Files
clang-p2996/llvm/lib/Target/RISCV/RISCVOptWInstrs.cpp
Alex Bradbury 3d2650bdeb [RISCV] Use addi rather than addiw for immediates materialised by lui+addi(w) pairs when possible (#141663)
The logic in RISCVMatInt would previously produce lui+addiw on RV64
whenever a 32-bit integer must be materialised and the Hi20 and Lo12
parts are non-zero. However, sometimes addi can be used equivalently
(whenever the sign extension behaviour of addiw would be a no-op). This
patch moves to using addiw only when necessary. Although there is
absolutely no advantage in terms of compressibility or performance, this
has the following advantages:
* It's more consistent with logic used elsewhere in the backend. For
instance, RISCVOptWInstrs will try to convert addiw to addi on the basis
it reduces test diffs vs RV32.
* This matches the lowering GCC does in its codegen path. Unlike LLVM,
GCC seems to have different expansion logic for the assembler vs
codegen. For codegen it will use lui+addi if possible, but expanding
`li` in the assembler will always produces lui+addiw as LLVM did prior
to this commit. As someone who has been looking at a lot of gcc vs clang
diffs lately, reducing unnecessary divergence is of at least some value.
* As the diff for fold-mem-offset.ll shows, we can fold memory offsets
in more cases when addi is used. Memory offset folding could be taught
to recognise when the addiw could be replaced with an addi, but that
seems unnecessary when we can simply change the logic in RISCVMatInt.

As pointed out by @topperc during review, making this change without
modifying RISCVOptWInstrs risks introducing some cases where we fail to
remove a sext.w that we removed before. I've incorporated a patch based
on a suggestion from Craig that avoids it, and also adds appropriate
RISCVOptWInstrs test cases.

The initial patch description noted that the main motivation was to
avoid unnecessary differences both for RV32/RV64 and when comparing GCC,
but noted that very occasionally we see a benefit from memory offset
folding kicking in when it didn't before. Looking at the dynamic
instruction count difference for SPEC benchmarks targeting rva22u64 and
it shows we actually get a meaningful
~4.3% reduction in dynamic icount for 519.lbm_r. Looking at the data
more closely, the codegen difference is in `LBM_performStreamCollideTRT`
which as a function accounts for ~98% for dynamically executed
instructions and the codegen diffs appear to be a knock-on effect of the
address merging reducing register pressure right from function entry
(for instance, we get a big reduction in dynamically executed loads in
that function).

Below is the icount data (rva22u64 -O3, no LTO):
```
Benchmark                Baseline            This PR   Diff (%)
============================================================
500.perlbench_r         174116601991    174115795810     -0.00%
502.gcc_r               218903280858    218903215788     -0.00%
505.mcf_r               131208029185    131207692803     -0.00%
508.namd_r              217497594322    217497594297     -0.00%
510.parest_r            289314486153    289313577652     -0.00%
511.povray_r             30640531048     30640765701      0.00%
519.lbm_r                95897914862     91712688050     -4.36%
520.omnetpp_r           134641549722    134867015683      0.17%
523.xalancbmk_r         281462762992    281432092673     -0.01%
525.x264_r              379776121941    379535558210     -0.06%
526.blender_r           659736022025    659738387343      0.00%
531.deepsjeng_r         349122867552    349122867481     -0.00%
538.imagick_r           238558760552    238558753269     -0.00%
541.leela_r             406578560612    406385135260     -0.05%
544.nab_r               400997131674    400996765827     -0.00%
557.xz_r                130079522194    129945515709     -0.10%

```

The instcounting setup I use doesn't have good support for drilling down
into functions from outside the linked executable (e.g. libc). The
difference in omnetpp all seems to come from there, and does not reflect
any degradation in codegen quality.

