Skip to content

[LoongArch][BF16] Add support for the __bf16 type #142548

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 1 commit into from
Jun 9, 2025
Merged

[LoongArch][BF16] Add support for the __bf16 type #142548

merged 1 commit into from
Jun 9, 2025

Conversation

@Ami-zhang
Copy link
Contributor

@Ami-zhang Ami-zhang commented Jun 3, 2025

The LoongArch psABI recently added __bf16 type support. Now we can enable this new type in clang.

Currently, bf16 operations are automatically supported by promoting to float. This patch adds bf16 support by ensuring that load extension / truncate store operations are properly expanded.

And this commit implements support for bf16 truncate/extend on hard FP targets. The extend operation is implemented by a shift just as in the standard legalization. This requires custom lowering of the truncate libcall on hard float ABIs (the normal libcall code path is used on soft ABIs).

@llvmbot llvmbot added clang Clang issues not falling into any other category clang:frontend Language frontend issues, e.g. anything involving "Sema" backend:loongarch labels Jun 3, 2025
@llvmbot
Copy link
Member

llvmbot commented Jun 3, 2025

@llvm/pr-subscribers-backend-loongarch

Author: None (Ami-zhang)

Changes

The LoongArch psABI recently added __bf16 type support. Now we can enable this new type in clang.

Currently, bf16 operations are automatically supported by promoting to float. This patch adds bf16 support by ensuring that load extension / truncate store operations are properly expanded.

And this commit implements support for bf16 truncate/extend on hard FP targets. The extend operation is implemented by a shift just as in the standard legalization. This requires custom lowering of the truncate libcall on hard float ABIs (the normal libcall code path is used on soft ABIs).

Patch is 92.03 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/142548.diff

8 Files Affected:

  • (modified) clang/docs/LanguageExtensions.rst (+1)
  • (modified) clang/lib/Basic/Targets/LoongArch.h (+5)
  • (added) clang/test/CodeGen/LoongArch/bfloat-abi.c (+611)
  • (added) clang/test/CodeGen/LoongArch/bfloat-mangle.cpp (+19)
  • (modified) llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp (+50-4)
  • (modified) llvm/lib/Target/LoongArch/LoongArchISelLowering.h (+2)
  • (added) llvm/test/CodeGen/LoongArch/bf16-promote.ll (+172)
  • (added) llvm/test/CodeGen/LoongArch/bf16.ll (+1048)
diff --git a/clang/docs/LanguageExtensions.rst b/clang/docs/LanguageExtensions.rst
index 01b45cf685959..34cea9cc0cf66 100644
--- a/clang/docs/LanguageExtensions.rst
+++ b/clang/docs/LanguageExtensions.rst
@@ -1009,6 +1009,7 @@ to ``float``; see below for more information on this emulation.
   * 64-bit ARM (AArch64)
   * RISC-V
   * X86 (when SSE2 is available)
+  * LoongArch
 
 (For X86, SSE2 is available on 64-bit and all recent 32-bit processors.)
 
diff --git a/clang/lib/Basic/Targets/LoongArch.h b/clang/lib/Basic/Targets/LoongArch.h
index 8a8c978ab89db..7e9affc98ac0f 100644
--- a/clang/lib/Basic/Targets/LoongArch.h
+++ b/clang/lib/Basic/Targets/LoongArch.h
@@ -49,6 +49,9 @@ class LLVM_LIBRARY_VISIBILITY LoongArchTargetInfo : public TargetInfo {
     HasFeatureLD_SEQ_SA = false;
     HasFeatureDiv32 = false;
     HasFeatureSCQ = false;
+    BFloat16Width = 16;
+    BFloat16Align = 16;
+    BFloat16Format = &llvm::APFloat::BFloat();
     LongDoubleWidth = 128;
     LongDoubleAlign = 128;
     LongDoubleFormat = &llvm::APFloat::IEEEquad();
@@ -99,6 +102,8 @@ class LLVM_LIBRARY_VISIBILITY LoongArchTargetInfo : public TargetInfo {
 
   bool hasBitIntType() const override { return true; }
 
+  bool hasBFloat16Type() const override { return true; }
+
   bool useFP16ConversionIntrinsics() const override { return false; }
 
