(cherry-pick) [Clang] Make --lto-partitions only default for HIP (llvm#133164)

Summary:
The default behavior for LTO on other targets does not specify the
number of LTO partitions. Recent changes made this default to 8 on
AMDGPU which had some issues with the `libc` project. The option to
disable this is HIP only so I think for now we should restrict this just
to HIP.

I'm definitely on board with getting some more parallelism here, but I
think it should probably be restricted to just offloading languages. The
new driver goes through the `--target=amdgcn-amd-amdhsa` for its output,
which means we'd need to forward the default somehow.

Author: jhuber6

clang/lib/Driver/ToolChains/AMDGPU.cpp

@@ -632,7 +632,7 @@ void amdgpu::Linker::ConstructJob(Compilation &C, const JobAction &JA,
     const bool ThinLTO = (C.getDriver().getLTOMode() == LTOK_Thin);
     addLTOOptions(getToolChain(), Args, CmdArgs, Output, Inputs[0], ThinLTO);
 
-    if (!ThinLTO)
+    if (!ThinLTO && JA.getOffloadingDeviceKind() == Action::OFK_HIP)
       addFullLTOPartitionOption(C.getDriver(), Args, CmdArgs);
   } else if (Args.hasArg(options::OPT_mcpu_EQ))
     CmdArgs.push_back(Args.MakeArgString(
@@ -645,14 +645,12 @@ void amdgpu::Linker::ConstructJob(Compilation &C, const JobAction &JA,
 }
 
 static unsigned getFullLTOPartitions(const Driver &D, const ArgList &Args) {
-  const Arg *A = Args.getLastArg(options::OPT_flto_partitions_EQ);
-  // In the absence of an option, use 8 as the default.
-  if (!A)
-    return 8;
   int Value = 0;
-  if (StringRef(A->getValue()).getAsInteger(10, Value) || (Value < 1)) {
+  StringRef A = Args.getLastArgValue(options::OPT_flto_partitions_EQ, "8");
+  if (A.getAsInteger(10, Value) || (Value < 1)) {
+    Arg *Arg = Args.getLastArg(options::OPT_flto_partitions_EQ);
     D.Diag(diag::err_drv_invalid_int_value)
-        << A->getAsString(Args) << A->getValue();
+        << Arg->getAsString(Args) << Arg->getValue();
     return 1;
   }
 
@@ -662,13 +660,8 @@ static unsigned getFullLTOPartitions(const Driver &D, const ArgList &Args) {
 void amdgpu::addFullLTOPartitionOption(const Driver &D,
                                        const llvm::opt::ArgList &Args,
                                        llvm::opt::ArgStringList &CmdArgs) {
-  // TODO: Should this be restricted to fgpu-rdc only ? Currently we'll
-  //       also do it for non gpu-rdc LTO
-
-  if (unsigned NumParts = getFullLTOPartitions(D, Args); NumParts > 1) {
-    CmdArgs.push_back(
-        Args.MakeArgString("--lto-partitions=" + Twine(NumParts)));
-  }
+  CmdArgs.push_back(Args.MakeArgString("--lto-partitions=" +
+                                       Twine(getFullLTOPartitions(D, Args))));
 }
 
 void amdgpu::getAMDGPUTargetFeatures(const Driver &D,

clang/test/Driver/amdgpu-toolchain.c

@@ -21,25 +21,12 @@
 // RUN: %clang -### --target=amdgcn-amd-amdhsa -mcpu=gfx906 -nogpulib \
 // RUN:   -L. -flto -fconvergent-functions %s 2>&1 | FileCheck -check-prefix=LTO %s
 // LTO: clang{{.*}} "-flto=full"{{.*}}"-fconvergent-functions"
-// LTO: ld.lld{{.*}}"-L."{{.*}}"-plugin-opt=mcpu=gfx906"{{.*}}"--lto-partitions={{[0-9]+}}"
+// LTO: ld.lld{{.*}}"-L."{{.*}}"-plugin-opt=mcpu=gfx906"{{.*}}"
+
 // RUN: %clang -### --target=amdgcn-amd-amdhsa -mcpu=gfx906 -nogpulib \
 // RUN:   -L. -fconvergent-functions %s 2>&1 | FileCheck -check-prefix=MCPU %s
 // MCPU: ld.lld{{.*}}"-L."{{.*}}"-plugin-opt=mcpu=gfx906"
 
