[clang] [clang] fix half && bfloat16 convert node expr codegen (PR #89051)

[Matt Arsenault via cfe-commits]

================
@@ -0,0 +1,194 @@
+// 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:%.*]] = fpext bfloat [[TMP0]] to float
+// CHECK-NEXT:    [[CONV1:%.*]] = fptrunc float [[CONV]] to half
+// CHECK-NEXT:    ret half [[CONV1]]
+//
+_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:%.*]] = fpext half [[TMP0]] to float
+// CHECK-NEXT:    [[CONV1:%.*]] = fptrunc float [[CONV]] to bfloat
+// CHECK-NEXT:    ret bfloat [[CONV1]]
+//
+__bf16 test_convert_from_fp16_to_bf16(_Float16 a) {
+    return (__bf16)a;
+}
+
+typedef _Float16 half2 __attribute__((ext_vector_type(2)));
+typedef _Float16 half4 __attribute__((ext_vector_type(4)));
+
+typedef __bf16 bfloat2 __attribute__((ext_vector_type(2)));
+typedef __bf16 bfloat4 __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:    [[TMP1:%.*]] = bitcast <2 x half> [[TMP0]] to <2 x bfloat>
+// CHECK-NEXT:    store <2 x bfloat> [[TMP1]], ptr [[RETVAL]], align 4
+// CHECK-NEXT:    [[TMP2:%.*]] = load i32, ptr [[RETVAL]], align 4
+// CHECK-NEXT:    ret i32 [[TMP2]]
+//
+bfloat2 test_cast_from_fp162_to_bf162(half2 in) {
+  return (bfloat2)in;
+}
+
+
+// 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:    [[TMP1:%.*]] = bitcast <4 x half> [[TMP0]] to <4 x bfloat>
+// CHECK-NEXT:    store <4 x bfloat> [[TMP1]], ptr [[RETVAL]], align 8
+// CHECK-NEXT:    [[TMP2:%.*]] = load double, ptr [[RETVAL]], align 8
+// CHECK-NEXT:    ret double [[TMP2]]
+//
+bfloat4 test_cast_from_fp164_to_bf164(half4 in) {
+  return (bfloat4)in;
+}
+
+// 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:    [[TMP1:%.*]] = bitcast <2 x bfloat> [[TMP0]] to <2 x half>
----------------
arsenm wrote:

This bitcast also doesn't look right. I'm shocked that the vector cast behavior seems to treat FP-to-int vectors as bitcast, radically different from the scalar case (which OpenCL doesn't even allow). 

The comment says it's allowing bitcast between fp/int of the same size, but that's not really what the cast is here. 

https://github.com/llvm/llvm-project/pull/89051