Skip to content

Commit e3965f7

Browse files
arcbbbarcbbb
committed
[RISCV][CostModel][NFC] Add getRISCVInstructionCost() to TTI for CostKind
Instruction cost for CodeSize and Latency/RecipThroughput can be very differnet. Considering the diversity of CostKind and vendor-specific cost, and how they are spread across various TTI functions, it's becoming quite a challenge to handle. This patch adds an interface getRISCVInstructionCost to address it.
1 parent d1652ff commit e3965f7

File tree

2 files changed

+108
-12
lines changed

2 files changed

+108
-12
lines changed

‎llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp

Lines changed: 90 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -34,6 +34,51 @@ static cl::opt<unsigned> SLPMaxVF(
3434
"exclusively by SLP vectorizer."),
3535
cl::Hidden);
3636

37+
InstructionCost
38+
RISCVTTIImpl::getRISCVInstructionCost(RISCVInstruction Inst, MVT VT,
39+
unsigned NumInstr,
40+
TTI::TargetCostKind CostKind) {
41+
if (CostKind == TTI::TCK_CodeSize)
42+
return NumInstr;
43+
44+
InstructionCost LMUL = TLI->getLMULCost(VT);
45+
InstructionCost Cost = LMUL * NumInstr;
46+
47+
if ((CostKind == TTI::TCK_RecipThroughput) ||
48+
(CostKind == TTI::TCK_Latency)) {
49+
switch (Inst) {
50+
case RISCVInstruction::VRGATHER_VI:
51+
return NumInstr * TLI->getVRGatherVICost(VT);
52+
case RISCVInstruction::VRGATHER_VV:
53+
return NumInstr * TLI->getVRGatherVVCost(VT);
54+
case RISCVInstruction::VSLIDE:
55+
return NumInstr * TLI->getVSlideCost(VT);
56+
case RISCVInstruction::VSIMPLE_INT_RED:
57+
case RISCVInstruction::VMINMAX_INTFP_RED:
58+
case RISCVInstruction::VUNORD_FP_RED: {
59+
unsigned VL = VT.getVectorMinNumElements();
60+
if (!VT.isFixedLengthVector()) {
61+
VL *= *getVScaleForTuning();
62+
}
63+
return Log2_32_Ceil(VL);
64+
}
65+
case RISCVInstruction::VORD_FP_RED: {
66+
unsigned VL = VT.getVectorMinNumElements();
67+
if (!VT.isFixedLengthVector()) {
68+
VL *= *getVScaleForTuning();
69+
}
70+
return VL;
71+
}
72+
case RISCVInstruction::VMERGE:
73+
case RISCVInstruction::VMV:
74+
case RISCVInstruction::VSIMPLE_INT:
75+
case RISCVInstruction::VNARROWING:
76+
return Cost;
77+
}
78+
}
79+
return Cost;
80+
}
81+
3782
InstructionCost RISCVTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
3883
TTI::TargetCostKind CostKind) {
3984
assert(Ty->isIntegerTy() &&
@@ -279,7 +324,9 @@ InstructionCost RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
279324
// Example sequence:
280325
// vnsrl.wi v10, v8, 0
281326
if (equal(DeinterleaveMask, Mask))
282-
return LT.first * TLI->getLMULCost(LT.second);
327+
return LT.first *
328+
getRISCVInstructionCost(RISCVInstruction::VNARROWING,
329+
LT.second, 1, CostKind);
283330
}
284331
}
285332
}
@@ -290,7 +337,9 @@ InstructionCost RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
290337
LT.second.getVectorNumElements() <= 256)) {
291338
VectorType *IdxTy = getVRGatherIndexType(LT.second, *ST, Tp->getContext());
292339
InstructionCost IndexCost = getConstantPoolLoadCost(IdxTy, CostKind);
293-
return IndexCost + TLI->getVRGatherVVCost(LT.second);
340+
return IndexCost +
341+
getRISCVInstructionCost(RISCVInstruction::VRGATHER_VV, LT.second,
342+
1, CostKind);
294343
}
295344
[[fallthrough]];
296345
}
@@ -308,7 +357,10 @@ InstructionCost RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
308357
VectorType *MaskTy = VectorType::get(IntegerType::getInt1Ty(C), EC);
309358
InstructionCost IndexCost = getConstantPoolLoadCost(IdxTy, CostKind);
310359
InstructionCost MaskCost = getConstantPoolLoadCost(MaskTy, CostKind);
311-
return 2 * IndexCost + 2 * TLI->getVRGatherVVCost(LT.second) + MaskCost;
360+
return 2 * IndexCost +
361+
getRISCVInstructionCost(RISCVInstruction::VRGATHER_VV, LT.second,
362+
2, CostKind) +
363+
MaskCost;
312364
}
313365
[[fallthrough]];
314366
}
@@ -363,19 +415,26 @@ InstructionCost RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
363415
// Example sequence:
364416
// vsetivli zero, 4, e8, mf2, tu, ma (ignored)
365417
// vslidedown.vi v8, v9, 2
366-
return LT.first * TLI->getVSlideCost(LT.second);
418+
return LT.first * getRISCVInstructionCost(RISCVInstruction::VSLIDE,
419+
LT.second, 1, CostKind);
367420
case TTI::SK_InsertSubvector:
368421
// Example sequence:
369422
// vsetivli zero, 4, e8, mf2, tu, ma (ignored)
370423
// vslideup.vi v8, v9, 2
371-
return LT.first * TLI->getVSlideCost(LT.second);
424+
return LT.first * getRISCVInstructionCost(RISCVInstruction::VSLIDE,
425+
LT.second, 1, CostKind);
372426
case TTI::SK_Select: {
373427
// Example sequence:
374428
// li a0, 90
375429
// vsetivli zero, 8, e8, mf2, ta, ma (ignored)
376430
// vmv.s.x v0, a0
377431
// vmerge.vvm v8, v9, v8, v0
378-
return LT.first * 3 * TLI->getLMULCost(LT.second);
432+
return LT.first *
433+
(TLI->getLMULCost(LT.second) + // FIXME: should be 1 for li
434+
getRISCVInstructionCost(RISCVInstruction::VMV, LT.second, 1,
435+
CostKind) +
436+
getRISCVInstructionCost(RISCVInstruction::VMERGE, LT.second, 1,
437+
CostKind));
379438
}
380439
case TTI::SK_Broadcast: {
381440
bool HasScalar = (Args.size() > 0) && (Operator::getOpcode(Args[0]) ==
@@ -387,7 +446,12 @@ InstructionCost RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
387446
// vsetivli zero, 2, e8, mf8, ta, ma (ignored)
388447
// vmv.v.x v8, a0
389448
// vmsne.vi v0, v8, 0
390-
return LT.first * TLI->getLMULCost(LT.second) * 3;
449+
return LT.first *
450+
(TLI->getLMULCost(LT.second) + // FIXME: should be 1 for andi
451+
getRISCVInstructionCost(RISCVInstruction::VMV, LT.second, 1,
452+
CostKind) +
453+
getRISCVInstructionCost(RISCVInstruction::VSIMPLE_INT,
454+
LT.second, 1, CostKind));
391455
}
392456
// Example sequence:
393457
// vsetivli zero, 2, e8, mf8, ta, mu (ignored)
@@ -398,24 +462,34 @@ InstructionCost RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
398462
// vmv.v.x v8, a0
399463
// vmsne.vi v0, v8, 0
400464

