[HIP] Fix lld failure when devie object is empty
authorYaxun (Sam) Liu <yaxun.liu@amd.com>
Thu, 17 Nov 2022 17:04:57 +0000 (12:04 -0500)
committerYaxun (Sam) Liu <yaxun.liu@amd.com>
Tue, 22 Nov 2022 15:38:42 +0000 (10:38 -0500)
When -fgpu-rdc is used for linking relocatable objects, clang driver launches
clang-offload-bundler to extract a device relocatable object from each input
relocatable object file and passes the extracted files to lld. The input relocatable
object file could either come from HIP program or C++ program. The relocatable
object file from C++ program does not contain device relocatable objects, therefore
clang-offload-bundler extracts an empty file and passes it to lld. lld treates
empty file as linker script. When there is no object input file to lld, lld
will emit error:

target emulation unknown: -m or at least one .o file required

This patch adds "elf64_amdgpu" to lld so that lld always know the target
no matter whether there are object input files or not.

Reviewed by: Artem Belevich, Fangrui Song

Differential Revision: https://reviews.llvm.org/D138221

clang/lib/Driver/ToolChains/HIPAMD.cpp
clang/test/Driver/hip-toolchain-device-only.hip
clang/test/Driver/hip-toolchain-no-rdc.hip
lld/ELF/Driver.cpp
lld/test/ELF/emulation-amdgpu.s [new file with mode: 0644]

index 1951ac1..c64421d 100644 (file)
@@ -108,7 +108,12 @@ void AMDGCN::Linker::constructLldCommand(Compilation &C, const JobAction &JA,
                                          const llvm::opt::ArgList &Args) const {
   // Construct lld command.
   // The output from ld.lld is an HSA code object file.
-  ArgStringList LldArgs{"-flavor", "gnu", "--no-undefined", "-shared",
+  ArgStringList LldArgs{"-flavor",
+                        "gnu",
+                        "-m",
+                        "elf64_amdgpu",
+                        "--no-undefined",
+                        "-shared",
                         "-plugin-opt=-amdgpu-internalize-symbols"};
 
   auto &TC = getToolChain();
index 9e4a36f..c6ff67b 100644 (file)
@@ -12,7 +12,7 @@
 // CHECK-SAME: "-target-cpu" "gfx803"
 // CHECK-SAME: {{.*}} "-o" [[OBJ_DEV_A_803:".*o"]] "-x" "hip"
 
-// CHECK: [[LLD: ".*lld.*"]] "-flavor" "gnu" "--no-undefined" "-shared"
+// CHECK: [[LLD: ".*lld.*"]] "-flavor" "gnu" "-m" "elf64_amdgpu" "--no-undefined" "-shared"
 // CHECK-SAME: "-o" "[[IMG_DEV_A_803:.*out]]" [[OBJ_DEV_A_803]]
 
 // CHECK: [[CLANG:".*clang.*"]] "-cc1"{{.*}} "-triple" "amdgcn-amd-amdhsa"
@@ -21,7 +21,7 @@
 // CHECK-SAME: "-target-cpu" "gfx900"
 // CHECK-SAME: {{.*}} "-o" [[OBJ_DEV_A_900:".*o"]] "-x" "hip"
 
-// CHECK: [[LLD]] "-flavor" "gnu" "--no-undefined" "-shared"
+// CHECK: [[LLD]] "-flavor" "gnu" "-m" "elf64_amdgpu" "--no-undefined" "-shared"
 // CHECK-SAME: "-o" "[[IMG_DEV_A_900:.*out]]" [[OBJ_DEV_A_900]]
 
 // CHECK: [[BUNDLER:".*clang-offload-bundler"]] "-type=o"
index 2cd44ca..4ae054b 100644 (file)
@@ -59,7 +59,7 @@
 // CHECK-NOT: {{".*opt"}}
 // CHECK-NOT: {{".*llc"}}
 
-// CHECK: [[LLD: ".*lld.*"]] "-flavor" "gnu" "--no-undefined" "-shared"
+// CHECK: [[LLD: ".*lld.*"]] "-flavor" "gnu" "-m" "elf64_amdgpu" "--no-undefined" "-shared"
 // CHECK-SAME: "-o" "[[IMG_DEV_A_803:.*out]]" [[OBJ_DEV_A_803]]
 
