alexfh added a comment.
Matt, did you see my comments on https://reviews.llvm.org/rGbb70b5d40652207c0bd3d385def10ef3ef1d45b4? The summary is that clang codegen time has significantly increased on some of our code after this patch, pushing the compile times far beyond the time limit. Here's a reduced test case: F24611479: q.cc <https://reviews.llvm.org/F24611479>
This is how the output of `clang --target=x86_64--linux-gnu -O1 -c q.cc -ftime-report` changes after this commit:
diff -u clang-llvmorg-16-init-4458-gc812b4a1d895.time clang-llvmorg-16-init-4459-gbb70b5d40652.time
--- clang-llvmorg-16-init-4458-gc812b4a1d895.time 2022-09-19 13:41:47.202346523 +0200
+++ clang-llvmorg-16-init-4459-gbb70b5d40652.time 2022-09-19 13:41:27.602138226 +0200
@@ -1,99 +1,99 @@
===-------------------------------------------------------------------------===
... Pass execution timing report ...
===-------------------------------------------------------------------------===
- Total Execution Time: 2.9992 seconds (2.9992 wall clock)
+ Total Execution Time: 2.9692 seconds (2.9693 wall clock)
...
===-------------------------------------------------------------------------===
Miscellaneous Ungrouped Timers
===-------------------------------------------------------------------------===
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name ---
- 3.2616 ( 97.5%) 0.4235 ( 98.4%) 3.6851 ( 97.6%) 3.6854 ( 97.6%) Code Generation Time
- 0.0839 ( 2.5%) 0.0070 ( 1.6%) 0.0909 ( 2.4%) 0.0909 ( 2.4%) LLVM IR Generation Time
- 3.3454 (100.0%) 0.4305 (100.0%) 3.7759 (100.0%) 3.7762 (100.0%) Total
+ 14.4539 ( 99.4%) 0.3530 ( 97.1%) 14.8068 ( 99.4%) 14.8079 ( 99.4%) Code Generation Time
+ 0.0818 ( 0.6%) 0.0105 ( 2.9%) 0.0924 ( 0.6%) 0.0924 ( 0.6%) LLVM IR Generation Time
+ 14.5357 (100.0%) 0.3635 (100.0%) 14.8992 (100.0%) 14.9002 (100.0%) Total
...
===-------------------------------------------------------------------------===
... Pass execution timing report ...
===-------------------------------------------------------------------------===
- Total Execution Time: 2.0628 seconds (2.0630 wall clock)
+ Total Execution Time: 13.1945 seconds (13.1955 wall clock)
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name ---
- 0.8411 ( 51.2%) 0.2109 ( 50.3%) 1.0520 ( 51.0%) 1.0521 ( 51.0%) X86 DAG->DAG Instruction Selection
...
+ 12.0451 ( 93.7%) 0.1443 ( 42.5%) 12.1893 ( 92.4%) 12.1902 ( 92.4%) X86 DAG->DAG Instruction Selection
...
===-------------------------------------------------------------------------===
Clang front-end time report
===-------------------------------------------------------------------------===
- Total Execution Time: 3.9323 seconds (3.9326 wall clock)
+ Total Execution Time: 15.0529 seconds (15.0539 wall clock)
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name ---
- 3.4902 (100.0%) 0.4421 (100.0%) 3.9323 (100.0%) 3.9326 (100.0%) Clang front-end timer
- 3.4902 (100.0%) 0.4421 (100.0%) 3.9323 (100.0%) 3.9326 (100.0%) Total
+ 14.6894 (100.0%) 0.3635 (100.0%) 15.0529 (100.0%) 15.0539 (100.0%) Clang front-end timer
+ 14.6894 (100.0%) 0.3635 (100.0%) 15.0529 (100.0%) 15.0539 (100.0%) Total
Clang's execution time profile looks like this:
- 99.51% 0.00% clang-llvmorg-1 clang-llvmorg-16-init-4459-gbb70b5d40652 [.] clang_main ▒
- clang_main ▒
- 99.50% clang::driver::Driver::ExecuteCompilation ▒
clang::driver::Compilation::ExecuteJobs ▒
clang::driver::Compilation::ExecuteCommand ▒
clang::driver::CC1Command::Execute ▒
llvm::CrashRecoveryContext::RunSafely ▒
llvm::function_ref<void ()>::callback_fn<clang::driver::CC1Command::Execute(llvm::ArrayRef<llvm::Optional<llvm::StringRef> >, std::__u::basic_string<char, std::__u::char_traits<char>, std::__u::allocator<char> >*, bool*) const::$_1> ▒
ExecuteCC1Tool ▒
- cc1_main ▒
- 99.47% clang::ExecuteCompilerInvocation ▒
- 99.47% clang::CompilerInstance::ExecuteAction ▒
- 99.46% clang::FrontendAction::Execute ▒
- clang::ParseAST ▒
- 97.83% clang::BackendConsumer::HandleTranslationUnit ▒
- 97.83% clang::EmitBackendOutput ▒
