[clang] [HIP][AMDGPU] Use non-LTO pipeline for non-RDC in the linker wrapper (PR #201135)
Yaxun Liu via cfe-commits
cfe-commits at lists.llvm.org
Tue Jun 2 10:50:12 PDT 2026
https://github.com/yxsamliu updated https://github.com/llvm/llvm-project/pull/201135
>From a8daecd6b38bed93bd6449643e85286d51ceef19 Mon Sep 17 00:00:00 2001
From: "Yaxun (Sam) Liu" <yaxun.liu at amd.com>
Date: Tue, 2 Jun 2026 10:12:23 -0400
Subject: [PATCH] [HIP][AMDGPU] Use non-LTO pipeline for non-RDC in the linker
wrapper
Non-RDC HIP does not need LTO, but the new offload driver compiles all
AMDGPU device code through the LTO pipeline. That makes non-RDC builds pay
full LTO codegen cost for no benefit.
Fix this in clang-linker-wrapper instead of the driver, so device codegen
still runs in the wrapper's parallel device-link step (kept fast by
--offload-jobs) rather than being serialized back in the driver.
The driver passes a new --no-lto flag for the non-RDC fat-binary job
(unless the user asked for -foffload-lto). With it, the wrapper drops -flto
and, because the device images are bitcode stored in object-extension
files, also passes -x ir so clang actually compiles them (cc1 -emit-obj)
instead of handing the bitcode to lld for an LTO link.
---
clang/lib/Driver/ToolChains/Clang.cpp | 11 +++++++++++
clang/test/Driver/hip-toolchain-no-rdc.hip | 9 +++++++++
clang/test/Driver/hipspv-toolchain.hip | 2 +-
.../linker-wrapper-hip-no-rdc.c | 12 ++++++++++++
.../clang-linker-wrapper/ClangLinkerWrapper.cpp | 16 ++++++++++++++--
.../clang-linker-wrapper/LinkerWrapperOpts.td | 3 +++
6 files changed, 50 insertions(+), 3 deletions(-)
diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp
index 7657afb14f077..dcd262a084753 100644
--- a/clang/lib/Driver/ToolChains/Clang.cpp
+++ b/clang/lib/Driver/ToolChains/Clang.cpp
@@ -9830,6 +9830,17 @@ void LinkerWrapper::ConstructJob(Compilation &C, const JobAction &JA,
JA.getType() == types::TY_Image);
if (JA.getType() == types::TY_HIP_FATBIN) {
CmdArgs.push_back("--emit-fatbin-only");
+ // Non-RDC HIP uses the conventional non-LTO pipeline unless the user opts
+ // into offload LTO. The device backend then runs in the linker wrapper's
+ // parallel device-link step rather than being deferred to the LTO link.
+ // Profile generation still needs LTO so the device profile runtime is
+ // linked and optimized together with the device code.
+ 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 (C.getDriver().getOffloadLTOMode() == LTOK_None && !UsesProfileGenerate)
+ CmdArgs.push_back("--no-lto");
CmdArgs.append({"-o", Output.getFilename()});
for (auto Input : Inputs)
CmdArgs.push_back(Input.getFilename());
diff --git a/clang/test/Driver/hip-toolchain-no-rdc.hip b/clang/test/Driver/hip-toolchain-no-rdc.hip
index f4cd703547ac0..d018cfe48b72e 100644
--- a/clang/test/Driver/hip-toolchain-no-rdc.hip
+++ b/clang/test/Driver/hip-toolchain-no-rdc.hip
@@ -46,6 +46,13 @@
// RUN: %s -nogpuinc -nogpulib \
// RUN: 2>&1 | FileCheck -check-prefixes=AMDGCNSPIRV-NEW %s
+// 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.
