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

[Andy Kaylor via cfe-commits]

================
@@ -1602,12 +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: {
----------------
andykaylor wrote:

> I too considered this way and I take it that it's your implementation of the reduce intrinsics right? But I think we are supposed to follow the already established pattern aren't we? I would be more than happy to implement this way though.

In general, we are trying to follow the established code patterns, but if the code can be improved it's OK to diverge. I like the suggestion from @badumbatish here.

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