   bool handleTargetFeatures(std::vector<std::string> &Features,
diff --git a/clang/test/CodeGen/LoongArch/bfloat-abi.c b/clang/test/CodeGen/LoongArch/bfloat-abi.c
new file mode 100644
index 0000000000000..9f0e25c17cc74
--- /dev/null
+++ b/clang/test/CodeGen/LoongArch/bfloat-abi.c
@@ -0,0 +1,611 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 2
+// RUN: %clang_cc1 -triple loongarch64 -emit-llvm %s -o - | FileCheck %s --check-prefixes=CHECK-LA64
+// RUN: %clang_cc1 -triple loongarch32 -emit-llvm %s -o - | FileCheck %s --check-prefixes=CHECK-LA32
+
+struct bfloat1 {
+  __bf16 a;
+};
+
+// CHECK-LA64-LABEL: define dso_local bfloat @h1
+// CHECK-LA64-SAME: (bfloat noundef [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT1:%.*]], align 2
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT1]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = getelementptr inbounds nuw { bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[TMP1]], align 2
+// CHECK-LA64-NEXT:    ret bfloat [[TMP2]]
+//
+// CHECK-LA32-LABEL: define dso_local bfloat @h1
+// CHECK-LA32-SAME: (bfloat noundef [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA32-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT1:%.*]], align 2
+// CHECK-LA32-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT1]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA32-NEXT:    [[TMP1:%.*]] = getelementptr inbounds nuw { bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[TMP1]], align 2
+// CHECK-LA32-NEXT:    ret bfloat [[TMP2]]
+//
+struct bfloat1 h1(__bf16 a) {
+  struct bfloat1 x;
+  x.a = a;
+  return x;
+}
+
+struct bfloat2 {
+  __bf16 a;
+  __bf16 b;
+};
+
+// CHECK-LA64-LABEL: define dso_local { bfloat, bfloat } @h2
+// CHECK-LA64-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]]) #[[ATTR0]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT2:%.*]], align 2
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT2]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT2]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw { bfloat, bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    [[TMP3:%.*]] = load bfloat, ptr [[TMP2]], align 2
+// CHECK-LA64-NEXT:    [[TMP4:%.*]] = getelementptr inbounds nuw { bfloat, bfloat }, ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    [[TMP5:%.*]] = load bfloat, ptr [[TMP4]], align 2
+// CHECK-LA64-NEXT:    [[TMP6:%.*]] = insertvalue { bfloat, bfloat } poison, bfloat [[TMP3]], 0
+// CHECK-LA64-NEXT:    [[TMP7:%.*]] = insertvalue { bfloat, bfloat } [[TMP6]], bfloat [[TMP5]], 1
+// CHECK-LA64-NEXT:    ret { bfloat, bfloat } [[TMP7]]
+//
+// CHECK-LA32-LABEL: define dso_local { bfloat, bfloat } @h2
+// CHECK-LA32-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]]) #[[ATTR0]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA32-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT2:%.*]], align 2
+// CHECK-LA32-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT2]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA32-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT2]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA32-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw { bfloat, bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    [[TMP3:%.*]] = load bfloat, ptr [[TMP2]], align 2
+// CHECK-LA32-NEXT:    [[TMP4:%.*]] = getelementptr inbounds nuw { bfloat, bfloat }, ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    [[TMP5:%.*]] = load bfloat, ptr [[TMP4]], align 2
+// CHECK-LA32-NEXT:    [[TMP6:%.*]] = insertvalue { bfloat, bfloat } poison, bfloat [[TMP3]], 0
+// CHECK-LA32-NEXT:    [[TMP7:%.*]] = insertvalue { bfloat, bfloat } [[TMP6]], bfloat [[TMP5]], 1
+// CHECK-LA32-NEXT:    ret { bfloat, bfloat } [[TMP7]]
+//
+struct bfloat2 h2(__bf16 a, __bf16 b) {
+  struct bfloat2 x;
+  x.a = a;
+  x.b = b;
+  return x;
+}
+
+struct bfloat3 {
+  __bf16 a;
+  __bf16 b;
+  __bf16 c;
+};
+
+// CHECK-LA64-LABEL: define dso_local i64 @h3
+// CHECK-LA64-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]]) #[[ATTR0]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT3:%.*]], align 2
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[C_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[RETVAL_COERCE:%.*]] = alloca i64, align 8
+// CHECK-LA64-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[C]], ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[C3:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 2
+// CHECK-LA64-NEXT:    store bfloat [[TMP2]], ptr [[C3]], align 2
+// CHECK-LA64-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[RETVAL_COERCE]], ptr align 2 [[RETVAL]], i64 6, i1 false)
+// CHECK-LA64-NEXT:    [[TMP3:%.*]] = load i64, ptr [[RETVAL_COERCE]], align 8
+// CHECK-LA64-NEXT:    ret i64 [[TMP3]]
+//
+// CHECK-LA32-LABEL: define dso_local [2 x i32] @h3
+// CHECK-LA32-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]]) #[[ATTR0]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA32-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT3:%.*]], align 2
+// CHECK-LA32-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[C_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[RETVAL_COERCE:%.*]] = alloca [2 x i32], align 4
+// CHECK-LA32-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[C]], ptr [[C_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA32-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA32-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[C_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[C3:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 2
+// CHECK-LA32-NEXT:    store bfloat [[TMP2]], ptr [[C3]], align 2
+// CHECK-LA32-NEXT:    call void @llvm.memcpy.p0.p0.i32(ptr align 4 [[RETVAL_COERCE]], ptr align 2 [[RETVAL]], i32 6, i1 false)
+// CHECK-LA32-NEXT:    [[TMP3:%.*]] = load [2 x i32], ptr [[RETVAL_COERCE]], align 4
+// CHECK-LA32-NEXT:    ret [2 x i32] [[TMP3]]
+//
+struct bfloat3 h3(__bf16 a, __bf16 b, __bf16 c) {
+  struct bfloat3 x;
+  x.a = a;
+  x.b = b;
+  x.c = c;
+  return x;
+}
+
+struct bfloat4 {
+  __bf16 a;
+  __bf16 b;
+  __bf16 c;
+  __bf16 d;
+};
+
+// CHECK-LA64-LABEL: define dso_local i64 @h4
+// CHECK-LA64-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]], bfloat noundef [[D:%.*]]) #[[ATTR0]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT4:%.*]], align 2
