[RISCV] Prevent re-ordering some adds after shifts

[lenary]

Summary

DAGCombine will normally turn a (shl (add x, c1), c2) into (add (shl x, c2), c1 << c2), where c1 and c2 are constants. This can be prevented by a callback in TargetLowering.

On RISC-V, materialising the constant c1 << c2 can be more expensive than materialising c1, because materialising the former may take more instructions, and may use a register, where materialising the latter would not.

This patch implements the hook in RISCVTargetLowering to prevent this transform, in the cases where:

• c1 fits into the immediate field in an addi instruction.
• c1 takes fewer instructions to materialise than c1 << c2.

In future, DAGCombine could do the check to see whether c1 fits into an add immediate, which might simplify more targets hooks than just RISC-V.

[lenary]

• Add commets about larger constants. These can be improved at a later date

[lenary]

• Generalise optimisation to check materialisation cost

[asb]

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

860

I was actually thinking it might be better to define getIntImmCost for RISC-V (which at least initially uses generateInstSeq, even if that might result in a little wasted work), then call that from here (that change might affect codegen in other areas, but should be an improvement). Arguably the introduction of getIntImmCost could make sense as a separate patch (that this one depends on), if a sensible standalone test case is straight forward.

[lenary]

• Abstract away calculation of Materialisation Cost

[lewis-revill]

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

869

Isn't there an inaccuracy in this method of checking the materialization cost since RISCVMatInt::generateInstSeq always calculates the cost of materializing into a register? In this case we have instructions which might use immediates, but this will calculate the cost as being the same as a single-instruction materialization into a register followed by an instruction using a register.

IE:

addi rd, rs1, C

would appear to be the same cost as:

lui rs2, (C >> 12)
add rd, rs1, rs2

[lenary]

• Remove out-of-date comments

[asb]

llvm/lib/Target/RISCV/Utils/RISCVMatInt.cpp

78 ↗

(On Diff #203314)

Should add a comment to document what this does, and to document that it really does calculate the cost of materialising an integer (i.e. doesn't take into account whether there might be an opportunity for merging it into an addi). You should also document that it is invalid to call this for a Val which can't be represented with 32Bits when Is64Bit is false (that triggers an assert in generateInstSeq - I think it's probably still ok to treat that as an API misuse rather than adding more explicit error handling).

[lenary]

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

869

On line 860, we check that C1 (which you're calling C) will fit into an add immediate. If it will, that counts as "free", and we definitely want to prevent the re-ordering, and we don't even check the materialisation cost.

In fact, I'm going to update the patch to always allow the re-ordering if C1 << C2 will fit into an add immediate, because then we also know that materialisation of that constant is "free", and so we should allow the re-ordering because it might help later dagcombines.

[lenary]

• Allow Combine if C1 << C2 will fit into an immediate
• Explain restrictions on RISCVMatInt::getIntMatCost

[asb]

I added some minor comments, along with a bigger suggestion. What do you think about adding a getIntMatCost taking APInt and IsRV64, which will split the immediate into XLEN-sized chunks (see comment for reference to similar code in AArch64)?

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

848

Actually, with the logic there now I guess we're checking that (add val, c1) is fewer instructions than (add, val', c1 << c2) which is a similar condition but not quite the same. i.e. both c1 and c1 << c2 might be materialisable with a single instruction, but if c1 << c2 is materialised using a single lui it's still unprofitable as we can't merge it into the add.

860

shift immediate -> add immediate

861

End comment with a full stop

870

I'm not totally liking calling getIntMatCost with something that isn't logically equivalent to IsRV64 (you could of course have MVT::i64 on RV32 pre-legalisation). The cost of materialising a 64-bit constant on RV64 also isn't necessarily the same as materialising an i64 split into two i32 on RV2.

In an ideal world, we'd have a getIntMatCost taking an APInt+IsRV64, with similar logic to int AArch64TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) - i.e. splitting into 32-bit/64-bit chunks. Hard to imagine this being a big deal for this particular case, but it's good infrastructure to have.

[lenary]

Address review feedback

• Introduce new getIntMatCost(const APInt &Val, bool IsRV64) API, which can materialise much wider constants than the previous method.
• Clarify and check grammar in comments.

[xbolva00]

In future, DAGCombine could do the check to see whether c1 fits into an add immediate, which might simplify more targets hooks than just RISC-V.

Why not do this the right way already?

@lebedev.ri @craig.topper

[asb]

Thanks for the update Sam, I think there might be a minor correctness issue with getIntMatCost - let me know what you think.

