[clang] [CIR][X86] Implement convert_half builtin (PR #171615)

[via cfe-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-clang

Author: Priyanshu Kumar (Priyanshu3820)

<details>
<summary>Changes</summary>

related to #<!-- -->167765 

Calls the LLVM intrinsic for the following builtins-

`BI__builtin_ia32_vcvtph2ps_mask`
`BI__builtin_ia32_vcvtph2ps256_mask`
`BI__builtin_ia32_vcvtph2ps512_mask`
`BI__builtin_ia32_cvtneps2bf16_128_mask`
`BI__builtin_ia32_cvtneps2bf16_256_mask`
`BI__builtin_ia32_cvtneps2bf16_512_mask`

---
Full diff: https://github.com/llvm/llvm-project/pull/171615.diff


2 Files Affected:

- (modified) clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp (+34-6) 
- (added) clang/test/CIR/CodeGen/X86/cir-convert-half.c (+55) 


``````````diff
diff --git a/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp b/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp
index fb17e31bf36d6..75022d4f93d4a 100644
--- a/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp
@@ -1514,12 +1514,40 @@ mlir::Value CIRGenFunction::emitX86BuiltinExpr(unsigned builtinID,
   case X86::BI__builtin_ia32_cmpnltsd:
   case X86::BI__builtin_ia32_cmpnlesd:
   case X86::BI__builtin_ia32_cmpordsd:
-  case X86::BI__builtin_ia32_vcvtph2ps_mask:
-  case X86::BI__builtin_ia32_vcvtph2ps256_mask:
-  case X86::BI__builtin_ia32_vcvtph2ps512_mask:
-  case X86::BI__builtin_ia32_cvtneps2bf16_128_mask:
-  case X86::BI__builtin_ia32_cvtneps2bf16_256_mask:
-  case X86::BI__builtin_ia32_cvtneps2bf16_512_mask:
+    cgm.errorNYI(expr->getSourceRange(),
+                 std::string("unimplemented X86 builtin call: ") +
+                     getContext().BuiltinInfo.getName(builtinID));
+    return {};
+  case X86::BI__builtin_ia32_vcvtph2ps_mask: {
+    mlir::Location loc = getLoc(expr->getExprLoc());
+    return emitIntrinsicCallOp(builder, loc, "x86.avx512.mask.vcvtph2ps.128",
+                               convertType(expr->getType()), ops);
+  }
+  case X86::BI__builtin_ia32_vcvtph2ps256_mask: {
+    mlir::Location loc = getLoc(expr->getExprLoc());
+    return emitIntrinsicCallOp(builder, loc, "x86.avx512.mask.vcvtph2ps.256",
+                               convertType(expr->getType()), ops);
+  }
+  case X86::BI__builtin_ia32_vcvtph2ps512_mask: {
+    mlir::Location loc = getLoc(expr->getExprLoc());
+    return emitIntrinsicCallOp(builder, loc, "x86.avx512.mask.vcvtph2ps.512",
+                               convertType(expr->getType()), ops);
+  }
+  case X86::BI__builtin_ia32_cvtneps2bf16_128_mask: {
+    mlir::Location loc = getLoc(expr->getExprLoc());
+    return emitIntrinsicCallOp(builder, loc, "x86.avx512bf16.mask.cvtneps2bf16.128",
+                               convertType(expr->getType()), ops);
+  }
+  case X86::BI__builtin_ia32_cvtneps2bf16_256_mask: {
+    mlir::Location loc = getLoc(expr->getExprLoc());
+    return emitIntrinsicCallOp(builder, loc, "x86.avx512bf16.cvtneps2bf16.256",
+                               convertType(expr->getType()), ops);
+  }
+  case X86::BI__builtin_ia32_cvtneps2bf16_512_mask: {
+    mlir::Location loc = getLoc(expr->getExprLoc());
+    return emitIntrinsicCallOp(builder, loc, "x86.avx512bf16.cvtneps2bf16.512",
