Check number of calls per kernel for inlining

For number of calls check for inlining check only calls in kernel for
given call site.

Author: stefan-il

IGC/Compiler/CISACodeGen/EstimateFunctionSize.cpp

@@ -28,7 +28,6 @@ SPDX-License-Identifier: MIT
 #include "common/LLVMWarningsPop.hpp"
 #include "Probe/Assertion.h"
 #include <deque>
-#include <iostream>
 #include <cfloat>
 #include <algorithm>
 #include <cmath>
@@ -1112,26 +1111,24 @@ std::size_t EstimateFunctionSize::getExpandedSize(const Function* F) const {
     return std::numeric_limits<std::size_t>::max();
 }
 
-bool EstimateFunctionSize::onlyCalledOnce(const Function* F) {
+bool EstimateFunctionSize::onlyCalledOnce(const Function* F, const Function* CallerF) {
     //IGC_ASSERT(IGC_IS_FLAG_DISABLED(ControlKernelTotalSize));
     auto I = ECG.find((Function*)F);
     if (I != ECG.end()) {
-        FunctionNode* Node = (FunctionNode*)I->second;
+        auto* Node = (FunctionNode*)I->second;
         IGC_ASSERT(F == Node->F);
         // one call-site and not a recursion
         if (Node->CallerList.size() == 1 &&
             Node->CallerList.begin()->second == 1 &&
             Node->CallerList.begin()->first != Node) {
             return true;
         }
-        // OpenCL specific, called once by each kernel
-        auto MdWrapper = getAnalysisIfAvailable<MetaDataUtilsWrapper>();
+        // OpenCL specific, called once by passed kernel
+        auto* MdWrapper = getAnalysisIfAvailable<MetaDataUtilsWrapper>();
         if (MdWrapper) {
-            auto pMdUtils = MdWrapper->getMetaDataUtils();
-            for (const auto &node : Node->CallerList) {
-                FunctionNode* Caller = node.first;
-                uint32_t cnt = node.second;
-                if (cnt > 1) {
+            auto* pMdUtils = MdWrapper->getMetaDataUtils();
+            for (const auto &[Caller, CallCount] : Node->CallerList) {
+                if (CallCount > 1 && Caller->F == CallerF) {
                     return false;
                 }
                 if (!isEntryFunc(pMdUtils, Caller->F)) {

IGC/Compiler/CISACodeGen/EstimateFunctionSize.h

@@ -13,11 +13,10 @@ SPDX-License-Identifier: MIT
 #include "llvm/Pass.h"
 #include <llvm/IR/InstVisitor.h>
 #include <llvm/ADT/StringRef.h>
-#include "llvm/ADT/DenseSet.h"
 #include "llvm/Support/ScaledNumber.h"
 #include "common/LLVMWarningsPop.hpp"
-#include <cstddef>
 #include "Probe/Assertion.h"
+#include <cstddef>
 #include <deque>
 #include <unordered_map>
 #include <unordered_set>
@@ -46,7 +45,7 @@ namespace IGC {
         ~EstimateFunctionSize();
         EstimateFunctionSize(const EstimateFunctionSize&) = delete;
         EstimateFunctionSize& operator=(const EstimateFunctionSize&) = delete;
-        virtual llvm::StringRef getPassName() const  override { return "Estimate Function Sizes"; }
+        llvm::StringRef getPassName() const  override { return "Estimate Function Sizes"; }
         void getAnalysisUsage(llvm::AnalysisUsage& AU) const override;
         bool runOnModule(llvm::Module& M) override;
 
@@ -56,7 +55,7 @@ namespace IGC {
         /// \brief Return the estimated function size after complete inlining.
         std::size_t getExpandedSize(const llvm::Function* F) const;
 
