[llvm] [AMDGPU] Enable amdgpu-sw-lower-lds pass to lower LDS accesses to use device global memory (PR #87265)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Apr 1 10:07:56 PDT 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Chaitanya (skc7)
<details>
<summary>Changes</summary>
This PR introduces new pass "amdgpu-sw-lower-lds". It lowers the local data store, LDS, uses in kernel and non-kernel functions in module with dynamically allocated device global memory.
**Replacement of Kernel LDS accesses:**
- For a kernel, LDS access can be static or dynamic which are direct (accessed within kernel) and indirect (accessed through non-kernels). A device global memory equal to size of all these LDS globals will be allocated. At the prologue of the kernel, a single work-item from the work-group, does a "malloc" and stores the pointer of the allocation in new LDS global that will be created for the kernel. This will be called "malloc LDS global" in this pass. Each LDS access corresponds to an offset in the allocated memory. All static LDS accesses will be allocated first and then dynamic LDS will occupy the device global memory. To store the offsets corresponding to all LDS accesses, another global variable is created which will be called "metadata global" in this pass.
- **Malloc LDS Global:**
It is LDS global of ptr type with name "llvm.amdgcn.sw.lds.{kernel-name}".
- **Metadata Global:**
It is of struct type, with n members. n equals the number of LDS globals accessed by the kernel(direct and indirect). Each member of struct is another struct of type {i32, i32}. First member corresponds to offset, second member corresponds to size of LDS global being replaced.
It will have name "llvm.amdgcn.sw.lds.{kernel-name}.md".
This global will have an intializer with static LDS related offsets and sizes initialized. But for dynamic LDS related entries, offsets will be intialized to previous static LDS allocation end offset. Sizes for them will be zero initially. These dynamic LDS offset and size values will be updated with in the kernel, since kernel can read the dynamic LDS size allocation done at runtime with query to "hidden_dynamic_lds_size" hidden kernel argument.
- LDS accesses within the kernel will be replaced by "gep" ptr to corresponding offset into allocated device global memory for the kernel.
- At the epilogue of kernel, allocated memory would be made free by the same single work-item.
**Replacement of non-kernel LDS accesses:**
- Multiple kernels can access the same non-kernel function. All the kernels accessing LDS through non-kernels are sorted and assigned a kernel-id. All the LDS globals accessed by non-kernels are sorted. This information is used to build two globals which are used as tables for query of LDS replacement:
- **Base table:**
Base table will have single row, with elements of the row placed as per kernel ID.
Each element in the row corresponds to addresss of "malloc LDS global" variable created for that kernel.
- **Offset table:**
Offset table will have multiple rows and columns.
Rows are assumed to be from 0 to (n-1). n is total number of kernels accessing the LDS through non-kernels. Each row will have m elements. m is the total number of unique LDS globals accessed by all non-kernels. Each element in the row correspond to the address of the replacement of LDS global done by that particular kernel. A LDS variable in non-kernel will be replaced based on the information from base and offset tables. Based on kernel-id query, address of "malloc LDS global" for that corresponding kernel is obtained from base table. The Offset into the base "malloc LDS global" is obtained from corresponding element in offset table. With this information, replacement value is obtained.
- Other changes:
"amdgpu-sw-lower-lds" pass is needed for address sanitizer instrumentation of LDS. So, this pass is enabled only if asan is enabled.
There are certain utility functions which can be reused from lower-module-lds pass. These functions are moved to AMDGPUMemoryUtils module for re-use in this new pass. lower-module-lds pass will be disabled if asan feature is enabled.