I can confirm with the current version of the PR there is no change in
the number of static sext.w across all the SPEC 2017 benchmarks
(rva22u64 O3)

Co-authored-by: Craig Topper <craig.topper@sifive.com>
2025-06-02 22:24:50 +01:00

813 lines
25 KiB
C++

//===- RISCVOptWInstrs.cpp - MI W instruction optimizations ---------------===//
//
// 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
//
//===---------------------------------------------------------------------===//
//
// This pass does some optimizations for *W instructions at the MI level.
//
// First it removes unneeded sext.w instructions. Either because the sign
// extended bits aren't consumed or because the input was already sign extended
// by an earlier instruction.
//
// Then:
// 1. Unless explicit disabled or the target prefers instructions with W suffix,
// it removes the -w suffix from opw instructions whenever all users are
// dependent only on the lower word of the result of the instruction.
// The cases handled are:
// * addw because c.add has a larger register encoding than c.addw.
// * addiw because it helps reduce test differences between RV32 and RV64
// w/o being a pessimization.
// * mulw because c.mulw doesn't exist but c.mul does (w/ zcb)
// * slliw because c.slliw doesn't exist and c.slli does
//
// 2. Or if explicit enabled or the target prefers instructions with W suffix,
// it adds the W suffix to the instruction whenever all users are dependent
// only on the lower word of the result of the instruction.
// The cases handled are:
// * add/addi/sub/mul.
// * slli with imm < 32.
// * ld/lwu.
//===---------------------------------------------------------------------===//
#include "RISCV.h"
#include "RISCVMachineFunctionInfo.h"
#include "RISCVSubtarget.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
using namespace llvm;
#define DEBUG_TYPE "riscv-opt-w-instrs"
#define RISCV_OPT_W_INSTRS_NAME "RISC-V Optimize W Instructions"
STATISTIC(NumRemovedSExtW, "Number of removed sign-extensions");
STATISTIC(NumTransformedToWInstrs,
"Number of instructions transformed to W-ops");
static cl::opt<bool> DisableSExtWRemoval("riscv-disable-sextw-removal",
cl::desc("Disable removal of sext.w"),
cl::init(false), cl::Hidden);
static cl::opt<bool> DisableStripWSuffix("riscv-disable-strip-w-suffix",
cl::desc("Disable strip W suffix"),
cl::init(false), cl::Hidden);
namespace {
class RISCVOptWInstrs : public MachineFunctionPass {
public:
static char ID;
RISCVOptWInstrs() : MachineFunctionPass(ID) {}
bool runOnMachineFunction(MachineFunction &MF) override;
bool removeSExtWInstrs(MachineFunction &MF, const RISCVInstrInfo &TII,
const RISCVSubtarget &ST, MachineRegisterInfo &MRI);
bool stripWSuffixes(MachineFunction &MF, const RISCVInstrInfo &TII,
const RISCVSubtarget &ST, MachineRegisterInfo &MRI);
bool appendWSuffixes(MachineFunction &MF, const RISCVInstrInfo &TII,
const RISCVSubtarget &ST, MachineRegisterInfo &MRI);
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
MachineFunctionPass::getAnalysisUsage(AU);
}
StringRef getPassName() const override { return RISCV_OPT_W_INSTRS_NAME; }
};
} // end anonymous namespace
char RISCVOptWInstrs::ID = 0;
INITIALIZE_PASS(RISCVOptWInstrs, DEBUG_TYPE, RISCV_OPT_W_INSTRS_NAME, false,
false)
FunctionPass *llvm::createRISCVOptWInstrsPass() {
return new RISCVOptWInstrs();
}
static bool vectorPseudoHasAllNBitUsers(const MachineOperand &UserOp,
unsigned Bits) {
const MachineInstr &MI = *UserOp.getParent();
unsigned MCOpcode = RISCV::getRVVMCOpcode(MI.getOpcode());
if (!MCOpcode)
return false;
const MCInstrDesc &MCID = MI.getDesc();
const uint64_t TSFlags = MCID.TSFlags;
if (!RISCVII::hasSEWOp(TSFlags))
return false;
assert(RISCVII::hasVLOp(TSFlags));
const unsigned Log2SEW = MI.getOperand(RISCVII::getSEWOpNum(MCID)).getImm();
if (UserOp.getOperandNo() == RISCVII::getVLOpNum(MCID))
return false;
auto NumDemandedBits =
RISCV::getVectorLowDemandedScalarBits(MCOpcode, Log2SEW);
return NumDemandedBits && Bits >= *NumDemandedBits;
}
// Checks if all users only demand the lower \p OrigBits of the original
// instruction's result.
