xref: /freebsd/contrib/llvm-project/llvm/lib/Target/NVPTX/NVPTXTargetTransformInfo.cpp (revision a90b9d0159070121c221b966469c3e36d912bf82)
1 //===-- NVPTXTargetTransformInfo.cpp - NVPTX specific TTI -----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include "NVPTXTargetTransformInfo.h"
10 #include "NVPTXUtilities.h"
11 #include "llvm/Analysis/LoopInfo.h"
12 #include "llvm/Analysis/TargetTransformInfo.h"
13 #include "llvm/Analysis/ValueTracking.h"
14 #include "llvm/CodeGen/BasicTTIImpl.h"
15 #include "llvm/CodeGen/CostTable.h"
16 #include "llvm/CodeGen/TargetLowering.h"
17 #include "llvm/IR/IntrinsicsNVPTX.h"
18 #include "llvm/Support/Debug.h"
19 #include <optional>
20 using namespace llvm;
21 
22 #define DEBUG_TYPE "NVPTXtti"
23 
24 // Whether the given intrinsic reads threadIdx.x/y/z.
25 static bool readsThreadIndex(const IntrinsicInst *II) {
26   switch (II->getIntrinsicID()) {
27     default: return false;
28     case Intrinsic::nvvm_read_ptx_sreg_tid_x:
29     case Intrinsic::nvvm_read_ptx_sreg_tid_y:
30     case Intrinsic::nvvm_read_ptx_sreg_tid_z:
31       return true;
32   }
33 }
34 
35 static bool readsLaneId(const IntrinsicInst *II) {
36   return II->getIntrinsicID() == Intrinsic::nvvm_read_ptx_sreg_laneid;
37 }
38 
39 // Whether the given intrinsic is an atomic instruction in PTX.
40 static bool isNVVMAtomic(const IntrinsicInst *II) {
41   switch (II->getIntrinsicID()) {
42     default: return false;
43     case Intrinsic::nvvm_atomic_load_inc_32:
44     case Intrinsic::nvvm_atomic_load_dec_32:
45 
46     case Intrinsic::nvvm_atomic_add_gen_f_cta:
47     case Intrinsic::nvvm_atomic_add_gen_f_sys:
48     case Intrinsic::nvvm_atomic_add_gen_i_cta:
49     case Intrinsic::nvvm_atomic_add_gen_i_sys:
50     case Intrinsic::nvvm_atomic_and_gen_i_cta:
51     case Intrinsic::nvvm_atomic_and_gen_i_sys:
52     case Intrinsic::nvvm_atomic_cas_gen_i_cta:
53     case Intrinsic::nvvm_atomic_cas_gen_i_sys:
54     case Intrinsic::nvvm_atomic_dec_gen_i_cta:
55     case Intrinsic::nvvm_atomic_dec_gen_i_sys:
56     case Intrinsic::nvvm_atomic_inc_gen_i_cta:
57     case Intrinsic::nvvm_atomic_inc_gen_i_sys:
58     case Intrinsic::nvvm_atomic_max_gen_i_cta:
59     case Intrinsic::nvvm_atomic_max_gen_i_sys:
60     case Intrinsic::nvvm_atomic_min_gen_i_cta:
61     case Intrinsic::nvvm_atomic_min_gen_i_sys:
62     case Intrinsic::nvvm_atomic_or_gen_i_cta:
63     case Intrinsic::nvvm_atomic_or_gen_i_sys:
64     case Intrinsic::nvvm_atomic_exch_gen_i_cta:
65     case Intrinsic::nvvm_atomic_exch_gen_i_sys:
66     case Intrinsic::nvvm_atomic_xor_gen_i_cta:
67     case Intrinsic::nvvm_atomic_xor_gen_i_sys:
68       return true;
69   }
70 }
71 
72 bool NVPTXTTIImpl::isSourceOfDivergence(const Value *V) {
73   // Without inter-procedural analysis, we conservatively assume that arguments
74   // to __device__ functions are divergent.
75   if (const Argument *Arg = dyn_cast<Argument>(V))
76     return !isKernelFunction(*Arg->getParent());
77 
78   if (const Instruction *I = dyn_cast<Instruction>(V)) {
79     // Without pointer analysis, we conservatively assume values loaded from
80     // generic or local address space are divergent.
