https://github.com/jhuber6 created https://github.com/llvm/llvm-project/pull/206961
I believe I fixed the immediate issues that were causing regressions with this, and now this is leading to its own regressions in building Pytorch apparently. This was explicitly marked as a hack, and this should resolve an ongoing issue so we should revert this for now. Should fix LCOMPILER-2396 This reverts commit eaff67c9838f98d221aa563c9e5d4e714d636674. >From ed7b0464594311ab47d8b99e0fbc430623fe1911 Mon Sep 17 00:00:00 2001 From: Joseph Huber <[email protected]> Date: Wed, 1 Jul 2026 06:57:30 -0500 Subject: [PATCH] Revert "[HIP][AMDGPU] Use non-LTO pipeline for non-RDC in the linker wrapper (#201135)" I believe I fixed the immediate issues that were causing regressions with this, and now this is leading to its own regressions in building Pytorch apparently. This was explicitly marked as a hack, and this should resolve an ongoing issue so we should revert this for now. Should fix LCOMPILER-2396 This reverts commit eaff67c9838f98d221aa563c9e5d4e714d636674. --- clang/lib/Driver/ToolChains/Clang.cpp | 13 ------------- clang/test/Driver/hip-toolchain-no-rdc.hip | 17 ----------------- .../linker-wrapper-hip-no-rdc.c | 12 ------------ .../clang-linker-wrapper/ClangLinkerWrapper.cpp | 13 ------------- .../clang-linker-wrapper/LinkerWrapperOpts.td | 3 --- 5 files changed, 58 deletions(-) diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp index bb3fbc3c585a6..475dcfcb4126c 100644 --- a/clang/lib/Driver/ToolChains/Clang.cpp +++ b/clang/lib/Driver/ToolChains/Clang.cpp @@ -9786,19 +9786,6 @@ void LinkerWrapper::ConstructJob(Compilation &C, const JobAction &JA, } } } - - if (JA.getType() == types::TY_HIP_FATBIN && Kind == Action::OFK_HIP) { - // Non-RDC HIP uses the conventional non-LTO pipeline unless the user - // opts into offload LTO. - bool UsesProfileGenerate = Args.hasArg( - options::OPT_fprofile_generate, options::OPT_fprofile_generate_EQ, - options::OPT_fprofile_instr_generate, - options::OPT_fprofile_instr_generate_EQ); - if (!Args.hasArg(options::OPT_foffload_lto_EQ, - options::OPT_fno_offload_lto) && - !UsesProfileGenerate && !TC->getTriple().isSPIRV()) - CmdArgs.push_back("--no-lto"); - } } } diff --git a/clang/test/Driver/hip-toolchain-no-rdc.hip b/clang/test/Driver/hip-toolchain-no-rdc.hip index 6f148ce01ff30..465722121fc67 100644 --- a/clang/test/Driver/hip-toolchain-no-rdc.hip +++ b/clang/test/Driver/hip-toolchain-no-rdc.hip @@ -46,21 +46,6 @@ // RUN: %s -nogpuinc -nogpulib \ // RUN: 2>&1 | FileCheck -check-prefixes=AMDGCNSPIRV-NEW %s -// SPIR-V requires the LTO pipeline; --no-lto must not be passed. -// RUN: %clang -### --target=x86_64-linux-gnu -fno-gpu-rdc \ -// RUN: --offload-arch=amdgcnspirv -nogpuinc -nogpulib \ -// RUN: %s 2>&1 | FileCheck -check-prefixes=SPIRV-NO-LTO %s -// SPIRV-NO-LTO: {{".*clang-linker-wrapper}} -// SPIRV-NO-LTO-NOT: "--no-lto" -// SPIRV-NO-LTO-SAME: "--emit-fatbin-only" - -// Profile generation needs LTO, so the linker wrapper must not get --no-lto. -// RUN: %clang -### --target=x86_64-linux-gnu -fno-gpu-rdc -fprofile-generate \ -// RUN: -x hip --cuda-gpu-arch=gfx900 -nogpuinc -nogpulib --offload-new-driver \ -// RUN: %s 2>&1 | FileCheck -check-prefixes=PGO -implicit-check-not="--no-lto" %s -// PGO: {{".*clang-linker-wrapper}} -// PGO-SAME: "--emit-fatbin-only" - // // Compile device code in a.cu to code object for gfx803. // @@ -123,7 +108,6 @@ // NEW: [[WRAPPER:".*clang-linker-wrapper]]" // NEW-SAME: "--device-compiler=amdgcn-amd-amdhsa=-flto=full" -// NEW-SAME: "--no-lto" // NEW-SAME: "--host-triple=x86_64-unknown-linux-gnu" // NEW-SAME: "--emit-fatbin-only" // NEW-SAME: "-o" "[[HIPFB_A:.*.hipfb]]" "[[PACKAGE_A]]" @@ -198,7 +182,6 @@ // NEW: [[WRAPPER:".*clang-linker-wrapper]]" // NEW-SAME: "--device-compiler=amdgcn-amd-amdhsa=-flto=full" -// NEW-SAME: "--no-lto" // NEW-SAME: "--emit-fatbin-only" // NEW-SAME: "-o" "[[HIPFB_B:.