// TODO: handle multiple interdependent transformations
static bool hasAllNBitUsers(const MachineInstr &OrigMI,
const RISCVSubtarget &ST,
const MachineRegisterInfo &MRI, unsigned OrigBits) {
SmallSet<std::pair<const MachineInstr *, unsigned>, 4> Visited;
SmallVector<std::pair<const MachineInstr *, unsigned>, 4> Worklist;
Worklist.push_back(std::make_pair(&OrigMI, OrigBits));
while (!Worklist.empty()) {
auto P = Worklist.pop_back_val();
const MachineInstr *MI = P.first;
unsigned Bits = P.second;
if (!Visited.insert(P).second)
continue;
// Only handle instructions with one def.
if (MI->getNumExplicitDefs() != 1)
return false;
Register DestReg = MI->getOperand(0).getReg();
if (!DestReg.isVirtual())
return false;
for (auto &UserOp : MRI.use_nodbg_operands(DestReg)) {
const MachineInstr *UserMI = UserOp.getParent();
unsigned OpIdx = UserOp.getOperandNo();
switch (UserMI->getOpcode()) {
default:
if (vectorPseudoHasAllNBitUsers(UserOp, Bits))
break;
return false;
case RISCV::ADDIW:
case RISCV::ADDW:
case RISCV::DIVUW:
case RISCV::DIVW:
case RISCV::MULW:
case RISCV::REMUW:
case RISCV::REMW:
case RISCV::SLLIW:
case RISCV::SLLW:
case RISCV::SRAIW:
case RISCV::SRAW:
case RISCV::SRLIW:
case RISCV::SRLW:
case RISCV::SUBW:
case RISCV::ROLW:
case RISCV::RORW:
case RISCV::RORIW:
case RISCV::CLZW:
case RISCV::CTZW:
case RISCV::CPOPW:
case RISCV::SLLI_UW:
case RISCV::FMV_W_X:
case RISCV::FCVT_H_W:
case RISCV::FCVT_H_W_INX:
case RISCV::FCVT_H_WU:
case RISCV::FCVT_H_WU_INX:
case RISCV::FCVT_S_W:
case RISCV::FCVT_S_W_INX:
case RISCV::FCVT_S_WU:
case RISCV::FCVT_S_WU_INX:
case RISCV::FCVT_D_W:
case RISCV::FCVT_D_W_INX:
case RISCV::FCVT_D_WU:
case RISCV::FCVT_D_WU_INX:
if (Bits >= 32)
break;
return false;
case RISCV::SEXT_B:
case RISCV::PACKH:
if (Bits >= 8)
break;
return false;
case RISCV::SEXT_H:
case RISCV::FMV_H_X:
case RISCV::ZEXT_H_RV32:
case RISCV::ZEXT_H_RV64:
case RISCV::PACKW:
if (Bits >= 16)
break;
return false;
case RISCV::PACK:
if (Bits >= (ST.getXLen() / 2))
break;
return false;
case RISCV::SRLI: {
// If we are shifting right by less than Bits, and users don't demand
// any bits that were shifted into [Bits-1:0], then we can consider this
// as an N-Bit user.
unsigned ShAmt = UserMI->getOperand(2).getImm();
if (Bits > ShAmt) {
Worklist.push_back(std::make_pair(UserMI, Bits - ShAmt));
break;
}
return false;
}
// these overwrite higher input bits, otherwise the lower word of output
// depends only on the lower word of input. So check their uses read W.
