[clang] [clang] fix half && bfloat16 convert node expr codegen (PR #89051)
via cfe-commits
cfe-commits at lists.llvm.org
Wed Apr 17 20:25:06 PDT 2024
https://github.com/JinjinLi868 updated https://github.com/llvm/llvm-project/pull/89051
>From 8d0ed7653c21315c5c920114b3a2d7686e54b9ca Mon Sep 17 00:00:00 2001
From: Jinjin Li <lijinjin.868 at bytedance.com>
Date: Wed, 17 Apr 2024 16:44:50 +0800
Subject: [PATCH] [clang] Fix half && bfloat16 convert node expr codegen
Data type conversion between fp16 and bf16 will generate fptrunc
and fpextend nodes, but they are actually bitcast nodes.
---
clang/lib/CodeGen/CGExprScalar.cpp | 11 +-
.../test/CodeGen/X86/bfloat16-convert-half.c | 109 ++++++++++++++++++
2 files changed, 117 insertions(+), 3 deletions(-)
create mode 100644 clang/test/CodeGen/X86/bfloat16-convert-half.c
diff --git a/clang/lib/CodeGen/CGExprScalar.cpp b/clang/lib/CodeGen/CGExprScalar.cpp
index 1f18e0d5ba409a..da5a410f040d1b 100644
--- a/clang/lib/CodeGen/CGExprScalar.cpp
+++ b/clang/lib/CodeGen/CGExprScalar.cpp
@@ -1431,9 +1431,12 @@ Value *ScalarExprEmitter::EmitScalarCast(Value *Src, QualType SrcType,
return Builder.CreateFPToUI(Src, DstTy, "conv");
}
- if (DstElementTy->getTypeID() < SrcElementTy->getTypeID())
+ if ((DstElementTy->is16bitFPTy() && SrcElementTy->is16bitFPTy()))
+ return Builder.CreateBitCast(Src, DstTy, "conv");
+ else if (DstElementTy->getTypeID() < SrcElementTy->getTypeID())
return Builder.CreateFPTrunc(Src, DstTy, "conv");
- return Builder.CreateFPExt(Src, DstTy, "conv");
+ else
+ return Builder.CreateFPExt(Src, DstTy, "conv");
}
/// Emit a conversion from the specified type to the specified destination type,
@@ -1906,7 +1909,9 @@ Value *ScalarExprEmitter::VisitConvertVectorExpr(ConvertVectorExpr *E) {
} else {
assert(SrcEltTy->isFloatingPointTy() && DstEltTy->isFloatingPointTy() &&
"Unknown real conversion");
- if (DstEltTy->getTypeID() < SrcEltTy->getTypeID())
+ if ((DstEltTy->is16bitFPTy() && SrcEltTy->is16bitFPTy()))
+ Res = Builder.CreateBitCast(Src, DstTy, "conv");
+ else if (DstEltTy->getTypeID() < SrcEltTy->getTypeID())
Res = Builder.CreateFPTrunc(Src, DstTy, "conv");
else
Res = Builder.CreateFPExt(Src, DstTy, "conv");
diff --git a/clang/test/CodeGen/X86/bfloat16-convert-half.c b/clang/test/CodeGen/X86/bfloat16-convert-half.c
new file mode 100644
index 00000000000000..138793c242db16
--- /dev/null
+++ b/clang/test/CodeGen/X86/bfloat16-convert-half.c
@@ -0,0 +1,109 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 4
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -target-feature +fullbf16 -S -emit-llvm %s -o - | FileCheck %s
+// CHECK-LABEL: define dso_local half @test_convert_from_bf16_to_fp16(
+// CHECK-SAME: bfloat noundef [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-NEXT: entry:
+// CHECK-NEXT: [[A_ADDR:%.*]] = alloca bfloat, align 2
+// CHECK-NEXT: store bfloat [[A]], ptr [[A_ADDR]], align 2
+// CHECK-NEXT: [[TMP0:%.*]] = load bfloat, ptr [[A_ADDR]], align 2
+// CHECK-NEXT: [[CONV:%.*]] = bitcast bfloat [[TMP0]] to half
+// CHECK-NEXT: ret half [[CONV]]
+//
+_Float16 test_convert_from_bf16_to_fp16(__bf16 a) {
+ return (_Float16)a;
+}
+
+// CHECK-LABEL: define dso_local bfloat @test_convert_from_fp16_to_bf16(
+// CHECK-SAME: half noundef [[A:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT: entry:
+// CHECK-NEXT: [[A_ADDR:%.*]] = alloca half, align 2
+// CHECK-NEXT: store half [[A]], ptr [[A_ADDR]], align 2
+// CHECK-NEXT: [[TMP0:%.*]] = load half, ptr [[A_ADDR]], align 2
+// CHECK-NEXT: [[CONV:%.*]] = bitcast half [[TMP0]] to bfloat
+// CHECK-NEXT: ret bfloat [[CONV]]
+//
+__bf16 test_convert_from_fp16_to_bf16(_Float16 a) {
+ return (__bf16)a;
+}
+
+typedef _Float16 _Float162 __attribute__((ext_vector_type(2)));
+typedef _Float16 _Float164 __attribute__((ext_vector_type(4)));
+
+typedef __bf16 __bf162 __attribute__((ext_vector_type(2)));
+typedef __bf16 __bf164 __attribute__((ext_vector_type(4)));
+
+// CHECK-LABEL: define dso_local i32 @test_cast_from_fp162_to_bf162(
