NEON support for _lane ops, and multiplies by scalar.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@105769 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/utils/TableGen/NeonEmitter.cpp b/utils/TableGen/NeonEmitter.cpp
index 4443376..1c5794c 100644
--- a/utils/TableGen/NeonEmitter.cpp
+++ b/utils/TableGen/NeonEmitter.cpp
@@ -137,6 +137,7 @@
usgn = true;
break;
case 's':
+ case 'a':
scal = true;
break;
case 'k':
@@ -442,14 +443,7 @@
return s;
}
-// Generate the definition for this intrinsic, e.g. "a + b" for OpAdd.
-// If structTypes is true, the NEON types are structs of vector types rather
-// than vector types, and the call becomes "a.val + b.val"
-static std::string GenOpString(OpKind op, const std::string &proto,
- StringRef typestr, bool structTypes = true) {
- std::string ts = TypeString(proto[0], typestr);
- std::string s = ts + " r; r";
-
+static std::string Duplicate(StringRef typestr, const std::string &a) {
bool dummy, quad = false;
char type = ClassifyType(typestr, quad, dummy, dummy);
unsigned nElts = 0;
@@ -462,6 +456,27 @@
case 'f': nElts = 2; break;
}
nElts <<= quad;
+
+ std::string s;
+
+ s = "(__neon_" + TypeString('d', typestr) + "){ ";
+ for (unsigned i = 0; i != nElts; ++i) {
+ s += a;
+ if ((i + 1) < nElts)
+ s += ", ";
+ }
+ s += " }";
+
+ return s;
+}
+
+// Generate the definition for this intrinsic, e.g. "a + b" for OpAdd.
+// If structTypes is true, the NEON types are structs of vector types rather
+// than vector types, and the call becomes "a.val + b.val"
+static std::string GenOpString(OpKind op, const std::string &proto,
+ StringRef typestr, bool structTypes = true) {
+ std::string ts = TypeString(proto[0], typestr);
+ std::string s = ts + " r; r";
if (structTypes)
s += ".val";
@@ -481,12 +496,18 @@
case OpSub:
s += a + " - " + b;
break;
+ case OpMulN:
+ b = Duplicate(typestr, "b");
case OpMul:
s += a + " * " + b;
break;
+ case OpMlaN:
+ c = Duplicate(typestr, "c");
case OpMla:
s += a + " + ( " + b + " * " + c + " )";
break;
+ case OpMlsN:
+ c = Duplicate(typestr, "c");
case OpMls:
s += a + " - ( " + b + " * " + c + " )";
break;
@@ -540,13 +561,7 @@
s += "(__neon_int64x1_t)(((__neon_int64x2_t)" + a + ")[0])";
break;
case OpDup:
- s += "(__neon_" + ts + "){ ";
- for (unsigned i = 0; i != nElts; ++i) {
- s += a;
- if ((i + 1) < nElts)
- s += ", ";
- }
- s += " }";
+ s += Duplicate(typestr, a);
break;
default:
throw "unknown OpKind!";
@@ -647,10 +662,17 @@
s += " = ";
}
- }
+ }
+
+ bool splat = proto.find('a') != std::string::npos;
s += "__builtin_neon_";
- s += MangleName(name, typestr, ck);
+ if (splat) {
+ std::string vname(name, 0, name.size()-2);
+ s += MangleName(vname, typestr, ck);
+ } else {
+ s += MangleName(name, typestr, ck);
+ }
s += "(";
for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
@@ -672,12 +694,18 @@
// Parenthesize the args from the macro.
if (define)
s.push_back('(');
- s.push_back(arg);
+
+ if (splat && (i + 1) == e)
+ s += Duplicate(typestr, std::string(&arg, 1));
+ else
+ s.push_back(arg);
+
+ // Parenthesize the args from the macro.
if (define)
s.push_back(')');
if (structTypes && proto[i] != 's' && proto[i] != 'i' && proto[i] != 'l' &&
- proto[i] != 'p' && proto[i] != 'c') {
+ proto[i] != 'p' && proto[i] != 'c' && proto[i] != 'a') {
s += ".val";
}
if ((i + 1) < e)
@@ -748,7 +776,6 @@
// Emit NEON-specific scalar typedefs.
// FIXME: probably need to do something better for polynomial types.
- // FIXME: is this the correct thing to do for float16?
OS << "typedef float float32_t;\n";
OS << "typedef uint8_t poly8_t;\n";
OS << "typedef uint16_t poly16_t;\n";
@@ -869,6 +896,9 @@
std::string Proto = R->getValueAsString("Prototype");
std::string Types = R->getValueAsString("Types");
+ if (Proto.find('a') != std::string::npos)
+ continue;
+
SmallVector<StringRef, 16> TypeVec;
ParseTypes(R, Types, TypeVec);