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[C_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[D_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[C]], ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[D]], ptr [[D_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[C3:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 2
+// CHECK-LA64-NEXT:    store bfloat [[TMP2]], ptr [[C3]], align 2
+// CHECK-LA64-NEXT:    [[TMP3:%.*]] = load bfloat, ptr [[D_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[D4:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 3
+// CHECK-LA64-NEXT:    store bfloat [[TMP3]], ptr [[D4]], align 2
+// CHECK-LA64-NEXT:    [[TMP4:%.*]] = load i64, ptr [[RETVAL]], align 2
+// CHECK-LA64-NEXT:    ret i64 [[TMP4]]
+//
+// CHECK-LA32-LABEL: define dso_local [2 x i32] @h4
+// CHECK-LA32-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]], bfloat noundef [[D:%.*]]) #[[ATTR0]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA32-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT4:%.*]], align 2
+// CHECK-LA32-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[C_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[D_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[C]], ptr [[C_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[D]], ptr [[D_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA32-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA32-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[C_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[C3:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 2
+// CHECK-LA32-NEXT:    store bfloat [[TMP2]], ptr [[C3]], align 2
+// CHECK-LA32-NEXT:    [[TMP3:%.*]] = load bfloat, ptr [[D_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[D4:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 3
+// CHECK-LA32-NEXT:    store bfloat [[TMP3]], ptr [[D4]], align 2
+// CHECK-LA32-NEXT:    [[TMP4:%.*]] = load [2 x i32], ptr [[RETVAL]], align 2
+// CHECK-LA32-NEXT:    ret [2 x i32] [[TMP4]]
+//
+struct bfloat4 h4(__bf16 a, __bf16 b, __bf16 c, __bf16 d) {
+  struct bfloat4 x;
+  x.a = a;
+  x.b = b;
+  x.c = c;
+  x.d = d;
+  return x;
+}
+
+struct floatbfloat {
+  float a;
+  __bf16 b;
+};
+
+// CHECK-LA64-LABEL: define dso_local { float, bfloat } @fh
+// CHECK-LA64-SAME: (float noundef [[A:%.*]], bfloat noundef [[B:%.*]]) #[[ATTR0]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_FLOATBFLOAT:%.*]], align 4
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca float, align 4
+// CHECK-LA64-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    store float [[A]], ptr [[A_ADDR]], align 4
+// CHECK-LA64-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load float, ptr [[A_ADDR]], align 4
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store float [[TMP0]], ptr [[A1]], align 4
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 4
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw { float, bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    [[TMP3:%.*]] = load float, ptr [[TMP2]], align 4
+// CHECK-LA64-NEXT:    [[TMP4:%.*]] = getelementptr inbounds nuw { float, bfloat }, ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    [[TMP5:%.*]] = load bfloat, ptr [[TMP4]], align 4
+// CHECK-LA64-NEXT:    [[TMP6:%.*]] = insertvalue { float, bfloat } poison, float [[TMP3]], 0
+// CHECK-LA64-NEXT:    [[TMP7:%.*]] = insertvalue { float, bfloat } [[TMP6]], bfloat [[TMP5]], 1
+// CHECK-LA64-NEXT:    ret { float, bfloat } [[TMP7]]
+//
+// CHECK-LA32-LABEL: define dso_local { float, bfloat } @fh
+// CHECK-LA32-SAME: (float noundef [[A:%.*]], bfloat noundef [[B:%.*]]) #[[ATTR0]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA32-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_FLOATBFLOAT:%.*]], align 4
+// CHECK-LA32-NEXT:    [[A_ADDR:%.*]] = alloca float, align 4
+// CHECK-LA32-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    store float [[A]], ptr [[A_ADDR]], align 4
+// CHECK-LA32-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[TMP0:%.*]] = load float, ptr [[A_ADDR]], align 4
+// CHECK-LA32-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    store float [[TMP0]], ptr [[A1]], align 4
+// CHECK-LA32-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 4
+// CHECK-LA32-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw { float, bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    [[TMP3:%.*]] = load float, ptr [[TMP2]], align 4
+// CHECK-LA32-NEXT:    [[TMP4:%.*]] = getelementptr inbounds nuw { float, bfloat }, ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    [[TMP5:%.*]] = load bfloat, ptr [[TMP4]], align 4
+// CHECK-LA32-NEXT:    [[TMP6:%.*]] = insertvalue { float, bfloat } poison, float [[TMP3]], 0
+// CHECK-LA32-NEXT:    [[TMP7:%.*]] = insertvalue { float, bfloat } [[TMP6]], bfloat [[TMP5]], 1
+// CHECK-LA32-NEXT:    ret { float, bfloat } [[TMP7]]
+//
+struct floatbfloat fh(float a, __bf16 b) {
+  struct floatbfloat x;
+  x.a = a;
+  x.b = b;
+  return x;
+}
+
+struct floatbfloat2 {
+  float a;
+  __bf16 b;
+  __bf16 c;
+};
+
+// CHECK-LA64-LABEL: define dso_local i64 @fh2
+// CHECK-LA64-SAME: (float noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]]) #[[ATTR0]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_FLOATBFLOAT2:%.*]], align 4
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca float, align 4
+// CHECK-LA64-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[C_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    store float [[A]], ptr [[A_ADDR]], align 4
+// CHECK-LA64-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[C]], ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load float, ptr [[A_ADDR]], align 4
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT2]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store float [[TMP0]], ptr [[A1]], align 4
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT2]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 4
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[C3:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT2]], ptr [[RETVAL]], i32 0, i32 2
+// CHECK-LA64-NEXT:    store bfloat [[TMP2]], ptr [[C3]], align 2
+// CHECK-LA64-NEXT:    [[TMP3:%.*]] = load i64, ptr [[RETVAL]], align 4
+// CHECK-LA64-NEXT:    ret i64 [[TMP3]]
+//
+// CHECK-LA32-LABEL: define dso_local [2 x i32] @fh2
+// CHECK-LA32-SAME: (float noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]]) #[[ATTR0]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA3...
[truncated]