//
@@ -109,6 +116,7 @@
// NEW: [[WRAPPER:".*clang-linker-wrapper]]"
// NEW-SAME: "--host-triple=x86_64-unknown-linux-gnu"
// NEW-SAME: "--emit-fatbin-only"
+// NEW-SAME: "--no-lto"
// NEW-SAME: "-o" "[[HIPFB_A:.*.hipfb]]" "[[PACKAGE_A]]"
// CHECK: [[CLANG]] "-cc1" "-triple" "x86_64-unknown-linux-gnu"
@@ -181,6 +189,7 @@
// NEW: [[WRAPPER:".*clang-linker-wrapper]]"
// NEW-SAME: "--emit-fatbin-only"
+// NEW-SAME: "--no-lto"
// NEW-SAME: "-o" "[[HIPFB_B:.*.hipfb]]" "[[PACKAGE_B]]"
// CHECK: [[CLANG]] "-cc1" "-triple" "x86_64-unknown-linux-gnu"
diff --git a/clang/test/Driver/hipspv-toolchain.hip b/clang/test/Driver/hipspv-toolchain.hip
index d2a7e9a3aeb3a..8eeb6e60cba50 100644
--- a/clang/test/Driver/hipspv-toolchain.hip
+++ b/clang/test/Driver/hipspv-toolchain.hip
@@ -38,7 +38,7 @@
// NEW-SAME: "--image=file=[[OBJ_DEV]],triple=[[TRIPLE]],arch=generic,kind=hip"
// NEW: {{".*clang-linker-wrapper"}} "--device-compiler=[[TRIPLE]]=--hip-path=[[HIP_PATH]]"
-// NEW-SAME: "--emit-fatbin-only" "-o" "[[BUNDLE:.*hipfb]]"
+// NEW-SAME: "--emit-fatbin-only" "--no-lto" "-o" "[[BUNDLE:.*hipfb]]"
// CHECK: [[CLANG]] "-cc1" "-triple" {{".*"}} "-aux-triple" "[[TRIPLE]]"
// CHECK-SAME: "-emit-obj"
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 aae330bd3f6de..5c5b7b1eabfab 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
@@ -55,3 +55,15 @@ __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{{.*}} -flto
+
+// 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 --implicit-check-not=-flto
+// NO-LTO: clang{{.*}} -mcpu=gfx1200{{.*}} -x ir
diff --git a/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp b/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp
index 24900a43dbcc2..c35aceff6aa43 100644
--- a/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp
+++ b/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp
@@ -535,8 +535,14 @@ Expected<StringRef> clang(ArrayRef<StringRef> InputFiles, const ArgList &Args,
Triple.isAMDGPU() ? CmdArgs.push_back(Args.MakeArgString("-mcpu=" + Arch))
: CmdArgs.push_back(Args.MakeArgString("-march=" + Arch));
- // AMDGPU is always in LTO mode currently.
- if (Triple.isAMDGPU())
+ // AMDGPU defaults to the LTO pipeline. Non-RDC HIP uses the conventional
+ // non-LTO pipeline so device codegen still runs here, in parallel, instead
+ // of being deferred to the LTO link.
+ // FIXME: This is a stop-gap for non-RDC. Longer term, RDC and non-RDC should
+ // share a unified interface so runtime libraries can be provided to non-RDC
+ // compilations without relying on -mlink-builtin-bitcode.
+ bool NonLTOAMDGPU = Triple.isAMDGPU() && Args.hasArg(OPT_no_lto);
+ if (Triple.isAMDGPU() && !NonLTOAMDGPU)
CmdArgs.push_back("-flto");
// Forward all of the `--offload-opt` and `-mllvm` options to the device.
@@ -548,6 +554,12 @@ 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.
+ if (NonLTOAMDGPU)
+ CmdArgs.append({"-x", "ir"});
for (StringRef InputFile : InputFiles)
CmdArgs.push_back(InputFile);
diff --git a/clang/tools/clang-linker-wrapper/LinkerWrapperOpts.td b/clang/tools/clang-linker-wrapper/LinkerWrapperOpts.td
index 53b6c596de291..87a26ca90a66f 100644
--- a/clang/tools/clang-linker-wrapper/LinkerWrapperOpts.td
+++ b/clang/tools/clang-linker-wrapper/LinkerWrapperOpts.td
@@ -39,6 +39,9 @@ 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=">,
More information about the cfe-commits
mailing list