*.hipfb]]" "[[PACKAGE_B]]" diff --git a/clang/test/OffloadTools/clang-linker-wrapper/linker-wrapper-hip-no-rdc.c b/clang/test/OffloadTools/clang-linker-wrapper/linker-wrapper-hip-no-rdc.c index 80ac493825aad..aae330bd3f6de 100644 --- a/clang/test/OffloadTools/clang-linker-wrapper/linker-wrapper-hip-no-rdc.c +++ b/clang/test/OffloadTools/clang-linker-wrapper/linker-wrapper-hip-no-rdc.c @@ -55,15 +55,3 @@ __attribute__((visibility("protected"), used)) int x; // RUN: test -s %t.gfx9-4-generic-xnack+.co // RUN: test -f %t.gfx1200.co // RUN: test -s %t.gfx1200.co - -// Without --no-lto the AMDGPU device compilation uses the LTO pipeline -// (-flto). -// RUN: clang-linker-wrapper --host-triple=x86_64-unknown-linux-gnu --wrapper-verbose --dry-run --emit-fatbin-only --linker-path=/usr/bin/ld %t.out -o %t.lto.hipfb 2>&1 | FileCheck %s --check-prefix=LTO -// LTO: clang{{.*}} -mcpu=gfx1200 - -// With --no-lto the AMDGPU device compilation uses the conventional non-LTO -// pipeline: -flto must not be passed, and '-x ir' must be passed so Clang -// compiles the bitcode (stored in an object-extension file) instead of -// handing it to the LTO link. -// RUN: clang-linker-wrapper --host-triple=x86_64-unknown-linux-gnu --wrapper-verbose --dry-run --no-lto --emit-fatbin-only --linker-path=/usr/bin/ld %t.out -o %t.nolto.hipfb 2>&1 | FileCheck %s --check-prefix=NO-LTO -// NO-LTO: clang{{.*}} -mcpu=gfx1200{{.*}} -x ir {{.*}}-flto=none diff --git a/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp b/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp index a4a67eed7d47f..76be0f776e561 100644 --- a/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp +++ b/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp @@ -543,16 +543,6 @@ Expected<StringRef> clang(ArrayRef<StringRef> InputFiles, const ArgList &Args, if (!Triple.isNVPTX() && !Triple.isSPIRV()) CmdArgs.push_back("-Wl,--no-undefined"); - // The device inputs are bitcode stored in files with an object extension. - // Force the IR input language so Clang runs the compile and backend phases - // instead of treating them as linker inputs, which would defer codegen to - // the LTO link and defeat the non-LTO pipeline. - // FIXME: This is a stop-gap for non-RDC. Longer term, RDC and non-RDC should - // share a unified interface. - // SPIR-V has no non-LTO pipeline so a --no-lto leaked from a concrete arch in - // a multi-target compile is ignored. Which is a workaround to remove. - if (Args.hasArg(OPT_no_lto) && !Triple.isSPIRV()) - CmdArgs.append({"-x", "ir"}); for (StringRef InputFile : InputFiles) CmdArgs.push_back(InputFile); @@ -615,9 +605,6 @@ Expected<StringRef> clang(ArrayRef<StringRef> InputFiles, const ArgList &Args, for (StringRef Arg : Args.getAllArgValues(OPT_compiler_arg_EQ)) CmdArgs.push_back(Args.MakeArgString(Arg)); - if (Args.hasArg(OPT_no_lto) && !Triple.isSPIRV()) - CmdArgs.append({"-flto=none", "-Wno-unused-command-line-argument"}); - if (Error Err = executeCommands(*ClangPath, CmdArgs)) return std::move(Err); diff --git a/clang/tools/clang-linker-wrapper/LinkerWrapperOpts.td b/clang/tools/clang-linker-wrapper/LinkerWrapperOpts.td index 87a26ca90a66f..53b6c596de291 100644 --- a/clang/tools/clang-linker-wrapper/LinkerWrapperOpts.td +++ b/clang/tools/clang-linker-wrapper/LinkerWrapperOpts.td @@ -39,9 +39,6 @@ def print_wrapped_module : Flag<["--"], "print-wrapped-module">, HelpText<"Print the wrapped module's IR for testing">; def save_temps : Flag<["--"], "save-temps">, Flags<[WrapperOnlyOption]>, HelpText<"Save intermediate results">; -def no_lto : Flag<["--"], "no-lto">, - Flags<[WrapperOnlyOption]>, - HelpText<"Use the non-LTO device compilation pipeline">; def compress : Flag<["--"], "compress">, Flags<[WrapperOnlyOption]>, HelpText<"Compress bundled files">; def compression_level_eq : Joined<["--"], "compression-level=">, _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