@llvmbot
Copy link
Member

llvmbot commented Jun 3, 2025

@llvm/pr-subscribers-clang

Author: None (Ami-zhang)

Changes

The LoongArch psABI recently added __bf16 type support. Now we can enable this new type in clang.

Currently, bf16 operations are automatically supported by promoting to float. This patch adds bf16 support by ensuring that load extension / truncate store operations are properly expanded.

And this commit implements support for bf16 truncate/extend on hard FP targets. The extend operation is implemented by a shift just as in the standard legalization. This requires custom lowering of the truncate libcall on hard float ABIs (the normal libcall code path is used on soft ABIs).

Patch is 92.03 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/142548.diff

8 Files Affected:

  • (modified) clang/docs/LanguageExtensions.rst (+1)
  • (modified) clang/lib/Basic/Targets/LoongArch.h (+5)
  • (added) clang/test/CodeGen/LoongArch/bfloat-abi.c (+611)
  • (added) clang/test/CodeGen/LoongArch/bfloat-mangle.cpp (+19)
  • (modified) llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp (+50-4)
  • (modified) llvm/lib/Target/LoongArch/LoongArchISelLowering.h (+2)
  • (added) llvm/test/CodeGen/LoongArch/bf16-promote.ll (+172)
  • (added) llvm/test/CodeGen/LoongArch/bf16.ll (+1048)
diff --git a/clang/docs/LanguageExtensions.rst b/clang/docs/LanguageExtensions.rst
index 01b45cf685959..34cea9cc0cf66 100644
--- a/clang/docs/LanguageExtensions.rst
+++ b/clang/docs/LanguageExtensions.rst
@@ -1009,6 +1009,7 @@ to ``float``; see below for more information on this emulation.
   * 64-bit ARM (AArch64)
   * RISC-V
   * X86 (when SSE2 is available)
+  * LoongArch
 
 (For X86, SSE2 is available on 64-bit and all recent 32-bit processors.)
 
diff --git a/clang/lib/Basic/Targets/LoongArch.h b/clang/lib/Basic/Targets/LoongArch.h
index 8a8c978ab89db..7e9affc98ac0f 100644
--- a/clang/lib/Basic/Targets/LoongArch.h
+++ b/clang/lib/Basic/Targets/LoongArch.h
@@ -49,6 +49,9 @@ class LLVM_LIBRARY_VISIBILITY LoongArchTargetInfo : public TargetInfo {
     HasFeatureLD_SEQ_SA = false;
     HasFeatureDiv32 = false;
     HasFeatureSCQ = false;
+    BFloat16Width = 16;
+    BFloat16Align = 16;
+    BFloat16Format = &llvm::APFloat::BFloat();
     LongDoubleWidth = 128;
     LongDoubleAlign = 128;
     LongDoubleFormat = &llvm::APFloat::IEEEquad();
@@ -99,6 +102,8 @@ class LLVM_LIBRARY_VISIBILITY LoongArchTargetInfo : public TargetInfo {
 
