AMDGPU: Enable _Float16

[b-sumner]

Looks good to me.

[rampitec]

LGTM

[foad]

Should have updated https://clang.llvm.org/docs/LanguageExtensions.html#half-precision-floating-point "_Float16 is currently only supported on the following targets, with further targets pending ABI standardization: ..."

[arsenm]

test/CodeGenCXX/amdgpu-float16.cpp

7

I thought clang tests were supposed to stop at the IR

[yaxunl]

In D61112#2584279, @foad wrote:

Should have updated https://clang.llvm.org/docs/LanguageExtensions.html#half-precision-floating-point "_Float16 is currently only supported on the following targets, with further targets pending ABI standardization: ..."

will update the doc

test/CodeGenCXX/amdgpu-float16.cpp

7

If you 'grep -- '-S ' clang/test/CodeGen/* | grep -v emit-llvm | grep clang', there are plenty of clang codegen tests checks assembly code. In this test we want to make sure f16 instructions are emitted for specific GPUs.

lib/Basic/Targets/AMDGPU.cpp

AMDGPUTargetInfo::AMDGPUTargetInfo(const llvm::Triple &Triple,
   resetDataLayout(isAMDGCN(getTriple()) ? DataLayoutStringAMDGCN
                                         : DataLayoutStringR600);
   assert(DataLayout->getAllocaAddrSpace() == Private);
 
   setAddressSpaceMap(Triple.getOS() == llvm::Triple::Mesa3D ||
                      !isAMDGCN(Triple));
   UseAddrSpaceMapMangling = true;
 
+  HasLegalHalfType = true;
+  HasFloat16 = true;
+
   // Set pointer width and alignment for target address space 0.
   PointerWidth = PointerAlign = DataLayout->getPointerSizeInBits();
   if (getMaxPointerWidth() == 64) {
     LongWidth = LongAlign = 64;
     SizeType = UnsignedLong;
     PtrDiffType = SignedLong;
     IntPtrType = SignedLong;
   }

test/CodeGenCXX/amdgpu-float16.cpp

+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -target-cpu gfx701 -S -o - %s | FileCheck %s -check-prefix=NOF16
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -target-cpu gfx803 -S -o - %s | FileCheck %s
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -target-cpu gfx900 -S -o - %s | FileCheck %s
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -target-cpu gfx906 -S -o - %s | FileCheck %s
+void f() {
+  _Float16 x, y, z;
+  // CHECK: v_add_f16_e64
+  // NOF16: v_add_f32_e64
+  z = x + y;
+  // CHECK: v_sub_f16_e64
+  // NOF16: v_sub_f32_e64
+  z = x - y;
+  // CHECK: v_mul_f16_e64
+  // NOF16: v_mul_f32_e64
+  z = x * y;
+  // CHECK: v_div_fixup_f16
+  // NOF16: v_div_fixup_f32
+  z = x / y;
+}