents an incoming formal argument to a Function.
Definition: Argument.h:31
bool isStringAttribute() const
Return true if the attribute is a string (target-dependent) attribute.
Definition: Attributes.cpp:347
bool getValueAsBool() const
Return the attribute's value as a boolean.
Definition: Attributes.cpp:378
StringRef getValueAsString() const
Return the attribute's value as a string.
Definition: Attributes.cpp:392
This is the shared class of boolean and integer constants.
Definition: Constants.h:81
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:63
MaybeAlign getAlign() const
Returns the alignment of the given variable or function.
Definition: GlobalObject.h:80
unsigned getAddressSpace() const
Definition: GlobalValue.h:205
std::optional< ConstantRange > getAbsoluteSymbolRange() const
If this is an absolute symbol reference, returns the range of the symbol, otherwise returns std::null...
Definition: Globals.cpp:407
Type * getValueType() const
Definition: GlobalValue.h:296
Metadata node.
Definition: Metadata.h:1069
const MDOperand & getOperand(unsigned I) const
Definition: Metadata.h:1430
unsigned getNumOperands() const
Return number of MDNode operands.
Definition: Metadata.h:1436
A Module instance is used to store all the information related to an LLVM module.
Definition: Module.h:65
SmallString - A SmallString is just a SmallVector with methods and accessors that make it work better...
Definition: SmallString.h:26
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:50
bool consumeInteger(unsigned Radix, T &Result)
Parse the current string as an integer of the specified radix.
Definition: StringRef.h:484
constexpr bool empty() const
empty - Check if the string is empty.
Definition: StringRef.h:134
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
@ REGION_ADDRESS
Address space for region memory. (GDS)
@ LOCAL_ADDRESS
Address space for local memory.
std::pair< unsigned, unsigned > getIntegerPairAttribute(const Function &F, StringRef Name, std::pair< unsigned, unsigned > Default, bool OnlyFirstRequired)
@ AMDGPU_KERNEL
Used for AMDGPU code object kernels.
Definition: CallingConv.h:200
@ SPIR_KERNEL
Used for SPIR kernel functions.
Definition: CallingConv.h:144
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
@ Offset
Definition: DWP.cpp:480
void report_fatal_error(Error Err, bool gen_crash_diag=true)
Report a serious error, calling any installed error handler.
Definition: Error.cpp:167
uint64_t alignTo(uint64_t Size, Align A)
Returns a multiple of A needed to store Size bytes.
Definition: Alignment.h:155
This struct is a compact representation of a valid (non-zero power of two) alignment.
Definition: Alignment.h:39