   bool hasBitIntType() const override { return true; }
 
+  bool hasBFloat16Type() const override { return true; }
+
   bool useFP16ConversionIntrinsics() const override { return false; }
 
   bool handleTargetFeatures(std::vector<std::string> &Features,
diff --git a/clang/test/CodeGen/LoongArch/bfloat-abi.c b/clang/test/CodeGen/LoongArch/bfloat-abi.c
new file mode 100644
index 0000000000000..9f0e25c17cc74
--- /dev/null
+++ b/clang/test/CodeGen/LoongArch/bfloat-abi.c
@@ -0,0 +1,611 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 2
+// RUN: %clang_cc1 -triple loongarch64 -emit-llvm %s -o - | FileCheck %s --check-prefixes=CHECK-LA64
+// RUN: %clang_cc1 -triple loongarch32 -emit-llvm %s -o - | FileCheck %s --check-prefixes=CHECK-LA32
+
+struct bfloat1 {
+  __bf16 a;
+};
+
+// CHECK-LA64-LABEL: define dso_local bfloat @h1
+// CHECK-LA64-SAME: (bfloat noundef [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT1:%.*]], align 2
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT1]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = getelementptr inbounds nuw { bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[TMP1]], align 2
+// CHECK-LA64-NEXT:    ret bfloat [[TMP2]]
+//
+// CHECK-LA32-LABEL: define dso_local bfloat @h1
+// CHECK-LA32-SAME: (bfloat noundef [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA32-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT1:%.*]], align 2
+// CHECK-LA32-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT1]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA32-NEXT:    [[TMP1:%.*]] = getelementptr inbounds nuw { bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[TMP1]], align 2
+// CHECK-LA32-NEXT:    ret bfloat [[TMP2]]
+//
+struct bfloat1 h1(__bf16 a) {
+  struct bfloat1 x;
+  x.a = a;
+  return x;
+}
+
+struct bfloat2 {
+  __bf16 a;
+  __bf16 b;
+};
+
+// CHECK-LA64-LABEL: define dso_local { bfloat, bfloat } @h2
+// CHECK-LA64-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]]) #[[ATTR0]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT2:%.*]], align 2
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT2]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT2]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw { bfloat, bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    [[TMP3:%.*]] = load bfloat, ptr [[TMP2]], align 2
+// CHECK-LA64-NEXT:    [[TMP4:%.*]] = getelementptr inbounds nuw { bfloat, bfloat }, ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    [[TMP5:%.*]] = load bfloat, ptr [[TMP4]], align 2
+// CHECK-LA64-NEXT:    [[TMP6:%.*]] = insertvalue { bfloat, bfloat } poison, bfloat [[TMP3]], 0
+// CHECK-LA64-NEXT:    [[TMP7:%.*]] = insertvalue { bfloat, bfloat } [[TMP6]], bfloat [[TMP5]], 1
+// CHECK-LA64-NEXT:    ret { bfloat, bfloat } [[TMP7]]
+//
+// CHECK-LA32-LABEL: define dso_local { bfloat, bfloat } @h2
+// CHECK-LA32-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]]) #[[ATTR0]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA32-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT2:%.*]], align 2
+// CHECK-LA32-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT2]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA32-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT2]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA32-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw { bfloat, bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    [[TMP3:%.*]] = load bfloat, ptr [[TMP2]], align 2
+// CHECK-LA32-NEXT:    [[TMP4:%.*]] = getelementptr inbounds nuw { bfloat, bfloat }, ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    [[TMP5:%.*]] = load bfloat, ptr [[TMP4]], align 2
+// CHECK-LA32-NEXT:    [[TMP6:%.*]] = insertvalue { bfloat, bfloat } poison, bfloat [[TMP3]], 0
+// CHECK-LA32-NEXT:    [[TMP7:%.*]] = insertvalue { bfloat, bfloat } [[TMP6]], bfloat [[TMP5]], 1
+// CHECK-LA32-NEXT:    ret { bfloat, bfloat } [[TMP7]]
+//
+struct bfloat2 h2(__bf16 a, __bf16 b) {
+  struct bfloat2 x;
+  x.a = a;
+  x.b = b;
+  return x;
+}
+
+struct bfloat3 {
+  __bf16 a;
+  __bf16 b;
+  __bf16 c;
+};
+
+// CHECK-LA64-LABEL: define dso_local i64 @h3
+// CHECK-LA64-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]]) #[[ATTR0]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT3:%.*]], align 2
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[C_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[RETVAL_COERCE:%.*]] = alloca i64, align 8
+// CHECK-LA64-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[C]], ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[C3:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 2