---
Patch is 122.46 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/87265.diff
18 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/AMDGPU.h (+9)
- (modified) llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp (+1-185)
- (modified) llvm/lib/Target/AMDGPU/AMDGPUPassRegistry.def (+1)
- (added) llvm/lib/Target/AMDGPU/AMDGPUSwLowerLDS.cpp (+865)
- (modified) llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp (+6)
- (modified) llvm/lib/Target/AMDGPU/CMakeLists.txt (+1)
- (modified) llvm/lib/Target/AMDGPU/Utils/AMDGPUMemoryUtils.cpp (+176)
- (modified) llvm/lib/Target/AMDGPU/Utils/AMDGPUMemoryUtils.h (+24)
- (added) llvm/test/CodeGen/AMDGPU/amdgpu-sw-lower-lds-dynamic-indirect-access.ll (+99)
- (added) llvm/test/CodeGen/AMDGPU/amdgpu-sw-lower-lds-dynamic-lds-test.ll (+57)
- (added) llvm/test/CodeGen/AMDGPU/amdgpu-sw-lower-lds-multi-static-dynamic-indirect-access.ll (+192)
- (added) llvm/test/CodeGen/AMDGPU/amdgpu-sw-lower-lds-multiple-blocks-return.ll (+79)
- (added) llvm/test/CodeGen/AMDGPU/amdgpu-sw-lower-lds-static-dynamic-indirect-access.ll (+101)
- (added) llvm/test/CodeGen/AMDGPU/amdgpu-sw-lower-lds-static-dynamic-lds-test.ll (+88)
- (added) llvm/test/CodeGen/AMDGPU/amdgpu-sw-lower-lds-static-indirect-access-function-param.ll (+61)
- (added) llvm/test/CodeGen/AMDGPU/amdgpu-sw-lower-lds-static-indirect-access-nested.ll (+212)
- (added) llvm/test/CodeGen/AMDGPU/amdgpu-sw-lower-lds-static-indirect-access.ll (+84)
- (added) llvm/test/CodeGen/AMDGPU/amdgpu-sw-lower-lds-static-lds-test.ll (+58)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.h b/llvm/lib/Target/AMDGPU/AMDGPU.h
index 6016bd5187d887..15ff74f7c53af3 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.h
@@ -263,6 +263,15 @@ struct AMDGPUAlwaysInlinePass : PassInfoMixin<AMDGPUAlwaysInlinePass> {
bool GlobalOpt;
};
+void initializeAMDGPUSwLowerLDSLegacyPass(PassRegistry &);
+extern char &AMDGPUSwLowerLDSLegacyPassID;
+ModulePass *createAMDGPUSwLowerLDSLegacyPass();
+
+struct AMDGPUSwLowerLDSPass : PassInfoMixin<AMDGPUSwLowerLDSPass> {
+ AMDGPUSwLowerLDSPass() {}
+ PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM);
+};
+
class AMDGPUCodeGenPreparePass
: public PassInfoMixin<AMDGPUCodeGenPreparePass> {
private:
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp b/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp
index 595f09664c55e4..f0456d3f62a816 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp
@@ -212,6 +212,7 @@
#define DEBUG_TYPE "amdgpu-lower-module-lds"
using namespace llvm;
+using namespace AMDGPU;
namespace {
@@ -234,17 +235,6 @@ cl::opt<LoweringKind> LoweringKindLoc(
clEnumValN(LoweringKind::hybrid, "hybrid",
"Lower via mixture of above strategies")));
-bool isKernelLDS(const Function *F) {
- // Some weirdness here. AMDGPU::isKernelCC does not call into
- // AMDGPU::isKernel with the calling conv, it instead calls into
- // isModuleEntryFunction which returns true for more calling conventions
- // than AMDGPU::isKernel does. There's a FIXME on AMDGPU::isKernel.
- // There's also a test that checks that the LDS lowering does not hit on
- // a graphics shader, denoted amdgpu_ps, so stay with the limited case.
- // Putting LDS in the name of the function to draw attention to this.