case RISCV::SLLI: {
unsigned ShAmt = UserMI->getOperand(2).getImm();
if (Bits >= (ST.getXLen() - ShAmt))
break;
Worklist.push_back(std::make_pair(UserMI, Bits + ShAmt));
break;
}
case RISCV::ANDI: {
uint64_t Imm = UserMI->getOperand(2).getImm();
if (Bits >= (unsigned)llvm::bit_width(Imm))
break;
Worklist.push_back(std::make_pair(UserMI, Bits));
break;
}
case RISCV::ORI: {
uint64_t Imm = UserMI->getOperand(2).getImm();
if (Bits >= (unsigned)llvm::bit_width<uint64_t>(~Imm))
break;
Worklist.push_back(std::make_pair(UserMI, Bits));
break;
}
case RISCV::SLL:
case RISCV::BSET:
case RISCV::BCLR:
case RISCV::BINV:
// Operand 2 is the shift amount which uses log2(xlen) bits.
if (OpIdx == 2) {
if (Bits >= Log2_32(ST.getXLen()))
break;
return false;
}
Worklist.push_back(std::make_pair(UserMI, Bits));
break;
case RISCV::SRA:
case RISCV::SRL:
case RISCV::ROL:
case RISCV::ROR:
// Operand 2 is the shift amount which uses 6 bits.
if (OpIdx == 2 && Bits >= Log2_32(ST.getXLen()))
break;
return false;
case RISCV::ADD_UW:
case RISCV::SH1ADD_UW:
case RISCV::SH2ADD_UW:
case RISCV::SH3ADD_UW:
// Operand 1 is implicitly zero extended.
if (OpIdx == 1 && Bits >= 32)
break;
Worklist.push_back(std::make_pair(UserMI, Bits));
break;
case RISCV::BEXTI:
if (UserMI->getOperand(2).getImm() >= Bits)
return false;
break;
case RISCV::SB:
// The first argument is the value to store.
if (OpIdx == 0 && Bits >= 8)
break;
return false;
case RISCV::SH:
// The first argument is the value to store.
if (OpIdx == 0 && Bits >= 16)
break;
return false;
case RISCV::SW:
// The first argument is the value to store.
if (OpIdx == 0 && Bits >= 32)
break;
return false;
// For these, lower word of output in these operations, depends only on
// the lower word of input. So, we check all uses only read lower word.
case RISCV::COPY:
case RISCV::PHI:
case RISCV::ADD:
case RISCV::ADDI:
case RISCV::AND:
case RISCV::MUL:
case RISCV::OR:
case RISCV::SUB:
case RISCV::XOR:
case RISCV::XORI:
case RISCV::ANDN:
case RISCV::BREV8:
case RISCV::CLMUL:
case RISCV::ORC_B:
case RISCV::ORN:
case RISCV::SH1ADD:
case RISCV::SH2ADD:
case RISCV::SH3ADD:
case RISCV::XNOR:
case RISCV::BSETI:
case RISCV::BCLRI:
case RISCV::BINVI:
Worklist.push_back(std::make_pair(UserMI, Bits));
break;
case RISCV::PseudoCCMOVGPR:
case RISCV::PseudoCCMOVGPRNoX0:
// Either operand 4 or operand 5 is returned by this instruction. If
// only the lower word of the result is used, then only the lower word
// of operand 4 and 5 is used.
if (OpIdx != 4 && OpIdx != 5)
return false;
Worklist.push_back(std::make_pair(UserMI, Bits));
break;
case RISCV::CZERO_EQZ:
case RISCV::CZERO_NEZ:
case RISCV::VT_MASKC:
case RISCV::VT_MASKCN:
if (OpIdx != 1)
return false;
Worklist.push_back(std::make_pair(UserMI, Bits));
break;
}
}
}
return true;
}
static bool hasAllWUsers(const MachineInstr &OrigMI, const RISCVSubtarget &ST,
const MachineRegisterInfo &MRI) {
return hasAllNBitUsers(OrigMI, ST, MRI, 32);
}
// This function returns true if the machine instruction always outputs a value
// where bits 63:32 match bit 31.
static bool isSignExtendingOpW(const MachineInstr &MI, unsigned OpNo) {
uint64_t TSFlags = MI.getDesc().TSFlags;
// Instructions that can be determined from opcode are marked in tablegen.
if (TSFlags & RISCVII::IsSignExtendingOpWMask)
return true;
// Special cases that require checking operands.
switch (MI.getOpcode()) {
// shifting right sufficiently makes the value 32-bit sign-extended
case RISCV::SRAI:
return MI.getOperand(2).getImm() >= 32;
case RISCV::SRLI:
return MI.getOperand(2).getImm() > 32;
// The LI pattern ADDI rd, X0, imm is sign extended.
case RISCV::ADDI:
return MI.getOperand(1).isReg() && MI.getOperand(1).getReg() == RISCV::X0;
// An ANDI with an 11 bit immediate will zero bits 63:11.
case RISCV::ANDI:
return isUInt<11>(MI.getOperand(2).getImm());
// An ORI with an >11 bit immediate (negative 12-bit) will set bits 63:11.
case RISCV::ORI:
return !isUInt<11>(MI.getOperand(2).getImm());
// A bseti with X0 is sign extended if the immediate is less than 31.
case RISCV::BSETI:
return MI.getOperand(2).getImm() < 31 &&
MI.getOperand(1).getReg() == RISCV::X0;
// Copying from X0 produces zero.
case RISCV::COPY:
return MI.getOperand(1).getReg() == RISCV::X0;
// Ignore the scratch register destination.
case RISCV::PseudoAtomicLoadNand32:
return OpNo == 0;
case RISCV::PseudoVMV_X_S: {
// vmv.x.s has at least 33 sign bits if log2(sew) <= 5.
int64_t Log2SEW = MI.getOperand(2).getImm();
assert(Log2SEW >= 3 && Log2SEW <= 6 && "Unexpected Log2SEW");
return Log2SEW <= 5;
}
}
return false;
}
static bool isSignExtendedW(Register SrcReg, const RISCVSubtarget &ST,
const MachineRegisterInfo &MRI,
SmallPtrSetImpl<MachineInstr *> &FixableDef) {
SmallSet<Register, 4> Visited;
SmallVector<Register, 4> Worklist;
auto AddRegToWorkList = [&](Register SrcReg) {
if (!SrcReg.isVirtual())
return false;
Worklist.push_back(SrcReg);
return true;
};
if (!AddRegToWorkList(SrcReg))
return false;
while (!Worklist.empty()) {
Register Reg = Worklist.pop_back_val();
// If we already visited this register, we don't need to check it again.
if (!Visited.insert(Reg).second)
continue;
MachineInstr *MI = MRI.getVRegDef(Reg);
if (!MI)
continue;
int OpNo = MI->findRegisterDefOperandIdx(Reg, /*TRI=*/nullptr);
assert(OpNo != -1 && "Couldn't find register");
// If this is a sign extending operation we don't need to look any further.
if (isSignExtendingOpW(*MI, OpNo))
continue;
// Is this an instruction that propagates sign extend?
switch (MI->getOpcode()) {
default:
// Unknown opcode, give up.
return false;
case RISCV::COPY: {
const MachineFunction *MF = MI->getMF();
const RISCVMachineFunctionInfo *RVFI =
MF->getInfo<RISCVMachineFunctionInfo>();
// If this is the entry block and the register is livein, see if we know
// it is sign extended.
if (MI->getParent() == &MF->front()) {
Register VReg = MI->getOperand(0).getReg();
if (MF->getRegInfo().isLiveIn(VReg) && RVFI->isSExt32Register(VReg))
continue;
}
Register CopySrcReg = MI->getOperand(1).getReg();
if (CopySrcReg == RISCV::X10) {
// For a method return value, we check the ZExt/SExt flags in attribute.
// We assume the following code sequence for method call.
// PseudoCALL @bar, ...
