[AArch64] Implement NEON FP8 intrinsics for fused multiply-add (indexed)

This patch adds the following intrinsics:

* Floating-point multiply-add long to half-precision (vector, by element)
    float16x8_t vmlalbq_lane_f16_mf8_fpm(float16x8_t vd, mfloat8x16_t vn, mfloat8x8_t vm, __builtin_constant_p(lane), fpm_t fpm)
    float16x8_t vmlalbq_laneq_f16_mf8_fpm(float16x8_t vd, mfloat8x16_t vn, mfloat8x16_t vm, __builtin_constant_p(lane), fpm_t fpm)
    float16x8_t vmlaltq_lane_f16_mf8_fpm(float16x8_t vd, mfloat8x16_t vn, mfloat8x8_t vm, __builtin_constant_p(lane), fpm_t fpm)
    float16x8_t vmlaltq_laneq_f16_mf8_fpm(float16x8_t vd, mfloat8x16_t vn, mfloat8x16_t vm, __builtin_constant_p(lane), fpm_t fpm)

* Floating-point multiply-add long-long to single-precision (vector, by element)

    float32x4_t vmlallbbq_lane_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x8_t vm, __builtin_constant_p(lane), fpm_t fpm)
    float32x4_t vmlallbbq_laneq_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x16_t vm, __builtin_constant_p(lane), fpm_t fpm)
    float32x4_t vmlallbtq_lane_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x8_t vm, __builtin_constant_p(lane), fpm_t fpm)
    float32x4_t vmlallbtq_laneq_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x16_t vm, __builtin_constant_p(lane), fpm_t fpm)
    float32x4_t vmlalltbq_lane_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x8_t vm, __builtin_constant_p(lane), fpm_t fpm)
    float32x4_t vmlalltbq_laneq_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x16_t vm, __builtin_constant_p(lane), fpm_t fpm)
    float32x4_t vmlallttq_lane_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x8_t vm, __builtin_constant_p(lane), fpm_t fpm)
    float32x4_t vmlallttq_laneq_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x16_t vm, __builtin_constant_p(lane), fpm_t fpm)

Author: momchil-velikov

clang/include/clang/Basic/arm_neon.td

@@ -2179,6 +2179,20 @@ let ArchGuard = "defined(__aarch64__)", TargetGuard = "fp8fma,neon" in {
   def VMLALLBT_F32_F8 : VInst<"vmlallbtq_f32_mf8_fpm", "(>>F)(>>F)..V", "Qm">;
   def VMLALLTB_F32_F8 : VInst<"vmlalltbq_f32_mf8_fpm", "(>>F)(>>F)..V", "Qm">;
   def VMLALLTT_F32_F8 : VInst<"vmlallttq_f32_mf8_fpm", "(>>F)(>>F)..V", "Qm">;
+
+  def VMLALB_F16_F8_LANE  : VInst<"vmlalbq_lane_f16_mf8_fpm",  "(>F)(>F).qIV", "Qm", [ImmCheck<3, ImmCheck0_7,  0>]>;
+  def VMLALB_F16_F8_LANEQ : VInst<"vmlalbq_laneq_f16_mf8_fpm", "(>F)(>F)..IV", "Qm", [ImmCheck<3, ImmCheck0_15, 0>]>;
+  def VMLALT_F16_F8_LANE  : VInst<"vmlaltq_lane_f16_mf8_fpm",  "(>F)(>F).qIV", "Qm", [ImmCheck<3, ImmCheck0_7,  0>]>;
+  def VMLALT_F16_F8_LANEQ : VInst<"vmlaltq_laneq_f16_mf8_fpm", "(>F)(>F)..IV", "Qm", [ImmCheck<3, ImmCheck0_15, 0>]>;
+
+  def VMLALLBB_F32_F8_LANE  : VInst<"vmlallbbq_lane_f32_mf8_fpm",  "(>>F)(>>F).qIV", "Qm", [ImmCheck<3, ImmCheck0_7,  0>]>;
+  def VMLALLBB_F32_F8_LANEQ : VInst<"vmlallbbq_laneq_f32_mf8_fpm", "(>>F)(>>F)..IV", "Qm", [ImmCheck<3, ImmCheck0_15, 0>]>;
+  def VMLALLBT_F32_F8_LANE  : VInst<"vmlallbtq_lane_f32_mf8_fpm",  "(>>F)(>>F).qIV", "Qm", [ImmCheck<3, ImmCheck0_7,  0>]>;
+  def VMLALLBT_F32_F8_LANEQ : VInst<"vmlallbtq_laneq_f32_mf8_fpm", "(>>F)(>>F)..IV", "Qm", [ImmCheck<3, ImmCheck0_15, 0>]>;
+  def VMLALLTB_F32_F8_LANE  : VInst<"vmlalltbq_lane_f32_mf8_fpm",  "(>>F)(>>F).qIV", "Qm", [ImmCheck<3, ImmCheck0_7,  0>]>;
+  def VMLALLTB_F32_F8_LANEQ : VInst<"vmlalltbq_laneq_f32_mf8_fpm", "(>>F)(>>F)..IV", "Qm", [ImmCheck<3, ImmCheck0_15, 0>]>;
+  def VMLALLTT_F32_F8_LANE  : VInst<"vmlallttq_lane_f32_mf8_fpm",  "(>>F)(>>F).qIV", "Qm", [ImmCheck<3, ImmCheck0_7,  0>]>;
+  def VMLALLTT_F32_F8_LANEQ : VInst<"vmlallttq_laneq_f32_mf8_fpm", "(>>F)(>>F)..IV", "Qm", [ImmCheck<3, ImmCheck0_15, 0>]>;
 }
 