It would be great to add a simple sanity check for the chunking logic. Perhaps there's a test case involving -1 that produces a different answer with the current version of the patch versus a version updated to use the type size for the bit size.

llvm/lib/Target/RISCV/Utils/RISCVMatInt.cpp

78 ↗

(On Diff #205064)

Given you have a description in the header, I think this repeated description for the implementation is redundant (see https://llvm.org/docs/CodingStandards.html#doxygen-use-in-documentation-comments "Don’t duplicate the documentation comment in the header file and in the implementation file"). Though maybe I missed a de facto standard elsewhere in the codebase for duplicating a reduced version. I don't feel strongly about this, so do feel free to keep if you prefer.

81 ↗

(On Diff #205064)

I think this isn't going to give the desired result if Val is e.g. -1. In that case, getMinSignedBits is going to return 1, meaning the loop is only executed once regardless of the original type width. If the original type was e.g. an i64 on RV32, then two instructions are required to materialise the constant, yet the logic in this function will only return 1.

I think the solutation is to more closely mirror the AArch64TTIImpl::getIntImmCost function I mentioned before, and add a Type parameter (and maybe also adopt similar logic for sign-extending constants to be a multiple of the PlatRegSize).

llvm/lib/Target/RISCV/Utils/RISCVMatInt.h

41 ↗

(On Diff #205064)

Nit: Shouldn't this be Is64Bit to to match the naming in the implementation?

[asb]

In D62857#1546071, @xbolva00 wrote:

In future, DAGCombine could do the check to see whether c1 fits into an add immediate, which might simplify more targets hooks than just RISC-V.

Why not do this the right way already?

@lebedev.ri @craig.topper

Now this patch has been extended to cover more cases (i.e. comparing materialisation cost rather than just identifying the isLegalAddImmediate case), it wouldn't have much effect on this backend. A patch that uses isLegalAddImmediate in the relevant DAGCombine might be a small benefit for targets that don't implement the isDesirableToCommuteWithShift hook, but I don't think it would affect the implementation here (it's not obvious to me the hook implementation should assume isLegalAddImmediate had already been checked). So if it does make sense to add, I think it's definitely a separate patch to this one.

[asb]

llvm/test/CodeGen/RISCV/add-before-shl.ll

14

The patch has since been updated to do a direct cost comparison, rather than just looking at the case where the constant fits into an immediate. These two paragraphs should be updated to reflect that

[lenary]

Address review feedback

• Update getIntMatCost to take an integer size (in bits). This ensures we chunk the constant correctly for the legal types on the target, and account for the costs of all required chunks. I was unable to devise a simple test case for when this behaviour would not match the behaivour using getMinSignedBits, due to legalisation always splitting the wider type before the isDesirableToCommuteWithShift callback is called.
  
  The chunking will automatically expand each chunk to be the platform register width, so we don't need to sign extend the constant to be a multiple of that width before we start chunking.

• Update and de-duplicate comments on tests and implementations
• Update naming of IsRV64 parameter in RISCVMatInt.{h,cpp}

[lenary]

llvm/lib/Target/RISCV/Utils/RISCVMatInt.cpp

81 ↗

(On Diff #205064)

I think I now cover chunking correctly. As in my message above, I don't need to sign extend Val before chunking, because each chunk is extended to be PlatRegSize within the loop.

[asb]

LGTM, thanks!

llvm/lib/Target/RISCV/RISCVISelLowering.h

ISD::NodeType getExtendForAtomicOps() const override {
     return ISD::SIGN_EXTEND;
   }
 
   bool shouldExpandShift(SelectionDAG &DAG, SDNode *N) const override {
     if (DAG.getMachineFunction().getFunction().hasMinSize())
       return false;
     return true;
   }
+  bool isDesirableToCommuteWithShift(const SDNode *N,
+                                     CombineLevel Level) const override;
 
 private:
   void analyzeInputArgs(MachineFunction &MF, CCState &CCInfo,
                         const SmallVectorImpl<ISD::InputArg> &Ins,
                         bool IsRet) const;
   void analyzeOutputArgs(MachineFunction &MF, CCState &CCInfo,
                          const SmallVectorImpl<ISD::OutputArg> &Outs,
                          bool IsRet, CallLoweringInfo *CLI) const;

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

 //===----------------------------------------------------------------------===//
 
 #include "RISCVISelLowering.h"
 #include "RISCV.h"
 #include "RISCVMachineFunctionInfo.h"
 #include "RISCVRegisterInfo.h"
 #include "RISCVSubtarget.h"
 #include "RISCVTargetMachine.h"
+#include "Utils/RISCVMatInt.h"
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/CodeGen/CallingConvLower.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/SelectionDAGISel.h"
return DCI.CombineTo(N,
                          DAG.getNode(ISD::AND, DL, MVT::i64, NewFMV,
                                      DAG.getConstant(~SignBit, DL, MVT::i64)));
   }
   }
 
