From fb7e000aa81db25ced208e290b6221f9e56f11c1 Mon Sep 17 00:00:00 2001
From: Tim Northover <tnorthover@apple.com>
Date: Tue, 25 Feb 2014 11:13:42 +0000
Subject: [PATCH] ARM NEON: add _f16 support to a couple of vector-shuffling
 intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@202137 91177308-0d34-0410-b5e6-96231b3b80d8
---
 include/clang/Basic/arm_neon.td    |  6 ++++--
 test/CodeGen/arm_neon_intrinsics.c | 24 ++++++++++++++++++++++++
 utils/TableGen/NeonEmitter.cpp     | 16 +++++++++++-----
 3 files changed, 39 insertions(+), 7 deletions(-)

diff --git a/include/clang/Basic/arm_neon.td b/include/clang/Basic/arm_neon.td
index eab3b50e38..b1ae1db0f6 100644
--- a/include/clang/Basic/arm_neon.td
+++ b/include/clang/Basic/arm_neon.td
@@ -410,9 +410,11 @@ def VCREATE : NoTestOpInst<"vcreate", "dl", "csihfUcUsUiUlPcPsl", OP_CAST>;
 // E.3.19 Set all lanes to same value
 let InstName = "vmov" in {
 def VDUP_N   : WOpInst<"vdup_n", "ds",
-                       "UcUsUicsiPcPsfQUcQUsQUiQcQsQiQPcQPsQflUlQlQUl", OP_DUP>;
+                       "UcUsUicsiPcPshfQUcQUsQUiQcQsQiQPcQPsQhQflUlQlQUl",
+                       OP_DUP>;
 def VMOV_N   : WOpInst<"vmov_n", "ds",
-                       "UcUsUicsiPcPsfQUcQUsQUiQcQsQiQPcQPsQflUlQlQUl", OP_DUP>;
+                       "UcUsUicsiPcPshfQUcQUsQUiQcQsQiQPcQPsQhQflUlQlQUl",
+                       OP_DUP>;
 }
 let InstName = "" in
 def VDUP_LANE: WOpInst<"vdup_lane", "dgi",
diff --git a/test/CodeGen/arm_neon_intrinsics.c b/test/CodeGen/arm_neon_intrinsics.c
index 1d76e8a575..d7473ed2bc 100644
--- a/test/CodeGen/arm_neon_intrinsics.c
+++ b/test/CodeGen/arm_neon_intrinsics.c
@@ -1887,6 +1887,12 @@ poly16x4_t test_vdup_n_p16(poly16_t a) {
   return vdup_n_p16(a);
 }
 
+// CHECK: test_vdup_n_f16
+// CHECK: vdup.16
+float16x4_t test_vdup_n_f16(float16_t *a) {
+  return vdup_n_f16(*a);
+}
+
 // CHECK: test_vdup_n_f32
 // CHECK: vmov 
 float32x2_t test_vdup_n_f32(float32_t a) {
@@ -1941,6 +1947,12 @@ poly16x8_t test_vdupq_n_p16(poly16_t a) {
   return vdupq_n_p16(a);
 }
 
+// CHECK: test_vdupq_n_f16
+// CHECK: vdup.16
+float16x8_t test_vdupq_n_f16(float16_t *a) {
+  return vdupq_n_f16(*a);
+}
+
 // CHECK: test_vdupq_n_f32
 // CHECK: vmov 
 float32x4_t test_vdupq_n_f32(float32_t a) {
@@ -4868,6 +4880,12 @@ poly16x4_t test_vmov_n_p16(poly16_t a) {
   return vmov_n_p16(a);
 }
 
+// CHECK: test_vmov_n_f16
+// CHECK: vdup.16
+float16x4_t test_vmov_n_f16(float16_t *a) {
+  return vmov_n_f16(*a);
+}
+
 // CHECK: test_vmov_n_f32
 // CHECK: vmov 
 float32x2_t test_vmov_n_f32(float32_t a) {
@@ -4922,6 +4940,12 @@ poly16x8_t test_vmovq_n_p16(poly16_t a) {
   return vmovq_n_p16(a);
 }
 
+// CHECK: test_vmovq_n_f16
+// CHECK: vdup.16
+float16x8_t test_vmovq_n_f16(float16_t *a) {
+  return vmovq_n_f16(*a);
+}
+
 // CHECK: test_vmovq_n_f32
 // CHECK: vmov 
 float32x4_t test_vmovq_n_f32(float32_t a) {
diff --git a/utils/TableGen/NeonEmitter.cpp b/utils/TableGen/NeonEmitter.cpp
index 932af9a8e8..165e749b33 100644
--- a/utils/TableGen/NeonEmitter.cpp
+++ b/utils/TableGen/NeonEmitter.cpp
@@ -1436,7 +1436,7 @@ static void GenerateChecksForIntrinsic(const std::string &Name,
 
 /// UseMacro - Examine the prototype string to determine if the intrinsic
 /// should be defined as a preprocessor macro instead of an inline function.
-static bool UseMacro(const std::string &proto) {
+static bool UseMacro(const std::string &proto, StringRef typestr) {
   // If this builtin takes an immediate argument, we need to #define it rather
   // than use a standard declaration, so that SemaChecking can range check
   // the immediate passed by the user.
@@ -1449,6 +1449,12 @@ static bool UseMacro(const std::string &proto) {
       proto.find('c') != std::string::npos)
     return true;
 
+  // It is not permitted to pass or return an __fp16 by value, so intrinsics
+  // taking a scalar float16_t must be implemented as macros.
+  if (typestr.find('h') != std::string::npos &&
+      proto.find('s') != std::string::npos)
+    return true;
+
   return false;
 }
 
@@ -1463,7 +1469,7 @@ static bool MacroArgUsedDirectly(const std::string &proto, unsigned i) {
 // Generate the string "(argtype a, argtype b, ...)"
 static std::string GenArgs(const std::string &proto, StringRef typestr,
                            const std::string &name) {
-  bool define = UseMacro(proto);
+  bool define = UseMacro(proto, typestr);
   char arg = 'a';
 
   std::string s;
@@ -1642,7 +1648,7 @@ static std::string GenOpString(const std::string &name, OpKind op,
                                const std::string &proto, StringRef typestr) {
   bool quad;
   unsigned nElts = GetNumElements(typestr, quad);
-  bool define = UseMacro(proto);
+  bool define = UseMacro(proto, typestr);
 
   std::string ts = TypeString(proto[0], typestr);
   std::string s;
@@ -2386,7 +2392,7 @@ static std::string GenBuiltin(const std::string &name, const std::string &proto,
   // sret-like argument.
   bool sret = IsMultiVecProto(proto[0]);
 
-  bool define = UseMacro(proto);
+  bool define = UseMacro(proto, typestr);
 
   // Check if the prototype has a scalar operand with the type of the vector
   // elements.  If not, bitcasting the args will take care of arg checking.
@@ -2531,7 +2537,7 @@ static std::string GenIntrinsic(const std::string &name,
                                 StringRef outTypeStr, StringRef inTypeStr,
                                 OpKind kind, ClassKind classKind) {
   assert(!proto.empty() && "");
-  bool define = UseMacro(proto) && kind != OpUnavailable;
+  bool define = UseMacro(proto, outTypeStr) && kind != OpUnavailable;
   std::string s;
 
   // static always inline + return type
-- 
2.40.0