- return AMDGPU::isKernel(F->getCallingConv());
-}
-
template <typename T> std::vector<T> sortByName(std::vector<T> &&V) {
llvm::sort(V.begin(), V.end(), [](const auto *L, const auto *R) {
return L->getName() < R->getName();
@@ -305,183 +295,9 @@ class AMDGPULowerModuleLDS {
Decl, {}, {OperandBundleDefT<Value *>("ExplicitUse", UseInstance)});
}
- static bool eliminateConstantExprUsesOfLDSFromAllInstructions(Module &M) {
- // Constants are uniqued within LLVM. A ConstantExpr referring to a LDS
- // global may have uses from multiple different functions as a result.
- // This pass specialises LDS variables with respect to the kernel that
- // allocates them.
-
- // This is semantically equivalent to (the unimplemented as slow):
- // for (auto &F : M.functions())
- // for (auto &BB : F)
- // for (auto &I : BB)
- // for (Use &Op : I.operands())
- // if (constantExprUsesLDS(Op))
- // replaceConstantExprInFunction(I, Op);
-
- SmallVector<Constant *> LDSGlobals;
- for (auto &GV : M.globals())
- if (AMDGPU::isLDSVariableToLower(GV))
- LDSGlobals.push_back(&GV);
-
- return convertUsersOfConstantsToInstructions(LDSGlobals);
- }
-
public:
AMDGPULowerModuleLDS(const AMDGPUTargetMachine &TM_) : TM(TM_) {}
- using FunctionVariableMap = DenseMap<Function *, DenseSet<GlobalVariable *>>;
-
- using VariableFunctionMap = DenseMap<GlobalVariable *, DenseSet<Function *>>;
-
- static void getUsesOfLDSByFunction(CallGraph const &CG, Module &M,
- FunctionVariableMap &kernels,
- FunctionVariableMap &functions) {
-
- // Get uses from the current function, excluding uses by called functions
- // Two output variables to avoid walking the globals list twice
- for (auto &GV : M.globals()) {
- if (!AMDGPU::isLDSVariableToLower(GV)) {
- continue;
- }
-
- for (User *V : GV.users()) {
- if (auto *I = dyn_cast<Instruction>(V)) {
- Function *F = I->getFunction();
- if (isKernelLDS(F)) {
- kernels[F].insert(&GV);
- } else {
- functions[F].insert(&GV);
- }
- }
- }
- }
- }
-
- struct LDSUsesInfoTy {
- FunctionVariableMap direct_access;
- FunctionVariableMap indirect_access;
- };
-
- static LDSUsesInfoTy getTransitiveUsesOfLDS(CallGraph const &CG, Module &M) {
-
- FunctionVariableMap direct_map_kernel;
- FunctionVariableMap direct_map_function;
- getUsesOfLDSByFunction(CG, M, direct_map_kernel, direct_map_function);
-
- // Collect variables that are used by functions whose address has escaped
- DenseSet<GlobalVariable *> VariablesReachableThroughFunctionPointer;
- for (Function &F : M.functions()) {
- if (!isKernelLDS(&F))
- if (F.hasAddressTaken(nullptr,
- /* IgnoreCallbackUses */ false,
- /* IgnoreAssumeLikeCalls */ false,
- /* IgnoreLLVMUsed */ true,
- /* IgnoreArcAttachedCall */ false)) {
- set_union(VariablesReachableThroughFunctionPointer,
- direct_map_function[&F]);
- }
- }
-
- auto functionMakesUnknownCall = [&](const Function *F) -> bool {
- assert(!F->isDeclaration());
- for (const CallGraphNode::CallRecord &R : *CG[F]) {
- if (!R.second->getFunction()) {
- return true;
- }
- }
- return false;
- };
-
- // Work out which variables are reachable through function calls
- FunctionVariableMap transitive_map_function = direct_map_function;
-
- // If the function makes any unknown call, assume the worst case that it can
- // access all variables accessed by functions whose address escaped
- for (Function &F : M.functions()) {
- if (!F.isDeclaration() && functionMakesUnknownCall(&F)) {
- if (!isKernelLDS(&F)) {
- set_union(transitive_map_function[&F],
- VariablesReachableThroughFunctionPointer);
- }
- }
- }
-