-        bool onlyCalledOnce(const llvm::Function* F);
+        bool onlyCalledOnce(const llvm::Function* F, const llvm::Function* CallerF);
 
         bool hasRecursion() const { return HasRecursion; }
 

IGC/Compiler/CISACodeGen/GenCodeGenModule.cpp

@@ -17,7 +17,6 @@ SPDX-License-Identifier: MIT
 #include "Compiler/MetaDataUtilsWrapper.h"
 #include "common/igc_regkeys.hpp"
 #include "common/LLVMWarningsPush.hpp"
-#include "llvm/Config/llvm-config.h"
 #include "llvm/IR/Argument.h"
 #include "llvm/IR/Attributes.h"
 #include "llvm/ADT/SetVector.h"
@@ -32,14 +31,12 @@ SPDX-License-Identifier: MIT
 #include "llvm/Transforms/IPO.h"
 #include "llvm/Transforms/IPO/Inliner.h"
 #include "llvmWrapper/IR/DerivedTypes.h"
-#include "llvmWrapper/Transforms/Utils/Cloning.h"
 #include "llvm/IR/DebugInfo.h"
 #include "llvm/IR/DIBuilder.h"
 #include "common/LLVMWarningsPop.hpp"
 #include "DebugInfo/VISADebugEmitter.hpp"
 #include <numeric>
 #include <utility>
-#include <iostream>
 #include "Probe/Assertion.h"
 
 using namespace llvm;
@@ -1458,11 +1455,11 @@ InlineCost SubroutineInliner::getInlineCost(IGCLLVM::CallSiteRef CS)
             {
                 return llvm::InlineCost::getAlways("Caller size smaller than per func. threshold");
             }
-            else if (isTrivialCall(Callee) || FSA->onlyCalledOnce(Callee))
+            if (isTrivialCall(Callee) || FSA->onlyCalledOnce(Callee, Caller))
             {
                 return llvm::InlineCost::getAlways("Callee is called only once");
             }
-            else if (!FSA->shouldEnableSubroutine())
+            if (!FSA->shouldEnableSubroutine())
             {
                 // This function returns true if the estimated total inlining size exceeds some module threshold.
                 // If we don't exceed it, and there's no preference on inline vs noinline, we just inline.

IGC/ocloc_tests/features/inline/inline_functions_multi_kernel.ll

@@ -0,0 +1,73 @@
+;=========================== begin_copyright_notice ============================
+;
+; Copyright (C) 2025 Intel Corporation
+;
+; SPDX-License-Identifier: MIT
+;
+;============================ end_copyright_notice =============================
+
+; REQUIRES: regkeys,pvc-supported,llvm-14-plus
+
+; Check that functions called once from kernel are inlined in it even if they are called multiple times from other kernels.
+
+; RUN: llvm-as %s -o %t.bc
+; RUN: ocloc compile -llvm_input -file %t.bc -options "-igc_opts 'DisableRecompilation=1, SubroutineThreshold=50, SubroutineInlinerThreshold=10, KernelTotalSizeThreshold=50, PrintToConsole=1, PrintBefore=EmitPass'" -device pvc 2>&1 | FileCheck %s
+; CHECK-LABEL: @_ZTS28Kernel_A_Supposed_2B_Inlined(
+; CHECK-NOT: call spir_func void @testInlineFn
+; CHECK: ret void
+
+; CHECK-LABEL: @_ZTS28Kernel_B_Supposed_2B_Inlined(
+; CHECK: call spir_func void @testInlineFn
+; CHECK: call spir_func void @testInlineFn
+; CHECK: ret void
+
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024"
+target triple = "spir64-unknown-unknown"
+
+define spir_kernel void @_ZTS28Kernel_A_Supposed_2B_Inlined() {
+  call spir_func void @testInlineFn()
+  ret void
+}
+
+define linkonce_odr spir_func void @testInlineFn() {
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  call spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+  ret void
+}
+
+declare spir_func void @_Z8mulAccumRP4DataRK5ColorS4_()
+
+define spir_kernel void @_ZTS28Kernel_B_Supposed_2B_Inlined() {
+  call spir_func void @testInlineFn()
+  call spir_func void @testInlineFn()
+  ret void
+}
+
+; uselistorder directives
+uselistorder void ()* @_Z8mulAccumRP4DataRK5ColorS4_, { 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0 }