[RISCV] Special case sign extended scalars when type legalizing nxvXi64 .vx instrinsics on RV32.

[craig.topper]

Summary

On RV32, we need to type legalize i64 scalar arguments to intrinsics.
We usually do this by splatting the value into a vector separately.
If the scalar happens to be sign extended, we can continue using a .vx
intrinsic.

We already special cased sign extended constants, this extends it
to any sign extended value.

I've only added tests for one case of vadd. Most intrinsics go
through the same check. I can add more tests if we're concerned.

[khchen]

Thanks, LGTM!

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

4625

nit: Need to update comment.

4751

nit: Maybe we remove this FIXME comment now?

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

 //===-- RISCVISelLowering.cpp - RISCV DAG Lowering Implementation  --------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
static SDValue lowerVectorIntrinsicScalars(SDValue Op, SelectionDAG &DAG,
   // element count from any operand or result and use it to construct the VT.
   assert(II->ScalarOperand > 0 && "Unexpected splat operand!");
   MVT VT = Op.getOperand(SplatOp - 1).getSimpleValueType();
 
   // The more complex case is when the scalar is larger than XLenVT.
   assert(XLenVT == MVT::i32 && OpVT == MVT::i64 &&
          VT.getVectorElementType() == MVT::i64 && "Unexpected VTs!");
 
   // If this is a sign-extended 32-bit constant, we can truncate it and rely
   // on the instruction to sign-extend since SEW>XLEN.
-  if (auto *CVal = dyn_cast<ConstantSDNode>(ScalarOp)) {
-    if (isInt<32>(CVal->getSExtValue())) {
-      ScalarOp = DAG.getConstant(CVal->getSExtValue(), DL, MVT::i32);
-      return DAG.getNode(Op->getOpcode(), DL, Op->getVTList(), Operands);
-    }
-  }
+  if (DAG.ComputeNumSignBits(ScalarOp) > 32) {
+    ScalarOp = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, ScalarOp);
 
   switch (IntNo) {
   case Intrinsic::riscv_vslide1up:
   case Intrinsic::riscv_vslide1down:
   case Intrinsic::riscv_vslide1up_mask:
   case Intrinsic::riscv_vslide1down_mask: {
     // We need to special case these when the scalar is larger than XLen.
     unsigned NumOps = Op.getNumOperands();
case Intrinsic::riscv_vslide1down_mask: {
     // TUMA or TUMU: Currently we always emit tumu policy regardless of tuma.
     // It's fine because vmerge does not care mask policy.
     return DAG.getNode(RISCVISD::VP_MERGE_VL, DL, VT, Mask, Vec, MaskedOff,
                        AVL);
   }
   }
 
   // We need to convert the scalar to a splat vector.
   // FIXME: Can we implicitly truncate the scalar if it is known to
   // be sign extended?
   SDValue VL = getVLOperand(Op);
   assert(VL.getValueType() == XLenVT);
   ScalarOp = splatSplitI64WithVL(DL, VT, SDValue(), ScalarOp, VL, DAG);
   return DAG.getNode(Op->getOpcode(), DL, Op->getVTList(), Operands);
 }
 
 SDValue RISCVTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,

llvm/test/CodeGen/RISCV/rvv/vadd.ll

%a = call <vscale x 1 x i64> @llvm.riscv.vadd.nxv1i64.i64(
     <vscale x 1 x i64> undef,
     <vscale x 1 x i64> %0,
     i64 %1,
     iXLen %2)
 
   ret <vscale x 1 x i64> %a
 }
 
+define <vscale x 1 x i64> @intrinsic_vadd_vx_sext_nxv1i64_nxv1i64_i64(<vscale x 1 x i64> %0, i32 %1, iXLen %2) nounwind {
+; RV32-LABEL: intrinsic_vadd_vx_sext_nxv1i64_nxv1i64_i64:
+; RV32:       # %bb.0: # %entry
+; RV32-NEXT:    vsetvli zero, a1, e64, m1, ta, mu
+; RV32-NEXT:    vadd.vx v8, v8, a0
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: intrinsic_vadd_vx_sext_nxv1i64_nxv1i64_i64:
+; RV64:       # %bb.0: # %entry
+; RV64-NEXT:    sext.w a0, a0
+; RV64-NEXT:    vsetvli zero, a1, e64, m1, ta, mu
+; RV64-NEXT:    vadd.vx v8, v8, a0
+; RV64-NEXT:    ret
+entry:
+  %ext = sext i32 %1 to i64
+  %a = call <vscale x 1 x i64> @llvm.riscv.vadd.nxv1i64.i64(
+    <vscale x 1 x i64> undef,
+    <vscale x 1 x i64> %0,
+    i64 %ext,
+    iXLen %2)
+
+  ret <vscale x 1 x i64> %a
+}
+
+define <vscale x 1 x i64> @intrinsic_vadd_vx_sextload_nxv1i64_nxv1i64_i64(<vscale x 1 x i64> %0, i32* %1, iXLen %2) nounwind {
+; CHECK-LABEL: intrinsic_vadd_vx_sextload_nxv1i64_nxv1i64_i64:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    lw a0, 0(a0)
+; CHECK-NEXT:    vsetvli zero, a1, e64, m1, ta, mu
+; CHECK-NEXT:    vadd.vx v8, v8, a0
+; CHECK-NEXT:    ret
+entry:
+  %load = load i32, i32* %1
+  %ext = sext i32 %load to i64
+  %a = call <vscale x 1 x i64> @llvm.riscv.vadd.nxv1i64.i64(
+    <vscale x 1 x i64> undef,
+    <vscale x 1 x i64> %0,
+    i64 %ext,
+    iXLen %2)
+
+  ret <vscale x 1 x i64> %a
+}
+
 declare <vscale x 1 x i64> @llvm.riscv.vadd.mask.nxv1i64.i64(
   <vscale x 1 x i64>,
   <vscale x 1 x i64>,
   i64,
   <vscale x 1 x i1>,
   iXLen, iXLen);
 
 define <vscale x 1 x i64> @intrinsic_vadd_mask_vx_nxv1i64_nxv1i64_i64(<vscale x 1 x i64> %0, <vscale x 1 x i64> %1, i64 %2, <vscale x 1 x i1> %3, iXLen %4) nounwind {