[Clang][XTHeadVector] Implement 13.5 vnclip/vnclipu (ruyisdk#90)

* [Clang][XTHeadVector] Implement 13.5 `vnclip/vnclipu` * [Clang][XTHeadVector] Implement 13.5 `vnclip/vnclipu` wrapper macros
imkiva · Feb 26, 2025 · 8f6ffa4 · 8f6ffa4
1 parent 18f8400
commit 8f6ffa4
Show file tree

Hide file tree

Showing 4 changed files with 866 additions and 0 deletions.
diff --git a/clang/include/clang/Basic/riscv_vector_xtheadv.td b/clang/include/clang/Basic/riscv_vector_xtheadv.td
@@ -1041,5 +1041,53 @@ let UnMaskedPolicyScheme = HasPassthruOperand,
   defm th_vsmul : RVVSignedBinBuiltinSetRoundingMode;
 }
 
+// 13.5 Vector Narrowing Fixed-Point Clip Operations
+let UnMaskedPolicyScheme = HasPassthruOperand,
+        MaskedPolicyScheme = HasPassthruOperand,
+        ManualCodegen = [{
+  {
+    // LLVM intrinsic
+    // Unmasked: (passthru, op0, op1, vxrm, vl)
+    // Masked:   (passthru, vector_in, vector_in/scalar_in, mask, vxrm, vl)
+
+    // Clang builtin
+    // Unmask:   (op1, op2, vxrm, vl)
+    // Mask:     (mask, op1, op2, vxrm, vl)
+
+    SmallVector<llvm::Value*, 7> Operands;
+    bool HasMaskedOff = !(
+        (IsMasked && (PolicyAttrs & RVV_VTA) && (PolicyAttrs & RVV_VMA)) ||
+        (!IsMasked && PolicyAttrs & RVV_VTA));
+    unsigned Offset = IsMasked ?
+        (HasMaskedOff ? 2 : 1) : (HasMaskedOff ? 1 : 0);
+
+    if (!HasMaskedOff)
+      Operands.push_back(llvm::PoisonValue::get(ResultType));
+    else
+      Operands.push_back(Ops[IsMasked ? 1 : 0]);
+
+    Operands.push_back(Ops[Offset]); // op0
+    Operands.push_back(Ops[Offset + 1]); // op1
+
+    if (IsMasked)
+      Operands.push_back(Ops[0]); // mask
+
+    Operands.push_back(Ops[Offset + 2]); // vxrm
+    Operands.push_back(Ops[Offset + 3]); // vl
+
+    // TODO: no policy in LLVM side for masked intrinsics.
+    // if (IsMasked)
+    //  Operands.push_back(ConstantInt::get(Ops.back()->getType(), PolicyAttrs));
+
+    IntrinsicTypes = {ResultType, Ops[Offset]->getType(), Ops[Offset + 1]->getType(),
+                      Ops.back()->getType()};
+    llvm::Function *F = CGM.getIntrinsic(ID, IntrinsicTypes);
+    return Builder.CreateCall(F, Operands, "");
+  }
+}] in {
+  // 13.5. Vector Narrowing Fixed-Point Clip Instructions
+  defm th_vnclipu : RVVUnsignedNShiftBuiltinSetRoundingMode;
+  defm th_vnclip  : RVVSignedNShiftBuiltinSetRoundingMode;
+}
 
 include "riscv_vector_xtheadv_wrappers.td"
diff --git a/clang/include/clang/Basic/riscv_vector_xtheadv_wrappers.td b/clang/include/clang/Basic/riscv_vector_xtheadv_wrappers.td
@@ -2189,3 +2189,87 @@ let HeaderCode =
 