// ADJCALLSTACKUP 0, 0, implicit-def dead $x2, implicit $x2
// %0:gpr = COPY $x10
//
// We use the PseudoCall to look up the IR function being called to find
// its return attributes.
const MachineBasicBlock *MBB = MI->getParent();
auto II = MI->getIterator();
if (II == MBB->instr_begin() ||
(--II)->getOpcode() != RISCV::ADJCALLSTACKUP)
return false;
const MachineInstr &CallMI = *(--II);
if (!CallMI.isCall() || !CallMI.getOperand(0).isGlobal())
return false;
auto *CalleeFn =
dyn_cast_if_present<Function>(CallMI.getOperand(0).getGlobal());
if (!CalleeFn)
return false;
auto *IntTy = dyn_cast<IntegerType>(CalleeFn->getReturnType());
if (!IntTy)
return false;
const AttributeSet &Attrs = CalleeFn->getAttributes().getRetAttrs();
unsigned BitWidth = IntTy->getBitWidth();
if ((BitWidth <= 32 && Attrs.hasAttribute(Attribute::SExt)) ||
(BitWidth < 32 && Attrs.hasAttribute(Attribute::ZExt)))
continue;
}
if (!AddRegToWorkList(CopySrcReg))
return false;
break;
}
// For these, we just need to check if the 1st operand is sign extended.
case RISCV::BCLRI:
case RISCV::BINVI:
case RISCV::BSETI:
if (MI->getOperand(2).getImm() >= 31)
return false;
[[fallthrough]];
case RISCV::REM:
case RISCV::ANDI:
case RISCV::ORI:
case RISCV::XORI:
// |Remainder| is always <= |Dividend|. If D is 32-bit, then so is R.
// DIV doesn't work because of the edge case 0xf..f 8000 0000 / (long)-1
// Logical operations use a sign extended 12-bit immediate.
if (!AddRegToWorkList(MI->getOperand(1).getReg()))
return false;
break;
case RISCV::PseudoCCADDW:
case RISCV::PseudoCCADDIW:
case RISCV::PseudoCCSUBW:
case RISCV::PseudoCCSLLW:
case RISCV::PseudoCCSRLW:
case RISCV::PseudoCCSRAW:
case RISCV::PseudoCCSLLIW:
case RISCV::PseudoCCSRLIW:
case RISCV::PseudoCCSRAIW:
// Returns operand 4 or an ADDW/SUBW/etc. of operands 5 and 6. We only
// need to check if operand 4 is sign extended.
if (!AddRegToWorkList(MI->getOperand(4).getReg()))
return false;
break;
case RISCV::REMU:
case RISCV::AND:
case RISCV::OR:
case RISCV::XOR:
case RISCV::ANDN:
case RISCV::ORN:
case RISCV::XNOR:
case RISCV::MAX:
case RISCV::MAXU:
case RISCV::MIN:
case RISCV::MINU:
case RISCV::PseudoCCMOVGPR:
case RISCV::PseudoCCMOVGPRNoX0:
case RISCV::PseudoCCAND:
case RISCV::PseudoCCOR:
case RISCV::PseudoCCXOR:
case RISCV::PHI: {
// If all incoming values are sign-extended, the output of AND, OR, XOR,
// MIN, MAX, or PHI is also sign-extended.
// The input registers for PHI are operand 1, 3, ...
// The input registers for PseudoCCMOVGPR(NoX0) are 4 and 5.
// The input registers for PseudoCCAND/OR/XOR are 4, 5, and 6.