- // Direct implementation of collecting all variables reachable from each
- // function
- for (Function &Func : M.functions()) {
- if (Func.isDeclaration() || isKernelLDS(&Func))
- continue;
-
- DenseSet<Function *> seen; // catches cycles
- SmallVector<Function *, 4> wip{&Func};
-
- while (!wip.empty()) {
- Function *F = wip.pop_back_val();
-
- // Can accelerate this by referring to transitive map for functions that
- // have already been computed, with more care than this
- set_union(transitive_map_function[&Func], direct_map_function[F]);
-
- for (const CallGraphNode::CallRecord &R : *CG[F]) {
- Function *ith = R.second->getFunction();
- if (ith) {
- if (!seen.contains(ith)) {
- seen.insert(ith);
- wip.push_back(ith);
- }
- }
- }
- }
- }
-
- // direct_map_kernel lists which variables are used by the kernel
- // find the variables which are used through a function call
- FunctionVariableMap indirect_map_kernel;
-
- for (Function &Func : M.functions()) {
- if (Func.isDeclaration() || !isKernelLDS(&Func))
- continue;
-
- for (const CallGraphNode::CallRecord &R : *CG[&Func]) {
- Function *ith = R.second->getFunction();
- if (ith) {
- set_union(indirect_map_kernel[&Func], transitive_map_function[ith]);
- } else {
- set_union(indirect_map_kernel[&Func],
- VariablesReachableThroughFunctionPointer);
- }
- }
- }
-
- // Verify that we fall into one of 2 cases:
- // - All variables are absolute: this is a re-run of the pass
- // so we don't have anything to do.
- // - No variables are absolute.
- std::optional<bool> HasAbsoluteGVs;
- for (auto &Map : {direct_map_kernel, indirect_map_kernel}) {
- for (auto &[Fn, GVs] : Map) {
- for (auto *GV : GVs) {
- bool IsAbsolute = GV->isAbsoluteSymbolRef();
- if (HasAbsoluteGVs.has_value()) {
- if (*HasAbsoluteGVs != IsAbsolute) {
- report_fatal_error(
- "Module cannot mix absolute and non-absolute LDS GVs");
- }
- } else
- HasAbsoluteGVs = IsAbsolute;
- }
- }
- }
-
- // If we only had absolute GVs, we have nothing to do, return an empty
- // result.
- if (HasAbsoluteGVs && *HasAbsoluteGVs)
- return {FunctionVariableMap(), FunctionVariableMap()};
-
- return {std::move(direct_map_kernel), std::move(indirect_map_kernel)};
- }
-
struct LDSVariableReplacement {
GlobalVariable *SGV = nullptr;
DenseMap<GlobalVariable *, Constant *> LDSVarsToConstantGEP;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUPassRegistry.def b/llvm/lib/Target/AMDGPU/AMDGPUPassRegistry.def
index 90f36fadf35903..eda4949d0296d5 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUPassRegistry.def
+++ b/llvm/lib/Target/AMDGPU/AMDGPUPassRegistry.def
@@ -22,6 +22,7 @@ MODULE_PASS("amdgpu-lower-buffer-fat-pointers",
AMDGPULowerBufferFatPointersPass(*this))
MODULE_PASS("amdgpu-lower-ctor-dtor", AMDGPUCtorDtorLoweringPass())
MODULE_PASS("amdgpu-lower-module-lds", AMDGPULowerModuleLDSPass(*this))
+MODULE_PASS("amdgpu-sw-lower-lds", AMDGPUSwLowerLDSPass())
MODULE_PASS("amdgpu-printf-runtime-binding", AMDGPUPrintfRuntimeBindingPass())
MODULE_PASS("amdgpu-unify-metadata", AMDGPUUnifyMetadataPass())
#undef MODULE_PASS
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSwLowerLDS.cpp b/llvm/lib/Target/AMDGPU/AMDGPUSwLowerLDS.cpp
new file mode 100644
index 00000000000000..ed3670fa1386d6
--- /dev/null
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSwLowerLDS.cpp
@@ -0,0 +1,865 @@
+//===-- AMDGPUSwLowerLDS.cpp -----------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass lowers the local data store, LDS, uses in kernel and non-kernel
+// functions in module with dynamically allocated device global memory.