 let ArchGuard = "defined(__aarch64__)", TargetGuard = "neon,faminmax" in {

clang/lib/CodeGen/CGBuiltin.cpp

@@ -6880,21 +6880,36 @@ Value *CodeGenFunction::EmitFP8NeonCall(unsigned IID,
 }
 
 llvm::Value *CodeGenFunction::EmitFP8NeonFDOTCall(
-    unsigned IID, bool ExtendLane, llvm::Type *RetTy,
+    unsigned IID, bool ExtendLaneArg, llvm::Type *RetTy,
     SmallVectorImpl<llvm::Value *> &Ops, const CallExpr *E, const char *name) {
 
   const unsigned ElemCount = Ops[0]->getType()->getPrimitiveSizeInBits() /
                              RetTy->getPrimitiveSizeInBits();
   llvm::Type *Tys[] = {llvm::FixedVectorType::get(RetTy, ElemCount),
                        Ops[1]->getType()};
-  if (ExtendLane) {
+  if (ExtendLaneArg) {
     auto *VT = llvm::FixedVectorType::get(Int8Ty, 16);
     Ops[2] = Builder.CreateInsertVector(VT, PoisonValue::get(VT), Ops[2],
                                         Builder.getInt64(0));
   }
   return EmitFP8NeonCall(IID, Tys, Ops, E, name);
 }
 
+llvm::Value *CodeGenFunction::EmitFP8NeonFMLACall(
+    unsigned IID, bool ExtendLaneArg, llvm::Type *RetTy,
+    SmallVectorImpl<llvm::Value *> &Ops, const CallExpr *E, const char *name) {
+
+  if (ExtendLaneArg) {
+    auto *VT = llvm::FixedVectorType::get(Int8Ty, 16);
+    Ops[2] = Builder.CreateInsertVector(VT, PoisonValue::get(VT), Ops[2],
+                                        Builder.getInt64(0));
+  }
+  const unsigned ElemCount = Ops[0]->getType()->getPrimitiveSizeInBits() /
+                             RetTy->getPrimitiveSizeInBits();
+  return EmitFP8NeonCall(IID, {llvm::FixedVectorType::get(RetTy, ElemCount)},
+                         Ops, E, name);
+}
+
 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty,
                                             bool neg) {
   int SV = cast<ConstantInt>(V)->getSExtValue();
@@ -12840,7 +12855,7 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
 
   unsigned Int;
   bool ExtractLow = false;
-  bool ExtendLane = false;
+  bool ExtendLaneArg = false;
   switch (BuiltinID) {
   default: return nullptr;
   case NEON::BI__builtin_neon_vbsl_v:
@@ -14115,24 +14130,24 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
                                Ops, E, "fdot2");
   case NEON::BI__builtin_neon_vdot_lane_f16_mf8_fpm:
   case NEON::BI__builtin_neon_vdotq_lane_f16_mf8_fpm:
-    ExtendLane = true;
+    ExtendLaneArg = true;
     LLVM_FALLTHROUGH;
   case NEON::BI__builtin_neon_vdot_laneq_f16_mf8_fpm:
   case NEON::BI__builtin_neon_vdotq_laneq_f16_mf8_fpm:
     return EmitFP8NeonFDOTCall(Intrinsic::aarch64_neon_fp8_fdot2_lane,
-                               ExtendLane, HalfTy, Ops, E, "fdot2_lane");
+                               ExtendLaneArg, HalfTy, Ops, E, "fdot2_lane");
   case NEON::BI__builtin_neon_vdot_f32_mf8_fpm:
   case NEON::BI__builtin_neon_vdotq_f32_mf8_fpm:
     return EmitFP8NeonFDOTCall(Intrinsic::aarch64_neon_fp8_fdot4, false,
                                FloatTy, Ops, E, "fdot4");
   case NEON::BI__builtin_neon_vdot_lane_f32_mf8_fpm:
   case NEON::BI__builtin_neon_vdotq_lane_f32_mf8_fpm:
-    ExtendLane = true;
+    ExtendLaneArg = true;
     LLVM_FALLTHROUGH;
   case NEON::BI__builtin_neon_vdot_laneq_f32_mf8_fpm:
   case NEON::BI__builtin_neon_vdotq_laneq_f32_mf8_fpm:
     return EmitFP8NeonFDOTCall(Intrinsic::aarch64_neon_fp8_fdot4_lane,
-                               ExtendLane, FloatTy, Ops, E, "fdot4_lane");
+                               ExtendLaneArg, FloatTy, Ops, E, "fdot4_lane");
 