// The input registers for others are operand 1 and 2.
unsigned B = 1, E = 3, D = 1;
switch (MI->getOpcode()) {
case RISCV::PHI:
E = MI->getNumOperands();
D = 2;
break;
case RISCV::PseudoCCMOVGPR:
case RISCV::PseudoCCMOVGPRNoX0:
B = 4;
E = 6;
break;
case RISCV::PseudoCCAND:
case RISCV::PseudoCCOR:
case RISCV::PseudoCCXOR:
B = 4;
E = 7;
break;
}
for (unsigned I = B; I != E; I += D) {
if (!MI->getOperand(I).isReg())
return false;
if (!AddRegToWorkList(MI->getOperand(I).getReg()))
return false;
}
break;
}
case RISCV::CZERO_EQZ:
case RISCV::CZERO_NEZ:
case RISCV::VT_MASKC:
case RISCV::VT_MASKCN:
// Instructions return zero or operand 1. Result is sign extended if
// operand 1 is sign extended.
if (!AddRegToWorkList(MI->getOperand(1).getReg()))
return false;
break;
case RISCV::ADDI: {
if (MI->getOperand(1).isReg() && MI->getOperand(1).getReg().isVirtual()) {
if (MachineInstr *SrcMI = MRI.getVRegDef(MI->getOperand(1).getReg())) {
if (SrcMI->getOpcode() == RISCV::LUI &&
SrcMI->getOperand(1).isImm()) {
uint64_t Imm = SrcMI->getOperand(1).getImm();
Imm = SignExtend64<32>(Imm << 12);
Imm += (uint64_t)MI->getOperand(2).getImm();
if (isInt<32>(Imm))
continue;
}
}
}
if (hasAllWUsers(*MI, ST, MRI)) {
FixableDef.insert(MI);
break;
}
return false;
}
// With these opcode, we can "fix" them with the W-version
// if we know all users of the result only rely on bits 31:0
case RISCV::SLLI:
// SLLIW reads the lowest 5 bits, while SLLI reads lowest 6 bits
if (MI->getOperand(2).getImm() >= 32)
return false;
[[fallthrough]];
case RISCV::ADD:
case RISCV::LD:
case RISCV::LWU:
case RISCV::MUL:
case RISCV::SUB:
if (hasAllWUsers(*MI, ST, MRI)) {
FixableDef.insert(MI);
break;
}
return false;
}
}
// If we get here, then every node we visited produces a sign extended value
// or propagated sign extended values. So the result must be sign extended.
return true;
}
static unsigned getWOp(unsigned Opcode) {
switch (Opcode) {
case RISCV::ADDI:
return RISCV::ADDIW;
case RISCV::ADD:
return RISCV::ADDW;
case RISCV::LD:
case RISCV::LWU:
return RISCV::LW;
case RISCV::MUL:
return RISCV::MULW;
case RISCV::SLLI:
return RISCV::SLLIW;
case RISCV::SUB:
return RISCV::SUBW;
default:
llvm_unreachable("Unexpected opcode for replacement with W variant");
}
}
bool RISCVOptWInstrs::removeSExtWInstrs(MachineFunction &MF,
const RISCVInstrInfo &TII,
const RISCVSubtarget &ST,
MachineRegisterInfo &MRI) {
if (DisableSExtWRemoval)
return false;
bool MadeChange = false;
for (MachineBasicBlock &MBB : MF) {
for (MachineInstr &MI : llvm::make_early_inc_range(MBB)) {
// We're looking for the sext.w pattern ADDIW rd, rs1, 0.
if (!RISCVInstrInfo::isSEXT_W(MI))
continue;
Register SrcReg = MI.getOperand(1).getReg();
SmallPtrSet<MachineInstr *, 4> FixableDefs;
// If all users only use the lower bits, this sext.w is redundant.
// Or if all definitions reaching MI sign-extend their output,
// then sext.w is redundant.
if (!hasAllWUsers(MI, ST, MRI) &&
!isSignExtendedW(SrcReg, ST, MRI, FixableDefs))
continue;
Register DstReg = MI.getOperand(0).getReg();
if (!MRI.constrainRegClass(SrcReg, MRI.getRegClass(DstReg)))
continue;
// Convert Fixable instructions to their W versions.