+//
+// Replacement of Kernel LDS accesses:
+// For a kernel, LDS access can be static or dynamic which are direct
+// (accessed within kernel) and indirect (accessed through non-kernels).
+// A device global memory equal to size of all these LDS globals will be
+// allocated. At the prologue of the kernel, a single work-item from the
+// work-group, does a "malloc" and stores the pointer of the allocation in
+// new LDS global that will be created for the kernel. This will be called
+// "malloc LDS global" in this pass.
+// Each LDS access corresponds to an offset in the allocated memory.
+// All static LDS accesses will be allocated first and then dynamic LDS
+// will occupy the device global memoery.
+// To store the offsets corresponding to all LDS accesses, another global
+// variable is created which will be called "metadata global" in this pass.
+// - Malloc LDS Global:
+// It is LDS global of ptr type with name
+// "llvm.amdgcn.sw.lds.<kernel-name>".
+// - Metadata Global:
+// It is of struct type, with n members. n equals the number of LDS
+// globals accessed by the kernel(direct and indirect). Each member of
+// struct is another struct of type {i32, i32}. First member corresponds
+// to offset, second member corresponds to size of LDS global being
+// replaced. It will have name "llvm.amdgcn.sw.lds.<kernel-name>.md".
+// This global will have an intializer with static LDS related offsets
+// and sizes initialized. But for dynamic LDS related entries, offsets
+// will be intialized to previous static LDS allocation end offset. Sizes
+// for them will be zero initially. These dynamic LDS offset and size
+// values will be updated with in the kernel, since kernel can read the
+// dynamic LDS size allocation done at runtime with query to
+// "hidden_dynamic_lds_size" hidden kernel argument.
+//
+// LDS accesses within the kernel will be replaced by "gep" ptr to
+// corresponding offset into allocated device global memory for the kernel.
+// At the epilogue of kernel, allocated memory would be made free by the same
+// single work-item.
+//
+// Replacement of non-kernel LDS accesses:
+// Multiple kernels can access the same non-kernel function.
+// All the kernels accessing LDS through non-kernels are sorted and
+// assigned a kernel-id. All the LDS globals accessed by non-kernels
+// are sorted. This information is used to build two tables:
+// - Base table:
+// Base table will have single row, with elements of the row
+// placed as per kernel ID. Each element in the row corresponds
+// to addresss of "malloc LDS global" variable created for
+// that kernel.
+// - Offset table:
+// Offset table will have multiple rows and columns.
+// Rows are assumed to be from 0 to (n-1). n is total number
+// of kernels accessing the LDS through non-kernels.
+// Each row will have m elements. m is the total number of
+// unique LDS globals accessed by all non-kernels.
+// Each element in the row correspond to the address of
+// the replacement of LDS global done by that particular kernel.
+// A LDS variable in non-kernel will be replaced based on the information
+// from base and offset tables. Based on kernel-id query, address of "malloc
+// LDS global" for that corresponding kernel is obtained from base table.
+// The Offset into the base "malloc LDS global" is obtained from
+// corresponding element in offset table. With this information, replacement
+// value is obtained.