- 87.02% llvm::legacy::PassManagerImpl::run ▒
- 86.96% llvm::FPPassManager::runOnModule ▒
- llvm::FPPassManager::runOnFunction ▒
- 85.90% llvm::MachineFunctionPass::runOnFunction ▒
- 80.45% (anonymous namespace)::X86DAGToDAGISel::runOnMachineFunction ▒
- llvm::SelectionDAGISel::runOnMachineFunction ▒
- 80.30% llvm::SelectionDAGISel::SelectAllBasicBlocks ▒
- 73.55% llvm::SelectionDAGISel::SelectBasicBlock ▒
- 73.53% llvm::SelectionDAGBuilder::visit ▒
- 72.72% llvm::SelectionDAGBuilder::visitLoad ▒
- 72.30% llvm::isDereferenceableAndAlignedPointer ▒
- 72.29% isDereferenceableAndAlignedPointer ▒
- 63.27% isDereferenceableAndAlignedPointer ▒
llvm::getKnowledgeForValue ▒
9.00% llvm::getKnowledgeForValue ▒
+ 6.70% llvm::SelectionDAGISel::CodeGenAndEmitDAG ▒
+ 1.88% llvm::X86AsmPrinter::runOnMachineFunction ▒
+ 10.79% (anonymous namespace)::EmitAssemblyHelper::RunOptimizationPipeline ▒
+ 1.03% clang::Parser::ParseTopLevelDecl ▒
+ 0.60% clang::BackendConsumer::HandleTopLevelDecl ▒
Before the commit it looked like this:
- 98.11% 0.00% clang-llvmorg-1 clang-llvmorg-16-init-4458-gc812b4a1d895 [.] clang_main ▒
- clang_main ▒
- 98.07% clang::driver::Driver::ExecuteCompilation ▒
clang::driver::Compilation::ExecuteJobs ▒
clang::driver::Compilation::ExecuteCommand ▒
clang::driver::CC1Command::Execute ▒
llvm::CrashRecoveryContext::RunSafely ▒
llvm::function_ref<void ()>::callback_fn<clang::driver::CC1Command::Execute(llvm::ArrayRef<llvm::Optional<llvm::StringRef> >, std::__u::basic_string<char, std::__u::char_traits<char>, std::__u::allocator<char> >*, bool*) const::$_1> ▒
ExecuteCC1Tool ▒
- cc1_main ▒
- 97.97% clang::ExecuteCompilerInvocation ▒
- clang::CompilerInstance::ExecuteAction ▒
- 97.95% clang::FrontendAction::Execute ▒
- clang::ParseAST ▒
- 91.46% clang::BackendConsumer::HandleTranslationUnit ▒
- 91.46% clang::EmitBackendOutput ▒
- 50.30% llvm::legacy::PassManagerImpl::run ▒
- 50.11% llvm::FPPassManager::runOnModule ▒
- llvm::FPPassManager::runOnFunction ▒
- 46.10% llvm::MachineFunctionPass::runOnFunction ▒
- 25.78% (anonymous namespace)::X86DAGToDAGISel::runOnMachineFunction ▒
- llvm::SelectionDAGISel::runOnMachineFunction ▒
- 25.27% llvm::SelectionDAGISel::SelectAllBasicBlocks ▒
- 22.34% llvm::SelectionDAGISel::CodeGenAndEmitDAG ▒
- 8.32% llvm::SelectionDAG::Combine ▒
- 5.12% (anonymous namespace)::DAGCombiner::combine ▒
+ 1.24% (anonymous namespace)::DAGCombiner::visitADD ▒
0.75% (anonymous namespace)::DAGCombiner::visitOR ▒
0.65% (anonymous namespace)::DAGCombiner::visitAND ▒
0.52% llvm::X86TargetLowering::PerformDAGCombine ▒
0.51% (anonymous namespace)::DAGCombiner::visitXOR ▒
1.19% (anonymous namespace)::DAGCombiner::AddToWorklist ▒
+ 3.13% llvm::Timer::stopTimer ▒
+ 2.73% llvm::Timer::startTimer ▒
+ 2.55% llvm::SelectionDAGISel::DoInstructionSelection ▒
+ 1.67% (anonymous namespace)::ScheduleDAGRRList::Schedule ▒
+ 1.31% llvm::ScheduleDAGSDNodes::EmitSchedule ▒
+ 0.88% llvm::SelectionDAG::LegalizeTypes ▒
0.74% llvm::SelectionDAG::Legalize ▒
+ 2.73% llvm::SelectionDAGISel::SelectBasicBlock ▒
- 7.05% llvm::X86AsmPrinter::runOnMachineFunction ▒
- 7.05% llvm::AsmPrinter::emitFunctionBody ▒
- 3.24% llvm::Timer::startTimer ▒
- 3.06% __GI___getrusage ▒
+ 2.52% entry_SYSCALL_64 ▒
0.53% syscall_return_via_sysret ▒
+ 3.08% llvm::Timer::stopTimer ▒
0.51% llvm::X86AsmPrinter::emitInstruction ▒
+ 1.23% (anonymous namespace)::MachineScheduler::runOnMachineFunction ▒
+ 1.04% (anonymous namespace)::DeadMachineInstructionElim::runOnMachineFunction ▒
+ 1.00% llvm::MachineDominatorTree::runOnMachineFunction ▒
+ 0.83% (anonymous namespace)::RegisterCoalescer::runOnMachineFunction ▒
0.74% llvm::RAGreedy::runOnMachineFunction ▒
0.70% (anonymous namespace)::MachineCSE::runOnMachineFunction ▒
+ 0.63% llvm::LiveIntervals::runOnMachineFunction ▒
This is currently blocking our internal release. If you don't see an obvious fix, could you revert the patch while working on this?
Thanks!
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D130042/new/
https://reviews.llvm.org/D130042