[RISCV] Fix incorrect extend type in vwmulsu combine.

[craig.topper]

Summary

If while matching widening multiply, we matched an extend from i8->i32,
i16->i64 or i8->i64 we need to reintroduce a narrower extend. If we're
matching a vwmulsu we need to use a sext for op0 and a zext for op1.

This bug exists in LLVM 14 and will need to be backported.

[benshi001]

LGTM

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

if (IsVWMULSU || Op0.getOpcode() == Op1.getOpcode()) {
     return SDValue();
 
   Op0 = Op0.getOperand(0);
 
   // Re-introduce narrower extends if needed.
   unsigned ExtOpc = IsSignExt ? RISCVISD::VSEXT_VL : RISCVISD::VZEXT_VL;
   if (Op0.getValueType() != NarrowVT)
     Op0 = DAG.getNode(ExtOpc, DL, NarrowVT, Op0, Mask, VL);
+  // vwmulsu requires second operand to be zero extended.
+  ExtOpc = IsVWMULSU ? RISCVISD::VZEXT_VL : ExtOpc;
   if (Op1.getValueType() != NarrowVT)
     Op1 = DAG.getNode(ExtOpc, DL, NarrowVT, Op1, Mask, VL);
 
   unsigned WMulOpc = RISCVISD::VWMULSU_VL;
   if (!IsVWMULSU)
     WMulOpc = IsSignExt ? RISCVISD::VWMUL_VL : RISCVISD::VWMULU_VL;
   return DAG.getNode(WMulOpc, DL, VT, Op0, Op1, Mask, VL);
 }

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vwmulsu.ll

 
 define <2 x i32> @vwmulsu_v2i32_v2i8(<2 x i8>* %x, <2 x i8>* %y) {
 ; CHECK-LABEL: vwmulsu_v2i32_v2i8:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetivli zero, 2, e8, mf8, ta, mu
 ; CHECK-NEXT:    vle8.v v8, (a0)
 ; CHECK-NEXT:    vle8.v v9, (a1)
 ; CHECK-NEXT:    vsetvli zero, zero, e16, mf4, ta, mu
-; CHECK-NEXT:    vsext.vf2 v10, v8
+; CHECK-NEXT:    vzext.vf2 v10, v8
 ; CHECK-NEXT:    vsext.vf2 v11, v9
 ; CHECK-NEXT:    vwmulsu.vv v8, v11, v10
 ; CHECK-NEXT:    ret
   %a = load <2 x i8>, <2 x i8>* %x
   %b = load <2 x i8>, <2 x i8>* %y
   %c = zext <2 x i8> %a to <2 x i32>
   %d = sext <2 x i8> %b to <2 x i32>
   %e = mul <2 x i32> %c, %d
   ret <2 x i32> %e
 }
 
 define <4 x i32> @vwmulsu_v4i32_v4i8_v4i16(<4 x i8>* %x, <4 x i16>* %y) {
 ; CHECK-LABEL: vwmulsu_v4i32_v4i8_v4i16:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetivli zero, 4, e8, mf4, ta, mu
 ; CHECK-NEXT:    vle8.v v8, (a0)
 ; CHECK-NEXT:    vle16.v v9, (a1)
 ; CHECK-NEXT:    vsetvli zero, zero, e16, mf2, ta, mu
-; CHECK-NEXT:    vsext.vf2 v10, v8
+; CHECK-NEXT:    vzext.vf2 v10, v8
 ; CHECK-NEXT:    vwmulsu.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %a = load <4 x i8>, <4 x i8>* %x
   %b = load <4 x i16>, <4 x i16>* %y
   %c = zext <4 x i8> %a to <4 x i32>
   %d = sext <4 x i16> %b to <4 x i32>
   %e = mul <4 x i32> %c, %d
   ret <4 x i32> %e