[clang] [AMDGPU][Clang] Support bfloat16 arithmetic for GFX950 and GFX1250. (PR #147541)

[Ivan Kosarev via cfe-commits]

https://github.com/kosarev updated https://github.com/llvm/llvm-project/pull/147541

>From 33e0843a1b66383e88f8bee4d51ce382209eb4a0 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nicolai=20H=C3=A4hnle?= <nicolai.haehnle at amd.com>
Date: Mon, 7 Jul 2025 11:34:03 +0100
Subject: [PATCH 1/2] [AMDGPU][Clang] Support bfloat16 arithmetic.

Co-authored-by: Ivan Kosarev <ivan.kosarev at amd.com>
---
 clang/lib/Basic/Targets/AMDGPU.h      |  2 ++
 clang/test/CodeGen/AMDGPU/full-bf16.c | 23 +++++++++++++++++++++++
 2 files changed, 25 insertions(+)
 create mode 100644 clang/test/CodeGen/AMDGPU/full-bf16.c

diff --git a/clang/lib/Basic/Targets/AMDGPU.h b/clang/lib/Basic/Targets/AMDGPU.h
index 509128f3cf070..006c2fe475a94 100644
--- a/clang/lib/Basic/Targets/AMDGPU.h
+++ b/clang/lib/Basic/Targets/AMDGPU.h
@@ -451,6 +451,8 @@ class LLVM_LIBRARY_VISIBILITY AMDGPUTargetInfo final : public TargetInfo {
         CUMode = false;
       else if (F == "+image-insts")
         HasImage = true;
+      else if (F == "+gfx950-insts" || F == "+gfx1250-insts")
+        HasFullBFloat16 = true;
       bool IsOn = F.front() == '+';
       StringRef Name = StringRef(F).drop_front();
       if (!llvm::is_contained(TargetIDFeatures, Name))
diff --git a/clang/test/CodeGen/AMDGPU/full-bf16.c b/clang/test/CodeGen/AMDGPU/full-bf16.c
new file mode 100644
index 0000000000000..7770d02c59df1
--- /dev/null
+++ b/clang/test/CodeGen/AMDGPU/full-bf16.c
@@ -0,0 +1,23 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa-gnu -target-cpu gfx950 -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa-gnu -target-cpu gfx1250 -emit-llvm -o - %s | FileCheck %s
+
+// CHECK-LABEL: define dso_local bfloat @div(
+// CHECK-SAME: bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    [[RETVAL:%.*]] = alloca bfloat, align 2, addrspace(5)
+// CHECK-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2, addrspace(5)
+// CHECK-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2, addrspace(5)
+// CHECK-NEXT:    [[RETVAL_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[RETVAL]] to ptr
+// CHECK-NEXT:    [[A_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[A_ADDR]] to ptr
+// CHECK-NEXT:    [[B_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[B_ADDR]] to ptr
+// CHECK-NEXT:    store bfloat [[A]], ptr [[A_ADDR_ASCAST]], align 2
+// CHECK-NEXT:    store bfloat [[B]], ptr [[B_ADDR_ASCAST]], align 2
+// CHECK-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR_ASCAST]], align 2
+// CHECK-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR_ASCAST]], align 2
+// CHECK-NEXT:    [[DIV:%.*]] = fdiv bfloat [[TMP0]], [[TMP1]]
+// CHECK-NEXT:    ret bfloat [[DIV]]
+//
+__bf16 div(__bf16 a, __bf16 b) {
+  return a / b;
+}

>From 4fedbbd3ea544d48d332adbf8555e7609b631ecd Mon Sep 17 00:00:00 2001
From: Ivan Kosarev <ivan.kosarev at amd.com>
Date: Tue, 8 Jul 2025 16:25:01 +0100
Subject: [PATCH 2/2] Enable bf16 arithmetic for all subtargets.

---
 clang/lib/Basic/Targets/AMDGPU.h      | 3 +--
 clang/test/CodeGen/AMDGPU/full-bf16.c | 5 +++++
 2 files changed, 6 insertions(+), 2 deletions(-)

diff --git a/clang/lib/Basic/Targets/AMDGPU.h b/clang/lib/Basic/Targets/AMDGPU.h
index 006c2fe475a94..1358abb70f984 100644
--- a/clang/lib/Basic/Targets/AMDGPU.h
+++ b/clang/lib/Basic/Targets/AMDGPU.h
@@ -439,6 +439,7 @@ class LLVM_LIBRARY_VISIBILITY AMDGPUTargetInfo final : public TargetInfo {
   // pre-defined macros.
   bool handleTargetFeatures(std::vector<std::string> &Features,
                             DiagnosticsEngine &Diags) override {
+    HasFullBFloat16 = true;
     auto TargetIDFeatures =
         getAllPossibleTargetIDFeatures(getTriple(), getArchNameAMDGCN(GPUKind));
     for (const auto &F : Features) {
@@ -451,8 +452,6 @@ class LLVM_LIBRARY_VISIBILITY AMDGPUTargetInfo final : public TargetInfo {
         CUMode = false;
       else if (F == "+image-insts")
         HasImage = true;
-      else if (F == "+gfx950-insts" || F == "+gfx1250-insts")
-        HasFullBFloat16 = true;
       bool IsOn = F.front() == '+';
       StringRef Name = StringRef(F).drop_front();
       if (!llvm::is_contained(TargetIDFeatures, Name))
diff --git a/clang/test/CodeGen/AMDGPU/full-bf16.c b/clang/test/CodeGen/AMDGPU/full-bf16.c
index 7770d02c59df1..d2ec34561cd8e 100644
--- a/clang/test/CodeGen/AMDGPU/full-bf16.c
+++ b/clang/test/CodeGen/AMDGPU/full-bf16.c
@@ -1,5 +1,10 @@
 // NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa-gnu -target-cpu tahiti -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa-gnu -target-cpu fiji -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa-gnu -target-cpu gfx900 -emit-llvm -o - %s | FileCheck %s
 // RUN: %clang_cc1 -triple amdgcn-amd-amdhsa-gnu -target-cpu gfx950 -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa-gnu -target-cpu gfx1010 -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa-gnu -target-cpu gfx1100 -emit-llvm -o - %s | FileCheck %s
 // RUN: %clang_cc1 -triple amdgcn-amd-amdhsa-gnu -target-cpu gfx1250 -emit-llvm -o - %s | FileCheck %s
 
 // CHECK-LABEL: define dso_local bfloat @div(