 }] in
 def th_single_width_fractional_multiply_with_rounding_and_saturation_wrapper_macros: RVVHeader;
+
+// 13.5. Vector Narrowing Fixed-Point Clip Operations
+
+let HeaderCode =
+[{
+// Vector Narrowing Fixed-Point Clip Operations
+#define __riscv_vnclip_wv_i8m1(op1, op2, rm, vl) __riscv_th_vnclip_wv_i8m1(op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i8m1(op1, op2, rm, vl) __riscv_th_vnclip_wx_i8m1(op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i8m2(op1, op2, rm, vl) __riscv_th_vnclip_wv_i8m2(op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i8m2(op1, op2, rm, vl) __riscv_th_vnclip_wx_i8m2(op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i8m4(op1, op2, rm, vl) __riscv_th_vnclip_wv_i8m4(op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i8m4(op1, op2, rm, vl) __riscv_th_vnclip_wx_i8m4(op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i16m1(op1, op2, rm, vl) __riscv_th_vnclip_wv_i16m1(op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i16m1(op1, op2, rm, vl) __riscv_th_vnclip_wx_i16m1(op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i16m2(op1, op2, rm, vl) __riscv_th_vnclip_wv_i16m2(op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i16m2(op1, op2, rm, vl) __riscv_th_vnclip_wx_i16m2(op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i16m4(op1, op2, rm, vl) __riscv_th_vnclip_wv_i16m4(op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i16m4(op1, op2, rm, vl) __riscv_th_vnclip_wx_i16m4(op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i32m1(op1, op2, rm, vl) __riscv_th_vnclip_wv_i32m1(op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i32m1(op1, op2, rm, vl) __riscv_th_vnclip_wx_i32m1(op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i32m2(op1, op2, rm, vl) __riscv_th_vnclip_wv_i32m2(op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i32m2(op1, op2, rm, vl) __riscv_th_vnclip_wx_i32m2(op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i32m4(op1, op2, rm, vl) __riscv_th_vnclip_wv_i32m4(op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i32m4(op1, op2, rm, vl) __riscv_th_vnclip_wx_i32m4(op1, op2, rm, vl)
+
+#define __riscv_vnclip_wv_i8m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wv_i8m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i8m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wx_i8m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i8m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wv_i8m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i8m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wx_i8m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i8m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wv_i8m4_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i8m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wx_i8m4_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i16m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wv_i16m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i16m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wx_i16m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i16m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wv_i16m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i16m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wx_i16m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i16m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wv_i16m4_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i16m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wx_i16m4_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i32m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wv_i32m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i32m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wx_i32m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i32m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wv_i32m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i32m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wx_i32m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wv_i32m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wv_i32m4_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclip_wx_i32m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclip_wx_i32m4_m(mask, op1, op2, rm, vl)
+
+#define __riscv_vnclipu_wv_u8m1(op1, op2, rm, vl) __riscv_th_vnclipu_wv_u8m1(op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u8m1(op1, op2, rm, vl) __riscv_th_vnclipu_wx_u8m1(op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u8m2(op1, op2, rm, vl) __riscv_th_vnclipu_wv_u8m2(op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u8m2(op1, op2, rm, vl) __riscv_th_vnclipu_wx_u8m2(op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u8m4(op1, op2, rm, vl) __riscv_th_vnclipu_wv_u8m4(op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u8m4(op1, op2, rm, vl) __riscv_th_vnclipu_wx_u8m4(op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u16m1(op1, op2, rm, vl) __riscv_th_vnclipu_wv_u16m1(op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u16m1(op1, op2, rm, vl) __riscv_th_vnclipu_wx_u16m1(op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u16m2(op1, op2, rm, vl) __riscv_th_vnclipu_wv_u16m2(op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u16m2(op1, op2, rm, vl) __riscv_th_vnclipu_wx_u16m2(op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u16m4(op1, op2, rm, vl) __riscv_th_vnclipu_wv_u16m4(op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u16m4(op1, op2, rm, vl) __riscv_th_vnclipu_wx_u16m4(op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u32m1(op1, op2, rm, vl) __riscv_th_vnclipu_wv_u32m1(op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u32m1(op1, op2, rm, vl) __riscv_th_vnclipu_wx_u32m1(op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u32m2(op1, op2, rm, vl) __riscv_th_vnclipu_wv_u32m2(op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u32m2(op1, op2, rm, vl) __riscv_th_vnclipu_wx_u32m2(op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u32m4(op1, op2, rm, vl) __riscv_th_vnclipu_wv_u32m4(op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u32m4(op1, op2, rm, vl) __riscv_th_vnclipu_wx_u32m4(op1, op2, rm, vl)
+
+#define __riscv_vnclipu_wv_u8m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wv_u8m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u8m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wx_u8m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u8m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wv_u8m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u8m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wx_u8m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u8m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wv_u8m4_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u8m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wx_u8m4_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u16m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wv_u16m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u16m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wx_u16m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u16m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wv_u16m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u16m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wx_u16m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u16m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wv_u16m4_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u16m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wx_u16m4_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u32m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wv_u32m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u32m1_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wx_u32m1_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u32m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wv_u32m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u32m2_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wx_u32m2_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wv_u32m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wv_u32m4_m(mask, op1, op2, rm, vl)
+#define __riscv_vnclipu_wx_u32m4_m(mask, op1, op2, rm, vl) __riscv_th_vnclipu_wx_u32m4_m(mask, op1, op2, rm, vl)
+
+}] in
+def th_narrowing_width_fixed_point_clip_wrapper_macros: RVVHeader;