 //
@@ -82,7 +82,7 @@
 // CHECK-NOT: {{".*opt"}}
 // CHECK-NOT: {{".*llc"}}
 
-// CHECK: [[LLD]] "-flavor" "gnu" "--no-undefined" "-shared"
+// CHECK: [[LLD]] "-flavor" "gnu" "-m" "elf64_amdgpu" "--no-undefined" "-shared"
 // CHECK-SAME: "-o" "[[IMG_DEV_A_900:.*out]]" [[OBJ_DEV_A_900]]
 
 //
 // CHECK-NOT: {{".*opt"}}
 // CHECK-NOT: {{".*llc"}}
 
-// CHECK: [[LLD]] "-flavor" "gnu" "--no-undefined" "-shared"
+// CHECK: [[LLD]] "-flavor" "gnu" "-m" "elf64_amdgpu" "--no-undefined" "-shared"
 // CHECK-SAME: "-o" "[[IMG_DEV_B_803:.*out]]" [[OBJ_DEV_B_803]]
 
 //
 // CHECK-NOT: {{".*opt"}}
 // CHECK-NOT: {{".*llc"}}
 
-// CHECK: [[LLD]] "-flavor" "gnu" "--no-undefined" "-shared"
+// CHECK: [[LLD]] "-flavor" "gnu" "-m" "elf64_amdgpu" "--no-undefined" "-shared"
 // CHECK-SAME: "-o" "[[IMG_DEV_B_900:.*out]]" [[OBJ_DEV_B_900]]
 
 //
index f9792ac..b8cc8ef 100644 (file)
@@ -176,12 +176,15 @@ static std::tuple<ELFKind, uint16_t, uint8_t> parseEmulation(StringRef emul) {
           .Case("elf_iamcu", {ELF32LEKind, EM_IAMCU})
           .Case("elf64_sparc", {ELF64BEKind, EM_SPARCV9})
           .Case("msp430elf", {ELF32LEKind, EM_MSP430})
+          .Case("elf64_amdgpu", {ELF64LEKind, EM_AMDGPU})
           .Default({ELFNoneKind, EM_NONE});
 
   if (ret.first == ELFNoneKind)
     error("unknown emulation: " + emul);
   if (ret.second == EM_MSP430)
     osabi = ELFOSABI_STANDALONE;
+  else if (ret.second == EM_AMDGPU)
+    osabi = ELFOSABI_AMDGPU_HSA;
   return std::make_tuple(ret.first, ret.second, osabi);
 }
 
diff --git a/lld/test/ELF/emulation-amdgpu.s b/lld/test/ELF/emulation-amdgpu.s
new file mode 100644 (file)
index 0000000..707f0ae
--- /dev/null
@@ -0,0 +1,36 @@
+# REQUIRES: amdgpu
+
+# RUN: llvm-mc -filetype=obj -triple=amdgcn-amd-amdhsa %s -o %t.o
+# RUN: ld.lld %t.o -o %t
+# RUN: llvm-readobj --file-headers %t | FileCheck %s
+# RUN: ld.lld -m elf64_amdgpu %t.o -o %t
+# RUN: llvm-readobj --file-headers %t | FileCheck %s
+
+# CHECK:      ElfHeader {
+# CHECK-NEXT:   Ident {
+# CHECK-NEXT:     Magic: (7F 45 4C 46)
+# CHECK-NEXT:     Class: 64-bit (0x2)
+# CHECK-NEXT:     DataEncoding: LittleEndian (0x1)
+# CHECK-NEXT:     FileVersion: 1
+# CHECK-NEXT:     OS/ABI: AMDGPU_HSA (0x40)
+# CHECK-NEXT:     ABIVersion: 2
+# CHECK-NEXT:     Unused: (00 00 00 00 00 00 00)
+# CHECK-NEXT:   }
+# CHECK-NEXT:   Type: Executable (0x2)
+# CHECK-NEXT:   Machine: EM_AMDGPU (0xE0)
+# CHECK-NEXT:   Version: 1
+# CHECK-NEXT:   Entry:
+# CHECK-NEXT:   ProgramHeaderOffset: 0x40
+# CHECK-NEXT:   SectionHeaderOffset:
+# CHECK-NEXT:   Flags [ (0x0)
+# CHECK-NEXT:   ]
+# CHECK-NEXT:   HeaderSize: 64
+# CHECK-NEXT:   ProgramHeaderEntrySize: 56
+# CHECK-NEXT:   ProgramHeaderCount:
+# CHECK-NEXT:   SectionHeaderEntrySize: 64
+# CHECK-NEXT:   SectionHeaderCount:
+# CHECK-NEXT:   StringTableSectionIndex:
+# CHECK-NEXT: }
+
+.globl _start
+_start: