[RISCV] Add DAG combine to detect opportunities to replace (i64 (any_extend (i32 X)) with sign_extend.

[craig.topper]

Summary

If type legalization is going to insert a sign_extend for other users
of X and we can fold the sign_extend into ADDW/MULW/SUBW, it is
better to replace the ANY_EXTEND so we don't end up with a separate
ADD/MUL/SUB instruction for the users of the ANY_EXTEND.

I'm only handling setcc uses right now, but there are other
instructions that force sign_extends like ashr.

There are probably other *W instructions we could use in addition
to ADDW/SUBW/MULW.

My motivating case was a loop terminating compare and a phi use
as seen in the new test file.

[asb]

Running this across the GCC torture suite I'm seeing a fair few cases where this triggers and improves codegen. This is a nice improvement, thanks!

[nathanchance]

This patch causes an infinite loop while compiling the RISC-V Linux kernel's allmodconfig target. A simplified reproducer:

$ cat aspeed-pwm-tacho.i
char aspeed_set_pwm_port_fan_ctrl_priv_0_0,
    aspeed_set_pwm_port_fan_ctrl_fan_ctrl;
pwm_store_period, pwm_store_dc_time_on;
pwm_store() {
  pwm_store_period = aspeed_set_pwm_port_fan_ctrl_priv_0_0;
  pwm_store_period += 1;
  pwm_store_dc_time_on =
      aspeed_set_pwm_port_fan_ctrl_fan_ctrl * pwm_store_period / 5;
  if (pwm_store_dc_time_on)
    aspeed_set_pwm_port_duty_rising_falling();
}

$ timeout 5s clang -O0 --target=riscv64-linux-gnu -c -o /dev/null aspeed-pwm-tacho.i; echo $?
...
0

$ timeout 5s clang -O1 --target=riscv64-linux-gnu -c -o /dev/null aspeed-pwm-tacho.i; echo $?
...
0

$ timeout 5s clang -O2 --target=riscv64-linux-gnu -c -o /dev/null aspeed-pwm-tacho.i; echo $?
...
124

[craig.topper]

In D104581#2902007, @nathanchance wrote:

This patch causes an infinite loop while compiling the RISC-V Linux kernel's allmodconfig target. A simplified reproducer:

$ cat aspeed-pwm-tacho.i
char aspeed_set_pwm_port_fan_ctrl_priv_0_0,
    aspeed_set_pwm_port_fan_ctrl_fan_ctrl;
pwm_store_period, pwm_store_dc_time_on;
pwm_store() {
  pwm_store_period = aspeed_set_pwm_port_fan_ctrl_priv_0_0;
  pwm_store_period += 1;
  pwm_store_dc_time_on =
      aspeed_set_pwm_port_fan_ctrl_fan_ctrl * pwm_store_period / 5;
  if (pwm_store_dc_time_on)
    aspeed_set_pwm_port_duty_rising_falling();
}

$ timeout 5s clang -O0 --target=riscv64-linux-gnu -c -o /dev/null aspeed-pwm-tacho.i; echo $?
...
0

$ timeout 5s clang -O1 --target=riscv64-linux-gnu -c -o /dev/null aspeed-pwm-tacho.i; echo $?
...
0

$ timeout 5s clang -O2 --target=riscv64-linux-gnu -c -o /dev/null aspeed-pwm-tacho.i; echo $?
...
124

This should prevent the transform for this case https://reviews.llvm.org/D106754

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
   setJumpIsExpensive();
 
   // We can use any register for comparisons
   setHasMultipleConditionRegisters();
 
   setTargetDAGCombine(ISD::AND);
   setTargetDAGCombine(ISD::OR);
   setTargetDAGCombine(ISD::XOR);
+  setTargetDAGCombine(ISD::ANY_EXTEND);
   if (Subtarget.hasStdExtV()) {
     setTargetDAGCombine(ISD::FCOPYSIGN);
     setTargetDAGCombine(ISD::MGATHER);
     setTargetDAGCombine(ISD::MSCATTER);
     setTargetDAGCombine(ISD::SRA);
     setTargetDAGCombine(ISD::SRL);
     setTargetDAGCombine(ISD::SHL);
   }
static SDValue performXORCombine(SDNode *N,
                                  const RISCVSubtarget &Subtarget) {
   SelectionDAG &DAG = DCI.DAG;
 
   // fold (xor (select_cc lhs, rhs, cc, 0, y), x) ->
   //      (select lhs, rhs, cc, x, (xor x, y))
   return combineSelectCCAndUseCommutative(N, DAG, false);
 }
 
+// Attempt to turn ANY_EXTEND into SIGN_EXTEND if the input to the ANY_EXTEND
+// has users that require SIGN_EXTEND and the SIGN_EXTEND can be done for free
+// by an instruction like ADDW/SUBW/MULW. Without this the ANY_EXTEND would be
+// removed during type legalization leaving an ADD/SUB/MUL use that won't use
+// ADDW/SUBW/MULW.
+static SDValue performANY_EXTENDCombine(SDNode *N,
+                                        TargetLowering::DAGCombinerInfo &DCI,
+                                        const RISCVSubtarget &Subtarget) {
+  if (!Subtarget.is64Bit())
+    return SDValue();
+
+  SelectionDAG &DAG = DCI.DAG;
+
+  SDValue Src = N->getOperand(0);
+  EVT VT = N->getValueType(0);
+  if (VT != MVT::i64 || Src.getValueType() != MVT::i32)
+    return SDValue();
+
+  // The opcode must be one that can implicitly sign_extend.
+  // FIXME: Additional opcodes.
+  switch (Src.getOpcode()) {
+  default:
+    return SDValue();
+  case ISD::MUL:
+    if (!Subtarget.hasStdExtM())
+      return SDValue();
+    LLVM_FALLTHROUGH;
+  case ISD::ADD:
+  case ISD::SUB:
+    break;
+  }
+
+  SmallVector<SDNode *, 4> SetCCs;
+  for (SDNode::use_iterator UI = Src.getNode()->use_begin(),
+                            UE = Src.getNode()->use_end();
+       UI != UE; ++UI) {
+    SDNode *User = *UI;
+    if (User == N)
+      continue;
+    if (UI.getUse().getResNo() != Src.getResNo())
+      continue;
+    // All i32 setccs are legalized by sign extending operands.
+    if (User->getOpcode() == ISD::SETCC) {
+      SetCCs.push_back(User);
+      continue;
+    }
+    // We don't know if we can extend this user.
+    break;
+  }
+
+  // If we don't have any SetCCs, this isn't worthwhile.
+  if (SetCCs.empty())
+    return SDValue();
+
+  SDLoc DL(N);
+  SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Src);
+  DCI.CombineTo(N, SExt);
+
+  // Promote all the setccs.
+  for (SDNode *SetCC : SetCCs) {
+    SmallVector<SDValue, 4> Ops;
+
+    for (unsigned j = 0; j != 2; ++j) {
+      SDValue SOp = SetCC->getOperand(j);
+      if (SOp == Src)
+        Ops.push_back(SExt);
+      else
+        Ops.push_back(DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, SOp));
+    }
+
+    Ops.push_back(SetCC->getOperand(2));
+    DCI.CombineTo(SetCC,
+                  DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
+  }
+  return SDValue(N, 0);
+}
+
 SDValue RISCVTargetLowering::PerformDAGCombine(SDNode *N,
                                                DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
 
   switch (N->getOpcode()) {
   default:
     break;
   case RISCVISD::SplitF64: {
return DAG.getNode(ISD::AND, DL, MVT::i64, NewFMV,
                        DAG.getConstant(~SignBit, DL, MVT::i64));
   }
   case ISD::AND:
     return performANDCombine(N, DCI, Subtarget);
   case ISD::OR:
     return performORCombine(N, DCI, Subtarget);
   case ISD::XOR:
     return performXORCombine(N, DCI, Subtarget);
+  case ISD::ANY_EXTEND:
+    return performANY_EXTENDCombine(N, DCI, Subtarget);
   case RISCVISD::SELECT_CC: {
     // Transform
     SDValue LHS = N->getOperand(0);
     SDValue RHS = N->getOperand(1);
     auto CCVal = static_cast<ISD::CondCode>(N->getConstantOperandVal(2));
     if (!ISD::isIntEqualitySetCC(CCVal))
       break;
 