+                               convertType(expr->getType()), ops);
+  }
   case X86::BI__cpuid:
   case X86::BI__cpuidex:
   case X86::BI__emul:
diff --git a/clang/test/CIR/CodeGen/X86/cir-convert-half.c b/clang/test/CIR/CodeGen/X86/cir-convert-half.c
new file mode 100644
index 0000000000000..4fad9aa02cfc1
--- /dev/null
+++ b/clang/test/CIR/CodeGen/X86/cir-convert-half.c
@@ -0,0 +1,55 @@
+// Test X86-specific convert_half builtins (4-argument form)
+
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -target-feature +avx512fp16 -target-feature +avx512bf16 -fclangir -emit-cir %s -o - | FileCheck --check-prefix=CIR %s
+
+typedef float __m512 __attribute__((__vector_size__(64), __aligned__(64)));
+typedef float __m256 __attribute__((__vector_size__(32), __aligned__(32)));
+typedef float __m128 __attribute__((__vector_size__(16), __aligned__(16)));
+typedef int __m256i __attribute__((__vector_size__(32), __aligned__(32)));
+typedef int __m128i __attribute__((__vector_size__(16), __aligned__(16)));
+typedef int __mmask16;
+typedef unsigned char __mmask8;
+typedef __bf16 __m256bh __attribute__((__vector_size__(32), __aligned__(32)));
+typedef __bf16 __m128bh __attribute__((__vector_size__(16), __aligned__(16)));
+
+// Test for __builtin_ia32_vcvtph2ps512_mask
+__m512 test_vcvtph2ps512_mask(__m256i a, __m512 src, __mmask16 k, __m512 passthru) {
+  return __builtin_ia32_vcvtph2ps512_mask(a, src, k, passthru);
+}
+// CIR-LABEL: cir.func {{.*}}@test_vcvtph2ps512_mask
+// CIR: cir.call @llvm.x86.avx512.mask.vcvtph2ps.512
+
+// Test for __builtin_ia32_vcvtph2ps256_mask
+__m256 test_vcvtph2ps256_mask(__m128i a, __m256 src, __mmask8 k, __m256 passthru) {
+  return __builtin_ia32_vcvtph2ps256_mask(a, src, k, passthru);
+}
+// CIR-LABEL: cir.func {{.*}}@test_vcvtph2ps256_mask
+// CIR: cir.call @llvm.x86.avx512.mask.vcvtph2ps.256
+
+// Test for __builtin_ia32_vcvtph2ps_mask
+__m128 test_vcvtph2ps_mask(__m128i a, __m128 src, __mmask8 k, __m128 passthru) {
+  return __builtin_ia32_vcvtph2ps_mask(a, src, k, passthru);
+}
+// CIR-LABEL: cir.func {{.*}}@test_vcvtph2ps_mask
+// CIR: cir.call @llvm.x86.avx512.mask.vcvtph2ps.128
+
+// Test for __builtin_ia32_cvtneps2bf16_512_mask
+__m256bh test_cvtneps2bf16_512_mask(__m512 a, __m256bh w, __mmask16 u, __m256bh passthru) {
+  return __builtin_ia32_cvtneps2bf16_512_mask(a, w, u, passthru);
+}
+// CIR-LABEL: cir.func {{.*}}@test_cvtneps2bf16_512_mask
+// CIR: cir.call @llvm.x86.avx512bf16.cvtneps2bf16.512
+
+// Test for __builtin_ia32_cvtneps2bf16_256_mask
+__m128bh test_cvtneps2bf16_256_mask(__m256 a, __m128bh w, __mmask8 u, __m128bh passthru) {
+  return __builtin_ia32_cvtneps2bf16_256_mask(a, w, u, passthru);
+}
+// CIR-LABEL: cir.func {{.*}}@test_cvtneps2bf16_256_mask
+// CIR: cir.call @llvm.x86.avx512bf16.cvtneps2bf16.256
+
+// Test for __builtin_ia32_cvtneps2bf16_128_mask
+__m128bh test_cvtneps2bf16_128_mask(__m128 a, __m128bh w, __mmask8 u, __m128bh passthru) {
+  return __builtin_ia32_cvtneps2bf16_128_mask(a, w, u, passthru);
+}
+// CIR-LABEL: cir.func {{.*}}@test_cvtneps2bf16_128_mask
+// CIR: cir.call @llvm.x86.avx512bf16.mask.cvtneps2bf16.128
\ No newline at end of file

``````````

</details>


https://github.com/llvm/llvm-project/pull/171615