+// CHECK-LA64-NEXT:    store bfloat [[TMP2]], ptr [[C3]], align 2
+// CHECK-LA64-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[RETVAL_COERCE]], ptr align 2 [[RETVAL]], i64 6, i1 false)
+// CHECK-LA64-NEXT:    [[TMP3:%.*]] = load i64, ptr [[RETVAL_COERCE]], align 8
+// CHECK-LA64-NEXT:    ret i64 [[TMP3]]
+//
+// CHECK-LA32-LABEL: define dso_local [2 x i32] @h3
+// CHECK-LA32-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]]) #[[ATTR0]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA32-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT3:%.*]], align 2
+// CHECK-LA32-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[C_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[RETVAL_COERCE:%.*]] = alloca [2 x i32], align 4
+// CHECK-LA32-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[C]], ptr [[C_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA32-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA32-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[C_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[C3:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT3]], ptr [[RETVAL]], i32 0, i32 2
+// CHECK-LA32-NEXT:    store bfloat [[TMP2]], ptr [[C3]], align 2
+// CHECK-LA32-NEXT:    call void @llvm.memcpy.p0.p0.i32(ptr align 4 [[RETVAL_COERCE]], ptr align 2 [[RETVAL]], i32 6, i1 false)
+// CHECK-LA32-NEXT:    [[TMP3:%.*]] = load [2 x i32], ptr [[RETVAL_COERCE]], align 4
+// CHECK-LA32-NEXT:    ret [2 x i32] [[TMP3]]
+//
+struct bfloat3 h3(__bf16 a, __bf16 b, __bf16 c) {
+  struct bfloat3 x;
+  x.a = a;
+  x.b = b;
+  x.c = c;
+  return x;
+}
+
+struct bfloat4 {
+  __bf16 a;
+  __bf16 b;
+  __bf16 c;
+  __bf16 d;
+};
+
+// CHECK-LA64-LABEL: define dso_local i64 @h4
+// CHECK-LA64-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]], bfloat noundef [[D:%.*]]) #[[ATTR0]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT4:%.*]], align 2
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[C_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[D_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[C]], ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[D]], ptr [[D_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[C3:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 2
+// CHECK-LA64-NEXT:    store bfloat [[TMP2]], ptr [[C3]], align 2
+// CHECK-LA64-NEXT:    [[TMP3:%.*]] = load bfloat, ptr [[D_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[D4:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 3
+// CHECK-LA64-NEXT:    store bfloat [[TMP3]], ptr [[D4]], align 2
+// CHECK-LA64-NEXT:    [[TMP4:%.*]] = load i64, ptr [[RETVAL]], align 2
+// CHECK-LA64-NEXT:    ret i64 [[TMP4]]
+//
+// CHECK-LA32-LABEL: define dso_local [2 x i32] @h4
+// CHECK-LA32-SAME: (bfloat noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]], bfloat noundef [[D:%.*]]) #[[ATTR0]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA32-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_BFLOAT4:%.*]], align 2
+// CHECK-LA32-NEXT:    [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[C_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    [[D_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[C]], ptr [[C_ADDR]], align 2
+// CHECK-LA32-NEXT:    store bfloat [[D]], ptr [[D_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    store bfloat [[TMP0]], ptr [[A1]], align 2
+// CHECK-LA32-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 2
+// CHECK-LA32-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[C_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[C3:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 2
+// CHECK-LA32-NEXT:    store bfloat [[TMP2]], ptr [[C3]], align 2
+// CHECK-LA32-NEXT:    [[TMP3:%.*]] = load bfloat, ptr [[D_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[D4:%.*]] = getelementptr inbounds nuw [[STRUCT_BFLOAT4]], ptr [[RETVAL]], i32 0, i32 3
+// CHECK-LA32-NEXT:    store bfloat [[TMP3]], ptr [[D4]], align 2
+// CHECK-LA32-NEXT:    [[TMP4:%.*]] = load [2 x i32], ptr [[RETVAL]], align 2
+// CHECK-LA32-NEXT:    ret [2 x i32] [[TMP4]]
+//
+struct bfloat4 h4(__bf16 a, __bf16 b, __bf16 c, __bf16 d) {
+  struct bfloat4 x;
+  x.a = a;
+  x.b = b;
+  x.c = c;
+  x.d = d;
+  return x;
+}
+
+struct floatbfloat {
+  float a;
+  __bf16 b;
+};
+
+// CHECK-LA64-LABEL: define dso_local { float, bfloat } @fh
+// CHECK-LA64-SAME: (float noundef [[A:%.*]], bfloat noundef [[B:%.*]]) #[[ATTR0]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_FLOATBFLOAT:%.*]], align 4