   return SDValue();
 }
 
+bool RISCVTargetLowering::isDesirableToCommuteWithShift(
+    const SDNode *N, CombineLevel Level) const {
+  // The following folds are only desirable if constant `c1 << c2` can be
+  // materialised in fewer instructions than the constant `c1`:
+  //   (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
+  //   (shl (or x, c1), c2) -> (or (shl x, c2), c1 << c2)
+  SDValue N0 = N->getOperand(0);
+  EVT VT = N0.getValueType();
+  if ((VT == MVT::i64 || VT == MVT::i32) &&
+      (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::OR)) {
+    auto *C1 = dyn_cast<ConstantSDNode>(N0->getOperand(1));
+    auto *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1));
+    if (C1 && C2) {
+      APInt C1Int = C1->getAPIntValue();
+      APInt C2Int = C2->getAPIntValue();
+      if (C1Int.isSignedIntN(64) && C2Int.isSignedIntN(64)) {
+        RISCVMatInt::InstSeq C1Seq;
+        RISCVMatInt::generateInstSeq(C1Int.getLimitedValue(), VT == MVT::i64,
+                                     C1Seq);
+
+        RISCVMatInt::InstSeq ShiftedC1Seq;
+        RISCVMatInt::generateInstSeq((C1Int << C2Int).getLimitedValue(),
+                                     VT == MVT::i64, ShiftedC1Seq);
+
+        if (C1Seq.size() <= ShiftedC1Seq.size())
+          return false;
+      }
+    }
+  }
+  return true;
+}
+
 unsigned RISCVTargetLowering::ComputeNumSignBitsForTargetNode(
     SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
     unsigned Depth) const {
   switch (Op.getOpcode()) {
   default:
     break;
   case RISCVISD::SLLW:
   case RISCVISD::SRAW:

llvm/test/CodeGen/RISCV/add-before-shl.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefix=RV32I %s
+; RUN: llc -mtriple=riscv64 -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefix=RV64I %s
+
+; These test that constant adds are not moved after shifts by DAGCombine,
+; if the constant can fit into an immediate.
+;
+; Materialising the large (shifted) constant produced for the new add
+; uses an extra register, and takes several instructions. It is more
+; efficient to perform the add before the shift if the constant to be
+; added fits into an immediate.
+
+define signext i32 @add_small_const(i32 signext %a) nounwind {
+; RV32I-LABEL: add_small_const:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi a0, a0, 1
+; RV32I-NEXT:    slli a0, a0, 24
+; RV32I-NEXT:    srai a0, a0, 24
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: add_small_const:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    addi a0, a0, 1
+; RV64I-NEXT:    slli a0, a0, 56
+; RV64I-NEXT:    srai a0, a0, 56
+; RV64I-NEXT:    ret
+  %1 = add i32 %a, 1
+  %2 = shl i32 %1, 24
+  %3 = ashr i32 %2, 24
+  ret i32 %3
+}
+
+; NOTE: This add constant does not fit into an add immediate, so we allow the
+; the transformation to fire. However, this introduces a second left shift,
+; which we wouldn't need if we did the add before the shl.
+define signext i32 @add_large_const(i32 signext %a) nounwind {
+; RV32I-LABEL: add_large_const:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    slli a0, a0, 16
+; RV32I-NEXT:    lui a1, 65520
+; RV32I-NEXT:    add a0, a0, a1
+; RV32I-NEXT:    srai a0, a0, 16
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: add_large_const:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    lui a1, 1
+; RV64I-NEXT:    addiw a1, a1, -1
+; RV64I-NEXT:    add a0, a0, a1
+; RV64I-NEXT:    slli a0, a0, 48
+; RV64I-NEXT:    srai a0, a0, 48
+; RV64I-NEXT:    ret
+  %1 = add i32 %a, 4095
+  %2 = shl i32 %1, 16
+  %3 = ashr i32 %2, 16
+  ret i32 %3
+}
+
+; NOTE: This add constant does not fit into an add immediate, so we allow the
+; the transformation to fire. However, this introduces a second left shift,
+; which we wouldn't need if we did the add before the shl.
+define signext i32 @add_huge_const(i32 signext %a) nounwind {
+; RV32I-LABEL: add_huge_const:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    slli a0, a0, 16
+; RV32I-NEXT:    lui a1, 524272
+; RV32I-NEXT:    add a0, a0, a1
+; RV32I-NEXT:    srai a0, a0, 16
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: add_huge_const:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    lui a1, 8
+; RV64I-NEXT:    addiw a1, a1, -1
+; RV64I-NEXT:    add a0, a0, a1
+; RV64I-NEXT:    slli a0, a0, 48
+; RV64I-NEXT:    srai a0, a0, 48
+; RV64I-NEXT:    ret
+  %1 = add i32 %a, 32767
+  %2 = shl i32 %1, 16
+  %3 = ashr i32 %2, 16
+  ret i32 %3
+}