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "Utils/AMDGPUMemoryUtils.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/DenseSet.h"
+#include "llvm/ADT/SetOperations.h"
+#include "llvm/ADT/SetVector.h"
+#include "llvm/ADT/StringRef.h"
+#include "llvm/Analysis/CallGraph.h"
+#include "llvm/Analysis/DomTreeUpdater.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicsAMDGPU.h"
+#include "llvm/IR/MDBuilder.h"
+#include "llvm/IR/ReplaceConstant.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Pass.h"
+#include "llvm/Transforms/Utils/ModuleUtils.h"
+
+#include <algorithm>
+
+#define DEBUG_TYPE "amdgpu-sw-lower-lds"
+
+using namespace llvm;
+using namespace AMDGPU;
+
+namespace {
+
+using DomTreeCallback = function_ref<DominatorTree *(Function &F)>;
+
+struct LDSAccessTypeInfo {
+ SetVector<GlobalVariable *> StaticLDSGlobals;
+ SetVector<GlobalVariable *> DynamicLDSGlobals;
+};
+
+// Struct to hold all the Metadata required for a kernel
+// to replace a LDS global uses with corresponding offset
+// in to device global memory.
+struct KernelLDSParameters {
+ GlobalVariable *MallocLDSGlobal{nullptr};
+ GlobalVariable *MallocMetadataGlobal{nullptr};
+ LDSAccessTypeInfo DirectAccess;
+ LDSAccessTypeInfo IndirectAccess;
+ DenseMap<GlobalVariable *, SmallVector<uint32_t, 3>>
+ LDSToReplacementIndicesMap;
+ int32_t KernelId{-1};
+ uint32_t MallocSize{0};
+};
+
+// Struct to store infor for creation of offset table
+// for all the non-kernel LDS accesses.
+struct NonKernelLDSParameters {
+ GlobalVariable *LDSBaseTable{nullptr};
+ GlobalVariable *LDSOffsetTable{nullptr};
+ SetVector<Function *> OrderedKernels;
+ SetVector<GlobalVariable *> OrdereLDSGlobals;
+};
+
+class AMDGPUSwLowerLDS {
+public:
+ AMDGPUSwLowerLDS(Module &mod, DomTreeCallback Callback)
+ : M(mod), IRB(M.getContext()), DTCallback(Callback) {}
+ bool Run();
+ void GetUsesOfLDSByNonKernels(CallGraph const &CG,
+ FunctionVariableMap &functions);
+ SetVector<Function *>
+ GetOrderedIndirectLDSAccessingKernels(SetVector<Function *> &&Kernels);
+ SetVector<GlobalVariable *>
+ GetOrderedNonKernelAllLDSGlobals(SetVector<GlobalVariable *> &&Variables);
+ void PopulateMallocLDSGlobal(Function *Func);
+ void PopulateMallocMetadataGlobal(Function *Func);
+ void PopulateLDSToReplacementIndicesMap(Function *Func);
+ void ReplaceKernelLDSAccesses(Function *Func);
+ void LowerKernelLDSAccesses(Function *Func, DomTreeUpdater &DTU);
+ void BuildNonKernelLDSOffsetTable(
+ std::shared_ptr<NonKernelLDSParameters> &NKLDSParams);
+ void BuildNonKernelLDSBaseTable(
+ std::shared_ptr<NonKernelLDSParameters> &NKLDSParams);
+ Constant *
+ GetAddressesOfVariablesInKernel(Function *Func,
+ SetVector<GlobalVariable *> &Variables);
+ void LowerNonKernelLDSAccesses(
+ Function *Func, SetVector<GlobalVariable *> &LDSGlobals,
+ std::shared_ptr<NonKernelLDSParameters> &NKLDSParams);
+
+private:
+ Module &M;
+ IRBuilder<> IRB;
+ DomTreeCallback DTCallback;
+ DenseMap<Function *, std::shared_ptr<KernelLDSParameters>>
+ KernelToLDSParametersMap;
+};
+
+template <typename T> SetVector<T> SortByName(std::vector<T> &&V) {
+ // Sort the vector of globals or Functions based on their name.
