[flang] Add AMDGPU target in flang

[jsjodin]

Summary

This is the first patch of several that will enable generating code for AMD GPUs. It adds the AMDGPU target so it can be used with the --target and -mcpu options.

[kiranchandramohan]

It will be helpful if you can share a godbolt link with the types that you see for complex argument and return type like in https://godbolt.org/z/x73xT3r83. I see some difference from what you have given here. Not sure whether it is because of an incorrect invocation.

Please add tests as well (refer to previous commits https://reviews.llvm.org/D136547).

[tschuett]

I would have expected something ala:

flang -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda example.f90

https://openmp.org/wp-content/uploads/SC18-BoothTalks-Ozen.pdf

gfx902 is not a cpu.

[jsjodin]

In D143102#4097753, @kiranchandramohan wrote:

It will be helpful if you can share a godbolt link with the types that you see for complex argument and return type like in https://godbolt.org/z/x73xT3r83. I see some difference from what you have given here. Not sure whether it is because of an incorrect invocation.

Please add tests as well (refer to previous commits https://reviews.llvm.org/D136547).

Yes, the code for complex is probably premature. It might be better to remove it and have stubs with TODO("adds support for complex ") instead. We cannot generate code for amdgcn right now because there are a lot of missing pieces (function attributes, metadata, address spaces) and without creating huge patch it is better to create small ones and incrementally build up the capability until we have something that works. Complex is pretty far down the list of priorities right now.

[jsjodin]

In D143102#4097773, @tschuett wrote:

I would have expected something ala:

flang -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda example.f90

https://openmp.org/wp-content/uploads/SC18-BoothTalks-Ozen.pdf

gfx902 is not a cpu.

This is not for OpenMP, this is just to allow compilation of a file using the AMDGPU target. You can try the same command in clang:
clang --target=amdgcn-amd-amdhsa -mcpu=gfx902 -c hello.c -###

"/home/jsjodin/git/trunk/llvm-project/build/bin/clang-17" "-cc1" "-triple" "amdgcn-amd-amdhsa" "-emit-obj" "-mrelax-all" "-disable-free" "-clear-ast-before-backend" "-main-file-name" "hello.c" "-mrelocation-model" "pic" "-pic-level" "2" "-fhalf-no-semantic-interposition" "-mframe-pointer=all" "-ffp-contract=on" "-fno-rounding-math" "-mconstructor-aliases" "-fvisibility=hidden" "-fapply-global-visibility-to-externs" "-mlink-builtin-bitcode" "/opt/rocm/amdgcn/bitcode/opencl.bc" "-mlink-builtin-bitcode" "/opt/rocm/amdgcn/bitcode/ocml.bc" "-mlink-builtin-bitcode" "/opt/rocm/amdgcn/bitcode/ockl.bc" "-mlink-builtin-bitcode" "/opt/rocm/amdgcn/bitcode/oclc_daz_opt_off.bc" "-mlink-builtin-bitcode" "/opt/rocm/amdgcn/bitcode/oclc_unsafe_math_off.bc" "-mlink-builtin-bitcode" "/opt/rocm/amdgcn/bitcode/oclc_finite_only_off.bc" "-mlink-builtin-bitcode" "/opt/rocm/amdgcn/bitcode/oclc_correctly_rounded_sqrt_off.bc" "-mlink-builtin-bitcode" "/opt/rocm/amdgcn/bitcode/oclc_wavefrontsize64_on.bc" "-mlink-builtin-bitcode" "/opt/rocm/amdgcn/bitcode/oclc_isa_version_902.bc" "-mlink-builtin-bitcode" "/opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" "-target-cpu" "gfx902" "-mllvm" "-treat-scalable-fixed-error-as-warning" "-debugger-tuning=gdb" "-resource-dir" "/home/jsjodin/git/trunk/llvm-project/build/lib/clang/17" "-fdebug-compilation-dir=/home/jsjodin/git/trunk/llvm-project/flang" "-ferror-limit" "19" "-fgnuc-version=4.2.1" "-fcolor-diagnostics" "-faddrsig" "-o" "hello.o" "-x" "c" "/home/jsjodin/test/fortran/hello.c"

As you can see there are lots more options being generated here, but the -triple and -target-cpu are there.

[jsjodin]

Remove complex arg/return type code.

[kiranchandramohan]

Would it be possible to add a test for the defaultWidth? Like in flang/test/Fir/target-rewrite-boxchar.fir?

[arsenm]

flang/lib/Optimizer/CodeGen/Target.cpp

476

default width of what?

[jsjodin]

In D143102#4103090, @kiranchandramohan wrote:

Would it be possible to add a test for the defaultWidth? Like in flang/test/Fir/target-rewrite-boxchar.fir?

Yes, I will update the patch with the added test.

[jsjodin]

flang/lib/Optimizer/CodeGen/Target.cpp

476

It is the default width of the index type for a string. I assumed that the offsets were 64-bits by default.

[jsjodin]

Add index width test.

[kiranchandramohan]

LGTM. Please wait for a day in case there are comments from others.