llvm/test/CodeGen/RISCV/aext-to-sext.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv64 -verify-machineinstrs < %s \
+; RUN:   | FileCheck %s -check-prefix=RV64I
+
+; Make sure we don't generate an addi in the loop in
+; addition to the addiw. Previously we type legalize the
+; setcc use using signext and the phi use using anyext.
+; We now detect when it would be beneficial to replace
+; anyext with signext.
+
+define void @quux(i32 signext %arg, i32 signext %arg1) nounwind {
+; RV64I-LABEL: quux:
+; RV64I:       # %bb.0: # %bb
+; RV64I-NEXT:    addi sp, sp, -32
+; RV64I-NEXT:    sd ra, 24(sp) # 8-byte Folded Spill
+; RV64I-NEXT:    sd s0, 16(sp) # 8-byte Folded Spill
+; RV64I-NEXT:    sd s1, 8(sp) # 8-byte Folded Spill
+; RV64I-NEXT:    beq a0, a1, .LBB0_3
+; RV64I-NEXT:  # %bb.1: # %bb2.preheader
+; RV64I-NEXT:    mv s0, a1
+; RV64I-NEXT:    mv s1, a0
+; RV64I-NEXT:  .LBB0_2: # %bb2
+; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
+; RV64I-NEXT:    call hoge@plt
+; RV64I-NEXT:    addiw s1, s1, 1
+; RV64I-NEXT:    bne s1, s0, .LBB0_2
+; RV64I-NEXT:  .LBB0_3: # %bb6
+; RV64I-NEXT:    ld s1, 8(sp) # 8-byte Folded Reload
+; RV64I-NEXT:    ld s0, 16(sp) # 8-byte Folded Reload
+; RV64I-NEXT:    ld ra, 24(sp) # 8-byte Folded Reload
+; RV64I-NEXT:    addi sp, sp, 32
+; RV64I-NEXT:    ret
+bb:
+  %tmp = icmp eq i32 %arg, %arg1
+  br i1 %tmp, label %bb6, label %bb2
+
+bb2:                                              ; preds = %bb2, %bb
+  %tmp3 = phi i32 [ %tmp4, %bb2 ], [ %arg, %bb ]
+  tail call void @hoge()
+  %tmp4 = add nsw i32 %tmp3, 1
+  %tmp5 = icmp eq i32 %tmp4, %arg1
+  br i1 %tmp5, label %bb6, label %bb2
+
+bb6:                                              ; preds = %bb2, %bb
+  ret void
+}
+
+declare void @hoge()

llvm/test/CodeGen/RISCV/rv64zbb.ll

 }
 
 define signext i32 @log2_ceil_i32(i32 signext %a) nounwind {
 ; RV64I-LABEL: log2_ceil_i32:
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    addi sp, sp, -16
 ; RV64I-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
 ; RV64I-NEXT:    sd s0, 0(sp) # 8-byte Folded Spill
-; RV64I-NEXT:    addiw a2, a0, -1
+; RV64I-NEXT:    addiw a0, a0, -1
 ; RV64I-NEXT:    addi s0, zero, 32
 ; RV64I-NEXT:    addi a1, zero, 32
-; RV64I-NEXT:    beqz a2, .LBB2_2
+; RV64I-NEXT:    beqz a0, .LBB2_2
 ; RV64I-NEXT:  # %bb.1: # %cond.false
-; RV64I-NEXT:    addi a0, a0, -1
 ; RV64I-NEXT:    srliw a1, a0, 1
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    srli a0, a0, 32
 ; RV64I-NEXT:    or a0, a0, a1
 ; RV64I-NEXT:    srli a1, a0, 2
 ; RV64I-NEXT:    or a0, a0, a1
 ; RV64I-NEXT:    srli a1, a0, 4
 ; RV64I-NEXT:    or a0, a0, a1