https://github.com/jhuber6 created https://github.com/llvm/llvm-project/pull/201585
Summary: This is related to reported regressions in the GROMACS suite when offloading to AMDGCN devices through the RDC / LTO interface. The application intentionally passes `-fno-slp-vectorize` to disable that pass, but there's currently no way to do this through the LTO pipline. This PR adds a few option to `lld` that suppresses SLP from being added to the pipeline. >From 5c50531b5ff8c2809f9becf58cbcf88dca32bae7 Mon Sep 17 00:00:00 2001 From: Joseph Huber <[email protected]> Date: Thu, 4 Jun 2026 08:20:30 -0500 Subject: [PATCH] [lld][LTO] Respect `-fno-slp-vectorize` for the LTO pipeline Summary: This is related to reported regressions in the GROMACS suite when offloading to AMDGCN devices through the RDC / LTO interface. The application intentionally passes `-fno-slp-vectorize` to disable that pass, but there's currently no way to do this through the LTO pipline. This PR adds a few option to `lld` that suppresses SLP from being added to the pipeline. --- clang/lib/Driver/ToolChains/Clang.cpp | 4 +++- clang/lib/Driver/ToolChains/CommonArgs.cpp | 11 +++++++++++ clang/test/Driver/lto.c | 8 ++++++++ lld/ELF/Config.h | 1 + lld/ELF/Driver.cpp | 2 ++ lld/ELF/LTO.cpp | 2 +- lld/ELF/Options.td | 7 +++++++ lld/test/ELF/lto/slp-vectorize-pm.ll | 3 +++ 8 files changed, 36 insertions(+), 2 deletions(-) diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp index 25bf84a2bde33..55519f6e1089e 100644 --- a/clang/lib/Driver/ToolChains/Clang.cpp +++ b/clang/lib/Driver/ToolChains/Clang.cpp @@ -9601,7 +9601,9 @@ void LinkerWrapper::ConstructJob(Compilation &C, const JobAction &JA, OPT_fsanitize_minimal_runtime, OPT_fno_sanitize_minimal_runtime, OPT_fsanitize_trap_EQ, - OPT_fno_sanitize_trap_EQ}; + OPT_fno_sanitize_trap_EQ, + OPT_fslp_vectorize, + OPT_fno_slp_vectorize}; const llvm::DenseSet<unsigned> LinkerOptions{OPT_mllvm, OPT_Zlinker_input}; auto ToolChainHasRT = [&](const ToolChain &TC, StringRef Name) { return TC.getVFS().exists( diff --git a/clang/lib/Driver/ToolChains/CommonArgs.cpp b/clang/lib/Driver/ToolChains/CommonArgs.cpp index 47953bc3a23b5..88175762f7d30 100644 --- a/clang/lib/Driver/ToolChains/CommonArgs.cpp +++ b/clang/lib/Driver/ToolChains/CommonArgs.cpp @@ -1157,6 +1157,17 @@ void tools::addLTOOptions(const ToolChain &ToolChain, const ArgList &Args, CmdArgs.push_back(Args.MakeArgString(Twine(PluginOptPrefix) + ParallelismOpt + Parallelism)); + // Forward SLP vectorization preference to the LTO backend. + if (Arg *A = Args.getLastArg(options::OPT_fslp_vectorize, + options::OPT_fno_slp_vectorize)) { + if (A->getOption().matches(options::OPT_fslp_vectorize)) + CmdArgs.push_back( + Args.MakeArgString(Twine(PluginOptPrefix) + "slp-vectorize")); + else + CmdArgs.push_back( + Args.MakeArgString(Twine(PluginOptPrefix) + "no-slp-vectorize")); + } + // Pass down GlobalISel options. if (Arg *A = Args.getLastArg(options::OPT_fglobal_isel, options::OPT_fno_global_isel)) { diff --git a/clang/test/Driver/lto.c b/clang/test/Driver/lto.c index 81165d3b9e8a3..44514c0cd830e 100644 --- a/clang/test/Driver/lto.c +++ b/clang/test/Driver/lto.c @@ -117,6 +117,14 @@ // CHECK-GISEL: "-plugin-opt=-global-isel=1" // CHECK-DISABLE-GISEL: "-plugin-opt=-global-isel=0" +// RUN: %clang --target=x86_64-unknown-linux-gnu -### %s -flto -fno-slp-vectorize 2> %t +// RUN: FileCheck --check-prefix=CHECK-NO-SLP < %t %s +// RUN: %clang --target=x86_64-unknown-linux-gnu -### %s -flto -fslp-vectorize 2> %t +// RUN: FileCheck --check-prefix=CHECK-SLP < %t %s +// +// CHECK-NO-SLP: "-plugin-opt=no-slp-vectorize" +// CHECK-SLP: "-plugin-opt=slp-vectorize" + // -flto passes -time-passes when -ftime-report is passed // RUN: %clang --target=x86_64-unknown-linux-gnu -### %s -flto -ftime-report 2> %t // RUN: FileCheck --check-prefix=CHECK-TIME-REPORT < %t %s diff --git a/lld/ELF/Config.h b/lld/ELF/Config.h index 12b16ded61fca..58eb12d87a36e 100644 --- a/lld/ELF/Config.h +++ b/lld/ELF/Config.h @@ -373,6 +373,7 @@ struct Config { bool ltoPGOWarnMismatch; bool ltoDebugPassManager; bool ltoEmitAsm; + bool ltoSLPVectorize; bool ltoUniqueBasicBlockSectionNames; bool ltoValidateAllVtablesHaveTypeInfos; bool ltoWholeProgramVisibility; diff --git a/lld/ELF/Driver.cpp b/lld/ELF/Driver.cpp index 7ec7dfcae6bca..9d1981986969f 100644 --- a/lld/ELF/Driver.cpp +++ b/lld/ELF/Driver.cpp @@ -1468,6 +1468,8 @@ static void readConfigs(Ctx &ctx, opt::InputArgList &args) { ctx.arg.ltoDebugPassManager = args.hasArg(OPT_lto_debug_pass_manager); ctx.arg.ltoEmitAsm = args.hasArg(OPT_lto_emit_asm); ctx.arg.ltoNewPmPasses = args.getLastArgValue(OPT_lto_newpm_passes); + ctx.arg.ltoSLPVectorize = + args.hasFlag(OPT_lto_slp_vectorize, OPT_no_lto_slp_vectorize, true); ctx.arg.ltoWholeProgramVisibility = args.hasFlag(OPT_lto_whole_program_visibility, OPT_no_lto_whole_program_visibility, false); diff --git a/lld/ELF/LTO.cpp b/lld/ELF/LTO.cpp index e40575bffec62..804f6959f87fc 100644 --- a/lld/ELF/LTO.cpp +++ b/lld/ELF/LTO.cpp @@ -116,7 +116,7 @@ static lto::Config createConfig(Ctx &ctx) { c.CGOptLevel = ctx.arg.ltoCgo; c.PTO.LoopVectorization = c.OptLevel > 1; - c.PTO.SLPVectorization = c.OptLevel > 1; + c.PTO.SLPVectorization = ctx.arg.ltoSLPVectorize && c.OptLevel > 1; // Set up a custom pipeline if we've been asked to. c.OptPipeline = std::string(ctx.arg.ltoNewPmPasses); diff --git a/lld/ELF/Options.td b/lld/ELF/Options.td index 64c42eb49607d..6b9ab89b9d987 100644 --- a/lld/ELF/Options.td +++ b/lld/ELF/Options.td @@ -672,6 +672,9 @@ defm lto_known_safe_vtables : EEq<"lto-known-safe-vtables", def lto_obj_path_eq: JJ<"lto-obj-path=">; def lto_sample_profile: JJ<"lto-sample-profile=">, HelpText<"Sample profile file path">; +defm lto_slp_vectorize: BB<"lto-slp-vectorize", + "Enable SLP vectorization during LTO (default at O2+)", + "Disable SLP vectorization during LTO">; defm lto_validate_all_vtables_have_type_infos: BB<"lto-validate-all-vtables-have-type-infos", "Validate that all vtables have type infos for LTO link", "Do not validate that all vtables have type infos for LTO link">; @@ -776,6 +779,10 @@ def: J<"plugin-opt=opt-remarks-hotness-threshold=">, def: J<"plugin-opt=sample-profile=">, Alias<lto_sample_profile>, HelpText<"Alias for --lto-sample-profile">; def: F<"plugin-opt=save-temps">, Alias<save_temps>, HelpText<"Alias for --save-temps">; +def: F<"plugin-opt=slp-vectorize">, + Alias<lto_slp_vectorize>, HelpText<"Alias for --lto-slp-vectorize">; +def: F<"plugin-opt=no-slp-vectorize">, + Alias<no_lto_slp_vectorize>, HelpText<"Alias for --no-lto-slp-vectorize">; def plugin_opt_stats_file: J<"plugin-opt=stats-file=">, HelpText<"Filename to write LTO statistics to">; def: J<"plugin-opt=time-trace=">, Alias<time_trace_eq>, HelpText<"Alias for --time-trace=">; diff --git a/lld/test/ELF/lto/slp-vectorize-pm.ll b/lld/test/ELF/lto/slp-vectorize-pm.ll index 2171092569536..f0576c98d9f46 100644 --- a/lld/test/ELF/lto/slp-vectorize-pm.ll +++ b/lld/test/ELF/lto/slp-vectorize-pm.ll @@ -14,10 +14,13 @@ ; RUN: ld.lld --plugin-opt=debug-pass-manager --plugin-opt=O3 --plugin-opt=save-temps -shared -o %t4.o %t.o 2>&1 | FileCheck %s --check-prefix=CHECK-O3-SLP ; RUN: llvm-dis %t.o.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-O3-LPV +; RUN: ld.lld --plugin-opt=debug-pass-manager --plugin-opt=O2 --no-lto-slp-vectorize -shared -o %t5.o %t.o 2>&1 | FileCheck %s --check-prefix=CHECK-O2-NOSLP + ; CHECK-O0-SLP-NOT: Running pass: SLPVectorizerPass ; CHECK-O1-SLP-NOT: Running pass: SLPVectorizerPass ; CHECK-O2-SLP: Running pass: SLPVectorizerPass ; CHECK-O3-SLP: Running pass: SLPVectorizerPass +; CHECK-O2-NOSLP-NOT: Running pass: SLPVectorizerPass ; CHECK-O0-LPV-NOT: = !{!"llvm.loop.isvectorized", i32 1} ; CHECK-O1-LPV-NOT: = !{!"llvm.loop.isvectorized", i32 1} ; CHECK-O2-LPV: = !{!"llvm.loop.isvectorized", i32 1} _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