for (MachineInstr *Fixable : FixableDefs) {
LLVM_DEBUG(dbgs() << "Replacing " << *Fixable);
Fixable->setDesc(TII.get(getWOp(Fixable->getOpcode())));
Fixable->clearFlag(MachineInstr::MIFlag::NoSWrap);
Fixable->clearFlag(MachineInstr::MIFlag::NoUWrap);
Fixable->clearFlag(MachineInstr::MIFlag::IsExact);
LLVM_DEBUG(dbgs() << " with " << *Fixable);
++NumTransformedToWInstrs;
}
LLVM_DEBUG(dbgs() << "Removing redundant sign-extension\n");
MRI.replaceRegWith(DstReg, SrcReg);
MRI.clearKillFlags(SrcReg);
MI.eraseFromParent();
++NumRemovedSExtW;
MadeChange = true;
}
}
return MadeChange;
}
bool RISCVOptWInstrs::stripWSuffixes(MachineFunction &MF,
const RISCVInstrInfo &TII,
const RISCVSubtarget &ST,
MachineRegisterInfo &MRI) {
bool MadeChange = false;
for (MachineBasicBlock &MBB : MF) {
for (MachineInstr &MI : MBB) {
unsigned Opc;
switch (MI.getOpcode()) {
default:
continue;
case RISCV::ADDW: Opc = RISCV::ADD; break;
case RISCV::ADDIW: Opc = RISCV::ADDI; break;
case RISCV::MULW: Opc = RISCV::MUL; break;
case RISCV::SLLIW: Opc = RISCV::SLLI; break;
}
if (hasAllWUsers(MI, ST, MRI)) {
MI.setDesc(TII.get(Opc));
MadeChange = true;
}
}
}
return MadeChange;
}
bool RISCVOptWInstrs::appendWSuffixes(MachineFunction &MF,
const RISCVInstrInfo &TII,
const RISCVSubtarget &ST,
MachineRegisterInfo &MRI) {
bool MadeChange = false;
for (MachineBasicBlock &MBB : MF) {
for (MachineInstr &MI : MBB) {
unsigned WOpc;
// TODO: Add more?
switch (MI.getOpcode()) {
default:
continue;
case RISCV::ADD:
WOpc = RISCV::ADDW;
break;
case RISCV::ADDI:
WOpc = RISCV::ADDIW;
break;
case RISCV::SUB:
WOpc = RISCV::SUBW;
break;
case RISCV::MUL:
WOpc = RISCV::MULW;
break;
case RISCV::SLLI:
// SLLIW reads the lowest 5 bits, while SLLI reads lowest 6 bits
if (MI.getOperand(2).getImm() >= 32)
continue;
WOpc = RISCV::SLLIW;
break;
case RISCV::LD:
case RISCV::LWU:
WOpc = RISCV::LW;
break;
}
if (hasAllWUsers(MI, ST, MRI)) {
LLVM_DEBUG(dbgs() << "Replacing " << MI);
MI.setDesc(TII.get(WOpc));
MI.clearFlag(MachineInstr::MIFlag::NoSWrap);
MI.clearFlag(MachineInstr::MIFlag::NoUWrap);
MI.clearFlag(MachineInstr::MIFlag::IsExact);
LLVM_DEBUG(dbgs() << " with " << MI);
++NumTransformedToWInstrs;
MadeChange = true;
}
}
}
return MadeChange;
}
bool RISCVOptWInstrs::runOnMachineFunction(MachineFunction &MF) {
if (skipFunction(MF.getFunction()))
return false;
MachineRegisterInfo &MRI = MF.getRegInfo();
const RISCVSubtarget &ST = MF.getSubtarget<RISCVSubtarget>();
const RISCVInstrInfo &TII = *ST.getInstrInfo();
if (!ST.is64Bit())
return false;
bool MadeChange = false;
MadeChange |= removeSExtWInstrs(MF, TII, ST, MRI);
if (!(DisableStripWSuffix || ST.preferWInst()))
MadeChange |= stripWSuffixes(MF, TII, ST, MRI);
if (ST.preferWInst())
MadeChange |= appendWSuffixes(MF, TII, ST, MRI);
return MadeChange;
}