401-
return LT.first * TLI->getLMULCost(LT.second) * 6;
465+
return LT.first *
466+
(TLI->getLMULCost(LT.second) + // FIXME: this should be 1 for andi
467+
getRISCVInstructionCost(RISCVInstruction::VMV, LT.second, 3,
468+
CostKind) +
469+
getRISCVInstructionCost(RISCVInstruction::VMERGE, LT.second, 1,
470+
CostKind) +
471+
getRISCVInstructionCost(RISCVInstruction::VSIMPLE_INT, LT.second,
472+
1, CostKind));
402473
}
403474

404475
if (HasScalar) {
405476
// Example sequence:
406477
// vmv.v.x v8, a0
407-
return LT.first * TLI->getLMULCost(LT.second);
478+
return LT.first * getRISCVInstructionCost(RISCVInstruction::VMV,
479+
LT.second, 1, CostKind);
408480
}
409481

410482
// Example sequence:
411483
// vrgather.vi v9, v8, 0
412-
return LT.first * TLI->getVRGatherVICost(LT.second);
484+
return LT.first * getRISCVInstructionCost(RISCVInstruction::VRGATHER_VI,
485+
LT.second, 1, CostKind);
413486
}
414487
case TTI::SK_Splice:
415488
// vslidedown+vslideup.
416489
// TODO: Multiplying by LT.first implies this legalizes into multiple copies
417490
// of similar code, but I think we expand through memory.
418-
return 2 * LT.first * TLI->getVSlideCost(LT.second);
491+
return LT.first * getRISCVInstructionCost(RISCVInstruction::VSLIDE,
492+
LT.second, 2, CostKind);
419493
case TTI::SK_Reverse: {
420494
// TODO: Cases to improve here:
421495
// * Illegal vector types
@@ -435,7 +509,11 @@ InstructionCost RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
435509
if (LT.second.isFixedLengthVector())
436510
// vrsub.vi has a 5 bit immediate field, otherwise an li suffices
437511
LenCost = isInt<5>(LT.second.getVectorNumElements() - 1) ? 0 : 1;
438-
InstructionCost GatherCost = 2 + TLI->getVRGatherVVCost(LT.second);
512+
// FIXME: replace the constant `2` below with cost of VSIMPLE_INT (vid.v &
513+
// vrsub.vx)
514+
InstructionCost GatherCost =
515+
2 + getRISCVInstructionCost(RISCVInstruction::VRGATHER_VV, LT.second, 1,
516+
CostKind);
439517
// Mask operation additionally required extend and truncate
440518
InstructionCost ExtendCost = Tp->getElementType()->isIntegerTy(1) ? 3 : 0;
441519
return LT.first * (LenCost + GatherCost + ExtendCost);

‎llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h

Lines changed: 18 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -48,6 +48,24 @@ class RISCVTTIImpl : public BasicTTIImplBase<RISCVTTIImpl> {
4848
/// actual target hardware.
4949
unsigned getEstimatedVLFor(VectorType *Ty);
5050

51+
enum class RISCVInstruction {
52+
VRGATHER_VV,
53+
VRGATHER_VI,
54+
VSLIDE,
55+
VMERGE,
56+
VMV,
57+
VSIMPLE_INT, // ICMP
58+
VNARROWING, // VNSRL
59+
VSIMPLE_INT_RED, // VREDSUM, VREDAND, VREDOR, VREDXOR
60+
VMINMAX_INTFP_RED, // VREDMAX, VREDMAXU, VREDMIN, VREDMINU
61+
VUNORD_FP_RED, // VFREDUSUM, VFREDMAX, VFREDMIN
62+
VORD_FP_RED, // VFREDOSUM
63+
};
64+
65+
InstructionCost getRISCVInstructionCost(RISCVInstruction Inst, MVT VT,
66+
unsigned NumInstr,
67+
TTI::TargetCostKind CostKind);
68+
5169
/// Return the cost of accessing a constant pool entry of the specified
5270
/// type.
5371
InstructionCost getConstantPoolLoadCost(Type *Ty,

0 commit comments

Comments
 (0)