[arsenm]

flang/lib/Optimizer/CodeGen/Target.cpp

476

Could use a better name or at least documentation

[jsjodin]

flang/lib/Optimizer/CodeGen/Target.cpp

476

Yes, I agree that the name is not very descriptive but it is already part of the code and this is just an override. I can add a comment.

flang/lib/Optimizer/CodeGen/Target.cpp

complexReturnType(mlir::Location loc, mlir::Type eleTy) const override {
     } else {
       TODO(loc, "complex for this precision");
     }
     return marshal;
   }
 };
 } // namespace
 
+//===----------------------------------------------------------------------===//
+// AMDGPU linux target specifics.
+//===----------------------------------------------------------------------===//
+
+namespace {
+struct TargetAMDGPU : public GenericTarget<TargetAMDGPU> {
+  using GenericTarget::GenericTarget;
+
+  // Default size (in bits) of the index type for strings.
+  static constexpr int defaultWidth = 64;
+
+  CodeGenSpecifics::Marshalling
+  complexArgumentType(mlir::Location loc, mlir::Type eleTy) const override {
+    CodeGenSpecifics::Marshalling marshal;
+    TODO(loc, "handle complex argument types");
+    return marshal;
+  }
+
+  CodeGenSpecifics::Marshalling
+  complexReturnType(mlir::Location loc, mlir::Type eleTy) const override {
+    CodeGenSpecifics::Marshalling marshal;
+    TODO(loc, "handle complex return types");
+    return marshal;
+  }
+};
+} // namespace
+
 // Instantiate the overloaded target instance based on the triple value.
 // TODO: Add other targets to this file as needed.
 std::unique_ptr<fir::CodeGenSpecifics>
 fir::CodeGenSpecifics::get(mlir::MLIRContext *ctx, llvm::Triple &&trp,
                            KindMapping &&kindMap) {
   switch (trp.getArch()) {
   default:
     break;
case llvm::Triple::ArchType::sparc:
     return std::make_unique<TargetSparc>(ctx, std::move(trp),
                                          std::move(kindMap));
   case llvm::Triple::ArchType::sparcv9:
     return std::make_unique<TargetSparcV9>(ctx, std::move(trp),
                                            std::move(kindMap));
   case llvm::Triple::ArchType::riscv64:
     return std::make_unique<TargetRISCV64>(ctx, std::move(trp),
                                            std::move(kindMap));
+  case llvm::Triple::ArchType::amdgcn:
+    return std::make_unique<TargetAMDGPU>(ctx, std::move(trp),
+                                          std::move(kindMap));
   }
   TODO(mlir::UnknownLoc::get(ctx), "target not implemented");
 }

flang/test/Driver/target-gpu-features.f90

+! REQUIRES: amdgpu-registered-target
+
+! Test that -mcpu are used and that the -target-cpu and -target-features
+! are also added to the fc1 command.
+
+! RUN: %flang --target=amdgcn-amd-amdhsa -mcpu=gfx902 -c %s -### 2>&1 \
+! RUN: | FileCheck %s -check-prefix=CHECK-AMDGCN
+
+! CHECK-AMDGCN: "-fc1" "-triple" "amdgcn-amd-amdhsa"
+! CHECK-AMDGCN-SAME: "-target-cpu" "gfx902"

flang/test/Fir/target-rewrite-boxchar.fir

 // RUN: fir-opt --target-rewrite="target=i386-unknown-linux-gnu" %s | FileCheck %s --check-prefix=INT32
 // RUN: fir-opt --target-rewrite="target=x86_64-unknown-linux-gnu" %s | FileCheck %s --check-prefix=INT64
 // RUN: fir-opt --target-rewrite="target=aarch64-unknown-linux-gnu" %s | FileCheck %s --check-prefix=INT64
 // RUN: fir-opt --target-rewrite="target=powerpc64le-unknown-linux-gnu" %s | FileCheck %s --check-prefix=INT64
+// RUN: fir-opt --target-rewrite="target=amdgcn-amd-amdhsa" %s | FileCheck %s --check-prefix=INT64
 
 // Test that we rewrite the signatures and bodies of functions that take boxchar
 // parameters.
 // INT32-LABEL: @boxcharparams
 // INT32-SAME: ([[ARG0:%[0-9A-Za-z]+]]: !fir.ref<!fir.char<1,?>>, [[ARG1:%[0-9A-Za-z]+]]: !fir.ref<!fir.char<1,?>>, [[ARG2:%[0-9A-Za-z]+]]: i32, [[ARG3:%[0-9A-Za-z]+]]: i32) -> i64
 // INT64-LABEL: @boxcharparams
 // INT64-SAME: ([[ARG0:%[0-9A-Za-z]+]]: !fir.ref<!fir.char<1,?>>, [[ARG1:%[0-9A-Za-z]+]]: !fir.ref<!fir.char<1,?>>, [[ARG2:%[0-9A-Za-z]+]]: i64, [[ARG3:%[0-9A-Za-z]+]]: i64) -> i64
 func.func @boxcharparams(%arg0 : !fir.boxchar<1>, %arg1 : !fir.boxchar<1>) -> i64 {