+// CHECK-SAME: i32 noundef [[IN_COERCE:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT: entry:
+// CHECK-NEXT: [[RETVAL:%.*]] = alloca <2 x bfloat>, align 4
+// CHECK-NEXT: [[IN:%.*]] = alloca <2 x half>, align 4
+// CHECK-NEXT: [[IN_ADDR:%.*]] = alloca <2 x half>, align 4
+// CHECK-NEXT: store i32 [[IN_COERCE]], ptr [[IN]], align 4
+// CHECK-NEXT: [[IN1:%.*]] = load <2 x half>, ptr [[IN]], align 4
+// CHECK-NEXT: store <2 x half> [[IN1]], ptr [[IN_ADDR]], align 4
+// CHECK-NEXT: [[TMP0:%.*]] = load <2 x half>, ptr [[IN_ADDR]], align 4
+// CHECK-NEXT: [[CONV:%.*]] = bitcast <2 x half> [[TMP0]] to <2 x bfloat>
+// CHECK-NEXT: store <2 x bfloat> [[CONV]], ptr [[RETVAL]], align 4
+// CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[RETVAL]], align 4
+// CHECK-NEXT: ret i32 [[TMP1]]
+//
+__bf162 test_cast_from_fp162_to_bf162(_Float162 in) {
+ return __builtin_convertvector(in, __bf162);
+}
+
+// CHECK-LABEL: define dso_local double @test_cast_from_fp164_to_bf164(
+// CHECK-SAME: double noundef [[IN_COERCE:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT: entry:
+// CHECK-NEXT: [[RETVAL:%.*]] = alloca <4 x bfloat>, align 8
+// CHECK-NEXT: [[IN:%.*]] = alloca <4 x half>, align 8
+// CHECK-NEXT: [[IN_ADDR:%.*]] = alloca <4 x half>, align 8
+// CHECK-NEXT: store double [[IN_COERCE]], ptr [[IN]], align 8
+// CHECK-NEXT: [[IN1:%.*]] = load <4 x half>, ptr [[IN]], align 8
+// CHECK-NEXT: store <4 x half> [[IN1]], ptr [[IN_ADDR]], align 8
+// CHECK-NEXT: [[TMP0:%.*]] = load <4 x half>, ptr [[IN_ADDR]], align 8
+// CHECK-NEXT: [[CONV:%.*]] = bitcast <4 x half> [[TMP0]] to <4 x bfloat>
+// CHECK-NEXT: store <4 x bfloat> [[CONV]], ptr [[RETVAL]], align 8
+// CHECK-NEXT: [[TMP1:%.*]] = load double, ptr [[RETVAL]], align 8
+// CHECK-NEXT: ret double [[TMP1]]
+//
+__bf164 test_cast_from_fp164_to_bf164(_Float164 in) {
+ return __builtin_convertvector(in, __bf164);
+}
+
+// CHECK-LABEL: define dso_local i32 @test_cast_from_bf162_to_fp162(
+// CHECK-SAME: i32 noundef [[IN_COERCE:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT: entry:
+// CHECK-NEXT: [[RETVAL:%.*]] = alloca <2 x half>, align 4
+// CHECK-NEXT: [[IN:%.*]] = alloca <2 x bfloat>, align 4
+// CHECK-NEXT: [[IN_ADDR:%.*]] = alloca <2 x bfloat>, align 4
+// CHECK-NEXT: store i32 [[IN_COERCE]], ptr [[IN]], align 4
+// CHECK-NEXT: [[IN1:%.*]] = load <2 x bfloat>, ptr [[IN]], align 4
+// CHECK-NEXT: store <2 x bfloat> [[IN1]], ptr [[IN_ADDR]], align 4
+// CHECK-NEXT: [[TMP0:%.*]] = load <2 x bfloat>, ptr [[IN_ADDR]], align 4
+// CHECK-NEXT: [[CONV:%.*]] = bitcast <2 x bfloat> [[TMP0]] to <2 x half>
+// CHECK-NEXT: store <2 x half> [[CONV]], ptr [[RETVAL]], align 4
+// CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[RETVAL]], align 4
+// CHECK-NEXT: ret i32 [[TMP1]]
+//
+_Float162 test_cast_from_bf162_to_fp162(__bf162 in) {
+ return __builtin_convertvector(in, _Float162);
+}
+
+// CHECK-LABEL: define dso_local double @test_cast_from_bf164_to_fp164(
+// CHECK-SAME: double noundef [[IN_COERCE:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT: entry:
+// CHECK-NEXT: [[RETVAL:%.*]] = alloca <4 x half>, align 8
+// CHECK-NEXT: [[IN:%.*]] = alloca <4 x bfloat>, align 8
+// CHECK-NEXT: [[IN_ADDR:%.*]] = alloca <4 x bfloat>, align 8
+// CHECK-NEXT: store double [[IN_COERCE]], ptr [[IN]], align 8
+// CHECK-NEXT: [[IN1:%.*]] = load <4 x bfloat>, ptr [[IN]], align 8
+// CHECK-NEXT: store <4 x bfloat> [[IN1]], ptr [[IN_ADDR]], align 8
+// CHECK-NEXT: [[TMP0:%.*]] = load <4 x bfloat>, ptr [[IN_ADDR]], align 8
+// CHECK-NEXT: [[CONV:%.*]] = bitcast <4 x bfloat> [[TMP0]] to <4 x half>
+// CHECK-NEXT: store <4 x half> [[CONV]], ptr [[RETVAL]], align 8
+// CHECK-NEXT: [[TMP1:%.*]] = load double, ptr [[RETVAL]], align 8
+// CHECK-NEXT: ret double [[TMP1]]
+//
+_Float164 test_cast_from_bf164_to_fp164(__bf164 in) {
+ return __builtin_convertvector(in, _Float164);
+}
More information about the cfe-commits
mailing list