 // RUN: %clang -### --target=amdgcn-amd-amdhsa -mcpu=gfx906 -nogpulib \
 // RUN:   -fuse-ld=ld %s 2>&1 | FileCheck -check-prefixes=LD %s
 // LD: ld.lld
-
-// Check --flto-partitions
-
-// RUN: %clang -### --target=amdgcn-amd-amdhsa -mcpu=gfx90a -nogpulib \
-// RUN:   -L. -flto --flto-partitions=42 %s 2>&1 | FileCheck -check-prefix=LTO_PARTS %s
-// LTO_PARTS: ld.lld{{.*}}"-L."{{.*}}"-plugin-opt=mcpu=gfx90a"{{.*}}"--lto-partitions=42"
-
-// RUN: not %clang -### --target=amdgcn-amd-amdhsa -mcpu=gfx90a -nogpulib \
-// RUN:   -L. -flto --flto-partitions=a %s 2>&1 | FileCheck -check-prefix=LTO_PARTS_INV0 %s
-// LTO_PARTS_INV0: clang: error: invalid integral value 'a' in '--flto-partitions=a'
-
-// RUN: not %clang -### --target=amdgcn-amd-amdhsa -mcpu=gfx90a -nogpulib \
-// RUN:   -L. -flto --flto-partitions=0 %s 2>&1 | FileCheck -check-prefix=LTO_PARTS_INV1 %s
-// LTO_PARTS_INV1: clang: error: invalid integral value '0' in '--flto-partitions=0'

clang/test/Driver/hip-toolchain-rdc.hip

@@ -161,3 +161,21 @@
 // output the executable
 // LNX: [[LD:".*ld.*"]] {{.*}}"-o" "a.out" {{.*}} [[A_OBJ_HOST]] [[B_OBJ_HOST]] [[OBJBUNDLE]]
 // MSVC: [[LD:".*lld-link.*"]] {{.*}}"-out:a.exe" {{.*}} [[A_OBJ_HOST]] [[B_OBJ_HOST]] [[OBJBUNDLE]]
+
+// Check --flto-partitions
+
+// RUN: %clang -### -fgpu-rdc --offload-arch=gfx90a -nogpulib -nogpuinc \
+// RUN:   -L. -foffload-lto %s 2>&1 | FileCheck -check-prefix=LTO_DEFAULT %s
+// LTO_DEFAULT: lld{{.*}}"--lto-partitions=8"
+
+// RUN: %clang -### -fgpu-rdc --offload-arch=gfx90a -nogpulib -nogpuinc \
+// RUN:   -L. -foffload-lto --flto-partitions=42 %s 2>&1 | FileCheck -check-prefix=LTO_PARTS %s
+// LTO_PARTS: lld{{.*}}"--lto-partitions=42"
+
+// RUN: not %clang -### -fgpu-rdc --offload-arch=gfx90a -nogpulib -nogpuinc \
+// RUN:   -L. -foffload-lto --flto-partitions=a %s 2>&1 | FileCheck -check-prefix=LTO_PARTS_INV0 %s
+// LTO_PARTS_INV0: clang: error: invalid integral value 'a' in '--flto-partitions=a'
+
+// RUN: not %clang -### -fgpu-rdc --offload-arch=gfx90a -nogpulib -nogpuinc \
+// RUN:   -L. -foffload-lto --flto-partitions=0 %s 2>&1 | FileCheck -check-prefix=LTO_PARTS_INV1 %s
+// LTO_PARTS_INV1: clang: error: invalid integral value '0' in '--flto-partitions=0'