[CIR][CIRGen][Builtin][X86] lower stmx and ldmx intrinsics

Author: RiverDave

clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp

@@ -327,11 +327,24 @@ mlir::Value CIRGenFunction::emitX86BuiltinExpr(unsigned BuiltinID,
   }
   case X86::BI_mm_setcsr:
   case X86::BI__builtin_ia32_ldmxcsr: {
-    llvm_unreachable("mm_setcsr NYI");
+    Address tmp =
+        CreateMemTemp(E->getArg(0)->getType(), getLoc(E->getExprLoc()));
+    builder.createStore(getLoc(E->getExprLoc()), Ops[0], tmp);
+    return builder
+        .create<cir::LLVMIntrinsicCallOp>(
+            getLoc(E->getExprLoc()), builder.getStringAttr("x86.sse.ldmxcsr"),
+            builder.getVoidTy(), tmp.getPointer())
+        .getResult();
   }
   case X86::BI_mm_getcsr:
   case X86::BI__builtin_ia32_stmxcsr: {
-    llvm_unreachable("mm_getcsr NYI");
+    Address tmp = CreateMemTemp(E->getType(), getLoc(E->getExprLoc()));
+    builder
+        .create<cir::LLVMIntrinsicCallOp>(
+            getLoc(E->getExprLoc()), builder.getStringAttr("x86.sse.stmxcsr"),
+            builder.getVoidTy(), tmp.getPointer())
+        .getResult();
+    return builder.createLoad(getLoc(E->getExprLoc()), tmp);
   }
 
   case X86::BI__builtin_ia32_xsave:

clang/test/CIR/CodeGen/X86/sse-builtins.c

@@ -29,3 +29,26 @@ void test_mm_sfence(void) {
   // CIR: {{%.*}} = cir.llvm.intrinsic "x86.sse.sfence" : () -> !void
   // LLVM: call void @llvm.x86.sse.sfence()
 }
+
+void test_mm_setcsr(unsigned int A) {
+  // CIR-LABEL: test_mm_setcsr
+  // CIR: cir.store {{.*}}, {{.*}} : !u32i
+  // CIR: cir.llvm.intrinsic "x86.sse.ldmxcsr" {{.*}} : (!cir.ptr<!u32i>) -> !void
+
+  // LLVM-LABEL: test_mm_setcsr 
+  // LLVM: store i32
+  // LLVM: call void @llvm.x86.sse.ldmxcsr(ptr {{.*}})
+  _mm_setcsr(A);
+}
+
+unsigned int test_mm_getcsr(void) {
+  // CIR-LABEL: test_mm_getcsr
+  // CIR: cir.llvm.intrinsic "x86.sse.stmxcsr" %{{.*}} : (!cir.ptr<!u32i>) -> !void
+  // CIR: cir.load {{.*}} : !cir.ptr<!u32i>, !u32i
+
+  // LLVM-LABEL: test_mm_getcsr
+  // LLVM: call void @llvm.x86.sse.stmxcsr(ptr %{{.*}})
+  // LLVM: load i32
+  return _mm_getcsr();
+}
+