From: Abderrazek Zaafrani <a.zaafrani@samsung.com>
Date: Tue, 20 Mar 2018 20:37:31 +0000 (+0000)
Subject: [AArch64] Add vmulxh_lane fp16 vector intrinsic
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=9901645365c4ce268ea03dd104e74ed6ec506e34;p=clang

[AArch64] Add vmulxh_lane fp16 vector intrinsic

https://reviews.llvm.org/D44591

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@328038 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/include/clang/Basic/arm_neon.td b/include/clang/Basic/arm_neon.td
index 37aac58a11..f4fe149782 100644
--- a/include/clang/Basic/arm_neon.td
+++ b/include/clang/Basic/arm_neon.td
@@ -1499,11 +1499,9 @@ let ArchGuard = "defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(__aarc
   def VMULX_LANEH   : IOpInst<"vmulx_lane", "ddgi", "hQh", OP_MULX_LN>;
   def VMULX_LANEQH  : IOpInst<"vmulx_laneq", "ddji", "hQh", OP_MULX_LN>;
   def VMULX_NH      : IOpInst<"vmulx_n", "dds", "hQh", OP_MULX_N>;
-  // TODO: Scalar floating point multiply extended (scalar, by element)
-  // Below ones are commented out because they need vmulx_f16(float16_t, float16_t)
-  // which will be implemented later with fp16 scalar intrinsic (arm_fp16.h)
-  //def SCALAR_FMULX_LANEH : IOpInst<"vmulx_lane", "ssdi", "Sh", OP_SCALAR_MUL_LN>;
-  //def SCALAR_FMULX_LANEQH : IOpInst<"vmulx_laneq", "ssji", "Sh", OP_SCALAR_MUL_LN>;
+  // Scalar floating point  mulx (scalar, by element)
+  def SCALAR_FMULX_LANEH : IInst<"vmulx_lane", "ssdi", "Sh">;
+  def SCALAR_FMULX_LANEQH : IInst<"vmulx_laneq", "ssji", "Sh">;
 
   // ARMv8.2-A FP16 reduction vector intrinsics.
   def VMAXVH   : SInst<"vmaxv", "sd", "hQh">;
diff --git a/lib/CodeGen/CGBuiltin.cpp b/lib/CodeGen/CGBuiltin.cpp
index 7c8f111d10..084bf50e1b 100644
--- a/lib/CodeGen/CGBuiltin.cpp
+++ b/lib/CodeGen/CGBuiltin.cpp
@@ -7238,6 +7238,16 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
     Int = Intrinsic::aarch64_neon_fmulx;
     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx");
   }
+  case NEON::BI__builtin_neon_vmulxh_lane_f16:
+  case NEON::BI__builtin_neon_vmulxh_laneq_f16: {
+    // vmulx_lane should be mapped to Neon scalar mulx after
+    // extracting the scalar element
+    Ops.push_back(EmitScalarExpr(E->getArg(2)));
+    Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract");
+    Ops.pop_back();
+    Int = Intrinsic::aarch64_neon_fmulx;
+    return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmulx");
+  }
   case NEON::BI__builtin_neon_vmul_lane_v:
   case NEON::BI__builtin_neon_vmul_laneq_v: {
     // v1f64 vmul_lane should be mapped to Neon scalar mul lane
diff --git a/test/CodeGen/aarch64-v8.2a-neon-intrinsics.c b/test/CodeGen/aarch64-v8.2a-neon-intrinsics.c
index 1a5c3a280f..c475816ad9 100644
--- a/test/CodeGen/aarch64-v8.2a-neon-intrinsics.c
+++ b/test/CodeGen/aarch64-v8.2a-neon-intrinsics.c
@@ -1223,27 +1223,25 @@ float16x8_t test_vmulxq_n_f16(float16x8_t a, float16_t b) {
   return vmulxq_n_f16(a, b);
 }
 
-/* TODO: Not implemented yet (needs scalar intrinsic from arm_fp16.h)
-// CCHECK-LABEL: test_vmulxh_lane_f16
-// CCHECK: [[CONV0:%.*]] = fpext half %a to float
-// CCHECK: [[CONV1:%.*]] = fpext half %{{.*}} to float
-// CCHECK: [[MUL:%.*]]   = fmul float [[CONV0:%.*]], [[CONV0:%.*]]
-// CCHECK: [[CONV3:%.*]] = fptrunc float %mul to half
-// CCHECK: ret half [[CONV3:%.*]]
+// CHECK-LABEL: test_vmulxh_lane_f16
+// CHECK: [[TMP0:%.*]] = bitcast <4 x half> %b to <8 x i8>
+// CHECK: [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x half>
+// CHECK: [[EXTR:%.*]] = extractelement <4 x half> [[TMP1]], i32 3
+// CHECK: [[MULX:%.*]] = call half @llvm.aarch64.neon.fmulx.f16(half %a, half [[EXTR]]
+// CHECK: ret half [[MULX]]
 float16_t test_vmulxh_lane_f16(float16_t a, float16x4_t b) {
   return vmulxh_lane_f16(a, b, 3);
 }
 
-// CCHECK-LABEL: test_vmulxh_laneq_f16
-// CCHECK: [[CONV0:%.*]] = fpext half %a to float
-// CCHECK: [[CONV1:%.*]] = fpext half %{{.*}} to float
-// CCHECK: [[MUL:%.*]]   = fmul float [[CONV0:%.*]], [[CONV0:%.*]]
-// CCHECK: [[CONV3:%.*]] = fptrunc float %mul to half
-// CCHECK: ret half [[CONV3:%.*]]
+// CHECK-LABEL: test_vmulxh_laneq_f16
+// CHECK: [[TMP0:%.*]] = bitcast <8 x half> %b to <16 x i8>
+// CHECK: [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x half>
+// CHECK: [[EXTR:%.*]] = extractelement <8 x half> [[TMP1]], i32 7
+// CHECK: [[MULX:%.*]] = call half @llvm.aarch64.neon.fmulx.f16(half %a, half [[EXTR]])
+// CHECK: ret half [[MULX]]
 float16_t test_vmulxh_laneq_f16(float16_t a, float16x8_t b) {
   return vmulxh_laneq_f16(a, b, 7);
 }
-*/
 
 // CHECK-LABEL: test_vmaxv_f16
 // CHECK: [[TMP0:%.*]] = bitcast <4 x half> %a to <8 x i8>