+ // Returns a SetVector of globals/Functions.
+ llvm::sort(V.begin(), V.end(), [](const auto *L, const auto *R) {
+ return L->getName() < R->getName();
+ });
+ return {std::move(SetVector<T>(V.begin(), V.end()))};
+}
+
+SetVector<GlobalVariable *> AMDGPUSwLowerLDS::GetOrderedNonKernelAllLDSGlobals(
+ SetVector<GlobalVariable *> &&Variables) {
+ // Sort all the non-kernel LDS accesses based on theor name.
+ SetVector<GlobalVariable *> Ordered = SortByName(
+ std::vector<GlobalVariable *>(Variables.begin(), Variables.end()));
+ return std::move(Ordered);
+}
+
+SetVector<Function *> AMDGPUSwLowerLDS::GetOrderedIndirectLDSAccessingKernels(
+ SetVector<Function *> &&Kernels) {
+ // Sort the non-kernels accessing LDS based on theor name.
+ // Also assign a kernel ID metadata based on the sorted order.
+ LLVMContext &Ctx = M.getContext();
+ if (Kernels.size() > UINT32_MAX) {
+ // 32 bit keeps it in one SGPR. > 2**32 kernels won't fit on the GPU
+ report_fatal_error("Unimplemented SW LDS lowering for > 2**32 kernels");
+ }
+ SetVector<Function *> OrderedKernels =
+ SortByName(std::vector<Function *>(Kernels.begin(), Kernels.end()));
+ for (size_t i = 0; i < Kernels.size(); i++) {
+ Metadata *AttrMDArgs[1] = {
+ ConstantAsMetadata::get(IRB.getInt32(i)),
+ };
+ Function *Func = OrderedKernels[i];
+ Func->setMetadata("llvm.amdgcn.lds.kernel.id",
+ MDNode::get(Ctx, AttrMDArgs));
+ auto &LDSParams = KernelToLDSParametersMap[Func];
+ assert(LDSParams);
+ LDSParams->KernelId = i;
+ }
+ return std::move(OrderedKernels);
+}
+
+void AMDGPUSwLowerLDS::GetUsesOfLDSByNonKernels(
+ CallGraph const &CG, FunctionVariableMap &functions) {
+ // Get uses from the current function, excluding uses by called functions
+ // Two output variables to avoid walking the globals list twice
+ for (auto &GV : M.globals()) {
+ if (!AMDGPU::isLDSVariableToLower(GV)) {
+ continue;
+ }
+
+ if (GV.isAbsoluteSymbolRef()) {
+ report_fatal_error(
+ "LDS variables with absolute addresses are unimplemented.");
+ }
+
+ for (User *V : GV.users()) {
+ User *FUU = V;
+ bool isCast = isa<BitCastOperator, AddrSpaceCastOperator>(FUU);
+ if (isCast && FUU->hasOneUse() && !FUU->user_begin()->user_empty())
+ FUU = *FUU->user_begin();
+ if (auto *I = dyn_cast<Instruction>(FUU)) {
+ Function *F = I->getFunction();
+ if (!isKernelLDS(F)) {
+ functions[F].insert(&GV);
+ }
+ }
+ }
+ }
+}
+
+void AMDGPUSwLowerLDS::PopulateMallocLDSGlobal(Function *Func) {
+ // Create new LDS global required for each kernel to store
+ // device global memory pointer.
+ auto &LDSParams = KernelToLDSParametersMap[Func];
+ assert(LDSParams);
+ // create new global pointer variable
+ LDSParams->MallocLDSGlobal = new GlobalVariable(
+ M, IRB.getPtrTy(), false, GlobalValue::InternalLinkage,
+ PoisonValue::get(IRB.getPtrTy()),
+ Twine("llvm.amdgcn.sw.lds." + F...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/87265
More information about the llvm-commits
mailing list