Media/intel-graphics-compiler
1
0
Fork 0
mirror of https://github.com/intel/intel-graphics-compiler.git synced 2025-10-30 08:18:26 +08:00
Code Issues Packages Projects Releases Wiki Activity

Improve disable inlining pass

Browse Source 
Remove always inline attribute from call instructions to prevent
LLVM inliner from inlining them.
This commit is contained in:
Stefan Ilic
2025-10-27 14:48:38 +00:00
committed by igcbot
parent e3fc880273
commit 6057876654
4 changed files with 20 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
IGC/Compiler/Optimizer/OpenCLPasses/DisableInlining/DisableInlining.cpp
View File

@ -12,6 +12,8 @@ SPDX-License-Identifier: MIT
#include "common/LLVMWarningsPush.hpp"
#include <llvm/IR/Function.h>
#include <llvm/IR/InstIterator.h>
#include <llvm/IR/Instructions.h>
#include "common/LLVMWarningsPop.hpp"
using namespace llvm;
@ -38,5 +40,14 @@ bool DisableInlining::runOnFunction(llvm::Function &F) {
F.removeFnAttr(llvm::Attribute::AlwaysInline);
}
F.addFnAttr(llvm::Attribute::NoInline);
for (auto &I : llvm::instructions(F)) {
if (auto *CI = dyn_cast<CallInst>(&I)) {
if (CI->hasFnAttr(llvm::Attribute::AlwaysInline)) {
CI->removeFnAttr(llvm::Attribute::AlwaysInline);
}
}
}
return true;
}
}

16
IGC/common/igc_flags.h
View File

@ -1330,19 +1330,17 @@ DECLARE_IGC_REGKEY(DWORD, ForceGroupSizeX, 8, "force group size along X", false)
DECLARE_IGC_REGKEY(DWORD, ForceGroupSizeY, 8, "force group size along Y", false)
DECLARE_IGC_REGKEY(bool, EnableThreadCombiningWithNoSLM, false, "Enable thread combining opt for shader without SLM",
false)
DECLARE_IGC_REGKEY(bool, DisableInlining, false, "Disable inlining of all functions", true)
DECLARE_IGC_REGKEY(bool, EnableDropTargetFunctions, false, "Enables pass for dropping targeted functions", true)
DECLARE_IGC_REGKEY(bool, DisableInlining, false, "Disable inlining of all functions", false)
DECLARE_IGC_REGKEY(bool, EnableDropTargetFunctions, false, "Enables pass for dropping targeted functions", false)
DECLARE_IGC_REGKEY(bool, VerboseDropTargetFunctions, false, "Enables verbose logging for dropping targeted functions",
true)
false)
DECLARE_IGC_REGKEY(bool, CrashOnDroppedFnAccess, false, "Enables crash on access to dropped functions",
true)
DECLARE_IGC_REGKEY(debugString, DropTargetFnListPath, 0, "Path to folder with lists of functions to drop", true)
DECLARE_IGC_REGKEY(bool, EnableDropTargetBBs, false, "Enables pass for dropping targeted BBs", true)
DECLARE_IGC_REGKEY(debugString, DropTargetFnListPath, 0, "Path to folder with lists of functions to drop", false)
DECLARE_IGC_REGKEY(bool, EnableDropTargetBBs, false, "Enables pass for dropping targeted BBs", false)
DECLARE_IGC_REGKEY(bool, VerboseDropTargetBBs, false, "Enables verbose logging for dropping targeted BBs",
true)
DECLARE_IGC_REGKEY(bool, CrashOnDroppedBBAccess, false, "Enables crash on access to dropped BBs",
true)
DECLARE_IGC_REGKEY(debugString, DropTargetBBListPath, 0, "Path to folder with lists of BBs to drop", true)
false)
DECLARE_IGC_REGKEY(debugString, DropTargetBBListPath, 0, "Path to folder with lists of BBs to drop", false)
DECLARE_IGC_REGKEY(DWORD, PrintFunctionSizeAnalysis, 0, "Print analysis data of function sizes", true)
DECLARE_IGC_REGKEY(DWORD, SubroutineThreshold, 110000, "Minimal kernel size to enable subroutines", false)
DECLARE_IGC_REGKEY(DWORD, SubroutineInlinerThreshold, 3000, "Subroutine inliner threshold", false)

2
IGC/ocloc_tests/features/inline/no_inline.ll
View File

@ -35,7 +35,7 @@ define linkonce_odr spir_func void @testInlineFn2(i32 %d) {
}
define spir_kernel void @testKernel(i32 %a, i32 %b, i32 %c) {
call spir_func void @testInlineFn1(i32 %b)
call spir_func void @testInlineFn1(i32 %b) #0
call spir_func void @testInlineFn2(i32 %c)
ret void
}

1
documentation/configuration_flags.md
View File

@ -702,7 +702,6 @@ $ export IGC_ShaderDumpEnable=1
| `ControlInlineTinySizeForSPGT` | Tiny function size for controlling kernel total size | Available |
| `ControlKernelTotalSize` | Control kernel total size | Available |
| `ControlUnitSize` | Control compilation unit size by unit trimming | Available |
| `CrashOnDroppedBBAccess` | Enables crash on access to dropped BBs | Available |
| `CrashOnDroppedFnAccess` | Enables crash on access to dropped functions | Available |
| `DelayEmuInt64AddLimit` | Delay emulating Int64 Add operations in vISA | - |
| `DetectCastToGAS` | Check if the module contains local/private to GAS (Gerneric Address Space) cast, it also check internal flags | Available |