[RISCV] Teach instruction selection to elide sext.w in some cases.

[craig.topper]

Summary

If a sext_inreg is up for isel, and all its users are W instructions,
we can skip emitting the sext_inreg. This helpful if the producing
instruction can't become a W instruction.

[asb]

LGTM!

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

 // may be able to use a W instruction and CSE with the other instruction if
 // this has happened. We could try to detect that the CSE opportunity exists
 // before doing this, but that would be more complicated.
 // TODO: Does this need to look through AND/OR/XOR to their users to find more
 // opportunities.
 bool RISCVDAGToDAGISel::hasAllNBitUsers(SDNode *Node, unsigned Bits) const {
   assert((Node->getOpcode() == ISD::ADD || Node->getOpcode() == ISD::SUB ||
           Node->getOpcode() == ISD::MUL || Node->getOpcode() == ISD::SHL ||
+          Node->getOpcode() == ISD::SIGN_EXTEND_INREG ||
           isa<ConstantSDNode>(Node)) &&
          "Unexpected opcode");
 
   for (auto UI = Node->use_begin(), UE = Node->use_end(); UI != UE; ++UI) {
     SDNode *User = *UI;
     // Users of this node should have already been instruction selected
     if (!User->isMachineOpcode())
       return false;

llvm/lib/Target/RISCV/RISCVInstrInfo.td

 // PatFrag to allow ADDW/SUBW/MULW/SLLW to be selected from i64 add/sub/mul/shl
 // if only the lower 32 bits of their result is used.
 class overflowingbinopw<SDPatternOperator operator>
     : PatFrag<(ops node:$lhs, node:$rhs),
               (operator node:$lhs, node:$rhs), [{
   return hasAllWUsers(Node);
 }]>;
 
+def sexti32_allwusers : PatFrag<(ops node:$src),
+                                (sext_inreg node:$src, i32), [{
+  return hasAllWUsers(Node);
+}]>;
+
 let Predicates = [IsRV64] in {
 
 /// sext and zext
 
+// Sign extend is not needed if all users are W instructions.
+def : Pat<(sexti32_allwusers GPR:$rs1), (XLenVT GPR:$rs1)>;
+
 def : Pat<(sext_inreg GPR:$rs1, i32), (ADDIW GPR:$rs1, 0)>;
 
 /// ALU operations
 
 def : Pat<(i64 (srl (and GPR:$rs1, 0xffffffff), uimm5:$shamt)),
           (SRLIW GPR:$rs1, uimm5:$shamt)>;
 def : Pat<(i64 (srl (shl GPR:$rs1, (i64 32)), uimm6gt32:$shamt)),
           (SRLIW GPR:$rs1, (ImmSub32 uimm6gt32:$shamt))>;

llvm/test/CodeGen/RISCV/rv64zba.ll

; RV64ZBA-NEXT:    ret
   ret i64 %5
 }
 
 ; Type legalization inserts a sext_inreg after the first add. That add will be
 ; selected as sh2add which does not sign extend. SimplifyDemandedBits is unable
 ; to remove the sext_inreg because it has multiple uses. The ashr will use the
 ; sext_inreg to become sraiw. This leaves the sext_inreg only used by the shl.
 ; If the shl is selected as sllw, we don't need the sext_inreg.
-; FIXME: We should not emit a sext.w.
 define i64 @sh2add_extra_sext(i32 %x, i32 %y, i32 %z) {
 ; RV64I-LABEL: sh2add_extra_sext:
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    slliw a0, a0, 2
 ; RV64I-NEXT:    addw a0, a0, a1
 ; RV64I-NEXT:    sllw a1, a2, a0
 ; RV64I-NEXT:    sraiw a0, a0, 2
 ; RV64I-NEXT:    mul a0, a1, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64B-LABEL: sh2add_extra_sext:
 ; RV64B:       # %bb.0:
 ; RV64B-NEXT:    sh2add a0, a0, a1
-; RV64B-NEXT:    sext.w a1, a0
-; RV64B-NEXT:    sllw a1, a2, a1
+; RV64B-NEXT:    sllw a1, a2, a0
 ; RV64B-NEXT:    sraiw a0, a0, 2
 ; RV64B-NEXT:    mul a0, a1, a0
 ; RV64B-NEXT:    ret
 ;
 ; RV64ZBA-LABEL: sh2add_extra_sext:
 ; RV64ZBA:       # %bb.0:
 ; RV64ZBA-NEXT:    sh2add a0, a0, a1
-; RV64ZBA-NEXT:    sext.w a1, a0
-; RV64ZBA-NEXT:    sllw a1, a2, a1
+; RV64ZBA-NEXT:    sllw a1, a2, a0
 ; RV64ZBA-NEXT:    sraiw a0, a0, 2
 ; RV64ZBA-NEXT:    mul a0, a1, a0
 ; RV64ZBA-NEXT:    ret
   %a = shl i32 %x, 2
   %b = add i32 %a, %y
   %c = shl i32 %z, %b
   %d = ashr i32 %b, 2
   %e = sext i32 %c to i64