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca float, align 4
+// CHECK-LA64-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    store float [[A]], ptr [[A_ADDR]], align 4
+// CHECK-LA64-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load float, ptr [[A_ADDR]], align 4
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store float [[TMP0]], ptr [[A1]], align 4
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 4
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw { float, bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    [[TMP3:%.*]] = load float, ptr [[TMP2]], align 4
+// CHECK-LA64-NEXT:    [[TMP4:%.*]] = getelementptr inbounds nuw { float, bfloat }, ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    [[TMP5:%.*]] = load bfloat, ptr [[TMP4]], align 4
+// CHECK-LA64-NEXT:    [[TMP6:%.*]] = insertvalue { float, bfloat } poison, float [[TMP3]], 0
+// CHECK-LA64-NEXT:    [[TMP7:%.*]] = insertvalue { float, bfloat } [[TMP6]], bfloat [[TMP5]], 1
+// CHECK-LA64-NEXT:    ret { float, bfloat } [[TMP7]]
+//
+// CHECK-LA32-LABEL: define dso_local { float, bfloat } @fh
+// CHECK-LA32-SAME: (float noundef [[A:%.*]], bfloat noundef [[B:%.*]]) #[[ATTR0]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA32-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_FLOATBFLOAT:%.*]], align 4
+// CHECK-LA32-NEXT:    [[A_ADDR:%.*]] = alloca float, align 4
+// CHECK-LA32-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA32-NEXT:    store float [[A]], ptr [[A_ADDR]], align 4
+// CHECK-LA32-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[TMP0:%.*]] = load float, ptr [[A_ADDR]], align 4
+// CHECK-LA32-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    store float [[TMP0]], ptr [[A1]], align 4
+// CHECK-LA32-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA32-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 4
+// CHECK-LA32-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw { float, bfloat }, ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA32-NEXT:    [[TMP3:%.*]] = load float, ptr [[TMP2]], align 4
+// CHECK-LA32-NEXT:    [[TMP4:%.*]] = getelementptr inbounds nuw { float, bfloat }, ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA32-NEXT:    [[TMP5:%.*]] = load bfloat, ptr [[TMP4]], align 4
+// CHECK-LA32-NEXT:    [[TMP6:%.*]] = insertvalue { float, bfloat } poison, float [[TMP3]], 0
+// CHECK-LA32-NEXT:    [[TMP7:%.*]] = insertvalue { float, bfloat } [[TMP6]], bfloat [[TMP5]], 1
+// CHECK-LA32-NEXT:    ret { float, bfloat } [[TMP7]]
+//
+struct floatbfloat fh(float a, __bf16 b) {
+  struct floatbfloat x;
+  x.a = a;
+  x.b = b;
+  return x;
+}
+
+struct floatbfloat2 {
+  float a;
+  __bf16 b;
+  __bf16 c;
+};
+
+// CHECK-LA64-LABEL: define dso_local i64 @fh2
+// CHECK-LA64-SAME: (float noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]]) #[[ATTR0]] {
+// CHECK-LA64-NEXT:  entry:
+// CHECK-LA64-NEXT:    [[RETVAL:%.*]] = alloca [[STRUCT_FLOATBFLOAT2:%.*]], align 4
+// CHECK-LA64-NEXT:    [[A_ADDR:%.*]] = alloca float, align 4
+// CHECK-LA64-NEXT:    [[B_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    [[C_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-LA64-NEXT:    store float [[A]], ptr [[A_ADDR]], align 4
+// CHECK-LA64-NEXT:    store bfloat [[B]], ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    store bfloat [[C]], ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[TMP0:%.*]] = load float, ptr [[A_ADDR]], align 4
+// CHECK-LA64-NEXT:    [[A1:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT2]], ptr [[RETVAL]], i32 0, i32 0
+// CHECK-LA64-NEXT:    store float [[TMP0]], ptr [[A1]], align 4
+// CHECK-LA64-NEXT:    [[TMP1:%.*]] = load bfloat, ptr [[B_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[B2:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT2]], ptr [[RETVAL]], i32 0, i32 1
+// CHECK-LA64-NEXT:    store bfloat [[TMP1]], ptr [[B2]], align 4
+// CHECK-LA64-NEXT:    [[TMP2:%.*]] = load bfloat, ptr [[C_ADDR]], align 2
+// CHECK-LA64-NEXT:    [[C3:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOATBFLOAT2]], ptr [[RETVAL]], i32 0, i32 2
+// CHECK-LA64-NEXT:    store bfloat [[TMP2]], ptr [[C3]], align 2
+// CHECK-LA64-NEXT:    [[TMP3:%.*]] = load i64, ptr [[RETVAL]], align 4
+// CHECK-LA64-NEXT:    ret i64 [[TMP3]]
+//
+// CHECK-LA32-LABEL: define dso_local [2 x i32] @fh2
+// CHECK-LA32-SAME: (float noundef [[A:%.*]], bfloat noundef [[B:%.*]], bfloat noundef [[C:%.*]]) #[[ATTR0]] {
+// CHECK-LA32-NEXT:  entry:
+// CHECK-LA3...
[truncated]