   case NEON::BI__builtin_neon_vmlalbq_f16_mf8_fpm:
     return EmitFP8NeonCall(Intrinsic::aarch64_neon_fp8_fmlalb,
@@ -14158,7 +14173,42 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
     return EmitFP8NeonCall(Intrinsic::aarch64_neon_fp8_fmlalltt,
                            {llvm::FixedVectorType::get(FloatTy, 4)}, Ops, E,
                            "vmlall");
-
+  case NEON::BI__builtin_neon_vmlalbq_lane_f16_mf8_fpm:
+    ExtendLaneArg = true;
+    LLVM_FALLTHROUGH;
+  case NEON::BI__builtin_neon_vmlalbq_laneq_f16_mf8_fpm:
+    return EmitFP8NeonFMLACall(Intrinsic::aarch64_neon_fp8_fmlalb_lane,
+                               ExtendLaneArg, HalfTy, Ops, E, "vmlal_lane");
+  case NEON::BI__builtin_neon_vmlaltq_lane_f16_mf8_fpm:
+    ExtendLaneArg = true;
+    LLVM_FALLTHROUGH;
+  case NEON::BI__builtin_neon_vmlaltq_laneq_f16_mf8_fpm:
+    return EmitFP8NeonFMLACall(Intrinsic::aarch64_neon_fp8_fmlalt_lane,
+                               ExtendLaneArg, HalfTy, Ops, E, "vmlal_lane");
+  case NEON::BI__builtin_neon_vmlallbbq_lane_f32_mf8_fpm:
+    ExtendLaneArg = true;
+    LLVM_FALLTHROUGH;
+  case NEON::BI__builtin_neon_vmlallbbq_laneq_f32_mf8_fpm:
+    return EmitFP8NeonFMLACall(Intrinsic::aarch64_neon_fp8_fmlallbb_lane,
+                               ExtendLaneArg, FloatTy, Ops, E, "vmlall_lane");
+  case NEON::BI__builtin_neon_vmlallbtq_lane_f32_mf8_fpm:
+    ExtendLaneArg = true;
+    LLVM_FALLTHROUGH;
+  case NEON::BI__builtin_neon_vmlallbtq_laneq_f32_mf8_fpm:
+    return EmitFP8NeonFMLACall(Intrinsic::aarch64_neon_fp8_fmlallbt_lane,
+                               ExtendLaneArg, FloatTy, Ops, E, "vmlall_lane");
+  case NEON::BI__builtin_neon_vmlalltbq_lane_f32_mf8_fpm:
+    ExtendLaneArg = true;
+    LLVM_FALLTHROUGH;
+  case NEON::BI__builtin_neon_vmlalltbq_laneq_f32_mf8_fpm:
+    return EmitFP8NeonFMLACall(Intrinsic::aarch64_neon_fp8_fmlalltb_lane,
+                               ExtendLaneArg, FloatTy, Ops, E, "vmlall_lane");
+  case NEON::BI__builtin_neon_vmlallttq_lane_f32_mf8_fpm:
+    ExtendLaneArg = true;
+    LLVM_FALLTHROUGH;
+  case NEON::BI__builtin_neon_vmlallttq_laneq_f32_mf8_fpm:
+    return EmitFP8NeonFMLACall(Intrinsic::aarch64_neon_fp8_fmlalltt_lane,
+                               ExtendLaneArg, FloatTy, Ops, E, "vmlall_lane");
   case NEON::BI__builtin_neon_vamin_f16:
   case NEON::BI__builtin_neon_vaminq_f16:
   case NEON::BI__builtin_neon_vamin_f32:

clang/lib/CodeGen/CodeGenFunction.h

@@ -4655,7 +4655,11 @@ class CodeGenFunction : public CodeGenTypeCache {
                                   llvm::Type *Ty1, bool Extract,
                                   SmallVectorImpl<llvm::Value *> &Ops,
                                   const CallExpr *E, const char *name);
-  llvm::Value *EmitFP8NeonFDOTCall(unsigned IID, bool ExtendLane,
+  llvm::Value *EmitFP8NeonFDOTCall(unsigned IID, bool ExtendLaneArg,
+                                   llvm::Type *RetTy,
+                                   SmallVectorImpl<llvm::Value *> &Ops,
+                                   const CallExpr *E, const char *name);
+  llvm::Value *EmitFP8NeonFMLACall(unsigned IID, bool ExtendLaneArg,
                                    llvm::Type *RetTy,
                                    SmallVectorImpl<llvm::Value *> &Ops,
                                    const CallExpr *E, const char *name);