heiher
heiher reviewed Jun 5, 2025
View reviewed changes
clang/test/CodeGen/LoongArch/bfloat-abi.c Outdated
Comment on lines 2 to 3
Copy link
Member

@heiher heiher Jun 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
// RUN: %clang_cc1 -triple loongarch64 -emit-llvm %s -o - | FileCheck %s --check-prefixes=CHECK-LA64
// RUN: %clang_cc1 -triple loongarch32 -emit-llvm %s -o - | FileCheck %s --check-prefixes=CHECK-LA32
// RUN: %clang_cc1 -triple loongarch64 -emit-llvm %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-LA64
// RUN: %clang_cc1 -triple loongarch32 -emit-llvm %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-LA32

Copy link
Contributor Author

@Ami-zhang Ami-zhang Jun 6, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done.

clang/test/CodeGen/LoongArch/bfloat-mangle.cpp Outdated
Comment on lines 2 to 3
Copy link
Member

@heiher heiher Jun 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
// RUN: %clang_cc1 -triple loongarch64 -emit-llvm -o - %s | FileCheck %s --check-prefixes=CHECK-LA64
// RUN: %clang_cc1 -triple loongarch32 -emit-llvm -o - %s | FileCheck %s --check-prefixes=CHECK-LA32
// RUN: %clang_cc1 -triple loongarch64 -emit-llvm -o - %s | FileCheck %s --check-prefixes=CHECK,CHECK-LA64
// RUN: %clang_cc1 -triple loongarch32 -emit-llvm -o - %s | FileCheck %s --check-prefixes=CHECK,CHECK-LA32

Copy link
Contributor Author

@Ami-zhang Ami-zhang Jun 6, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Both LA32 and LA64 follow the C++ ABI for __bf16 name mangling, so the default FileCheck prefix can be used for verification.

@Ami-zhang
[LoongArch][BF16] Add support for the __bf16 type
fef9547 
The LoongArch psABI recently added __bf16 type support.
Now we can enable this new type in clang.

Currently, bf16 operations are automatically supported by promoting to float.
This patch adds bf16 support by ensuring that load extension / truncate store
operations are properly expanded.

And this commit implements support for bf16 truncate/extend on hard FP targets.
The extend operation is implemented by a shift just as in the standard legalization.
This requires custom lowering of the truncate libcall on hard float ABIs (the
normal libcall code path is used on soft ABIs).
heiher
heiher approved these changes Jun 6, 2025
View reviewed changes
Copy link
Member

@heiher heiher left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM.

@Ami-zhang Ami-zhang merged commit 0ed5d9a into llvm:main Jun 9, 2025
8 checks passed
@llvm-ci
Copy link
Collaborator

llvm-ci commented Jun 9, 2025

LLVM Buildbot has detected a new failure on builder llvm-clang-x86_64-gcc-ubuntu-no-asserts running on doug-worker-6 while building clang,llvm at step 6 "test-build-unified-tree-check-all".

Full details are available at: https://lab.llvm.org/buildbot/#/builders/202/builds/1690

Here is the relevant piece of the build log for the reference 
Step 6 (test-build-unified-tree-check-all) failure: 1200 seconds without output running [b'ninja', b'check-all'], attempting to kill
...
PASS: lit :: unique-output-file.py (88707 of 88717)
PASS: lit :: usage.py (88708 of 88717)
PASS: lit :: use-llvm-tool.py (88709 of 88717)
PASS: lit :: selecting.py (88710 of 88717)
PASS: lit :: xunit-output-report-failures-only.py (88711 of 88717)
PASS: lit :: xunit-output.py (88712 of 88717)
PASS: lit :: xfail-cl.py (88713 of 88717)
PASS: lit :: shtest-timeout.py (88714 of 88717)
PASS: lit :: shtest-define.py (88715 of 88717)
PASS: lit :: shtest-shell.py (88716 of 88717)
command timed out: 1200 seconds without output running [b'ninja', b'check-all'], attempting to kill
process killed by signal 9
program finished with exit code -1
elapsedTime=3873.821627

tomtor pushed a commit to tomtor/llvm-project that referenced this pull request Jun 14, 2025
@Ami-zhang @tomtor
[LoongArch][BF16] Add support for the __bf16 type (llvm#142548)
20461ea 
The LoongArch psABI recently added __bf16 type support. Now we can
enable this new type in clang.

Currently, bf16 operations are automatically supported by promoting to
float. This patch adds bf16 support by ensuring that load extension /
truncate store operations are properly expanded.

And this commit implements support for bf16 truncate/extend on hard FP
targets. The extend operation is implemented by a shift just as in the
standard legalization. This requires custom lowering of the truncate
libcall on hard float ABIs (the normal libcall code path is used on soft
ABIs).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@heiher heiher heiher approved these changes

@wangleiat wangleiat Awaiting requested review from wangleiat

@SixWeining SixWeining Awaiting requested review from SixWeining

Assignees

No one assigned

Labels

backend:loongarch clang:frontend Language frontend issues, e.g. anything involving "Sema" clang Clang issues not falling into any other category

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@Ami-zhang @llvmbot @llvm-ci @heiher