81     if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
82       unsigned AS = LI->getPointerAddressSpace();
83       return AS == ADDRESS_SPACE_GENERIC || AS == ADDRESS_SPACE_LOCAL;
84     }
85     // Atomic instructions may cause divergence. Atomic instructions are
86     // executed sequentially across all threads in a warp. Therefore, an earlier
87     // executed thread may see different memory inputs than a later executed
88     // thread. For example, suppose *a = 0 initially.
89     //
90     //   atom.global.add.s32 d, [a], 1
91     //
92     // returns 0 for the first thread that enters the critical region, and 1 for
93     // the second thread.
94     if (I->isAtomic())
95       return true;
96     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
97       // Instructions that read threadIdx are obviously divergent.
98       if (readsThreadIndex(II) || readsLaneId(II))
99         return true;
100       // Handle the NVPTX atomic intrinsics that cannot be represented as an
101       // atomic IR instruction.
102       if (isNVVMAtomic(II))
103         return true;
104     }
105     // Conservatively consider the return value of function calls as divergent.
106     // We could analyze callees with bodies more precisely using
107     // inter-procedural analysis.
108     if (isa<CallInst>(I))
109       return true;
110   }
111 
112   return false;
113 }
114 
115 // Convert NVVM intrinsics to target-generic LLVM code where possible.
116 static Instruction *simplifyNvvmIntrinsic(IntrinsicInst *II, InstCombiner &IC) {
117   // Each NVVM intrinsic we can simplify can be replaced with one of:
118   //
119   //  * an LLVM intrinsic,
120   //  * an LLVM cast operation,
121   //  * an LLVM binary operation, or
122   //  * ad-hoc LLVM IR for the particular operation.
123 
124   // Some transformations are only valid when the module's
125   // flush-denormals-to-zero (ftz) setting is true/false, whereas other
126   // transformations are valid regardless of the module's ftz setting.
127   enum FtzRequirementTy {
128     FTZ_Any,       // Any ftz setting is ok.
129     FTZ_MustBeOn,  // Transformation is valid only if ftz is on.
130     FTZ_MustBeOff, // Transformation is valid only if ftz is off.
131   };
132   // Classes of NVVM intrinsics that can't be replaced one-to-one with a
133   // target-generic intrinsic, cast op, or binary op but that we can nonetheless
134   // simplify.
135   enum SpecialCase {
136     SPC_Reciprocal,
137   };
138 
139   // SimplifyAction is a poor-man's variant (plus an additional flag) that
140   // represents how to replace an NVVM intrinsic with target-generic LLVM IR.
141   struct SimplifyAction {
142     // Invariant: At most one of these Optionals has a value.
143     std::optional<Intrinsic::ID> IID;
144     std::optional<Instruction::CastOps> CastOp;
145     std::optional<Instruction::BinaryOps> BinaryOp;
146     std::optional<SpecialCase> Special;
147 
148     FtzRequirementTy FtzRequirement = FTZ_Any;
149     // Denormal handling is guarded by different attributes depending on the
150     // type (denormal-fp-math vs denormal-fp-math-f32), take note of halfs.
151     bool IsHalfTy = false;
152 
153     SimplifyAction() = default;
154 
155     SimplifyAction(Intrinsic::ID IID, FtzRequirementTy FtzReq,
156                    bool IsHalfTy = false)
157         : IID(IID), FtzRequirement(FtzReq), IsHalfTy(IsHalfTy) {}
158 
159     // Cast operations don't have anything to do with FTZ, so we skip that
160     // argument.
161     SimplifyAction(Instruction::CastOps CastOp) : CastOp(CastOp) {}
162 
163     SimplifyAction(Instruction::BinaryOps BinaryOp, FtzRequirementTy FtzReq)
164         : BinaryOp(BinaryOp), FtzRequirement(FtzReq) {}
165 
166     SimplifyAction(SpecialCase Special, FtzRequirementTy FtzReq)
167         : Special(Special), FtzRequirement(FtzReq) {}
168   };
169 
170   // Try to generate a SimplifyAction describing how to replace our
171   // IntrinsicInstr with target-generic LLVM IR.
172   const SimplifyAction Action = [II]() -> SimplifyAction {
173     switch (II->getIntrinsicID()) {
174     // NVVM intrinsics that map directly to LLVM intrinsics.
175     case Intrinsic::nvvm_ceil_d:
176       return {Intrinsic::ceil, FTZ_Any};
177     case Intrinsic::nvvm_ceil_f:
178       return {Intrinsic::ceil, FTZ_MustBeOff};
179     case Intrinsic::nvvm_ceil_ftz_f:
180       return {Intrinsic::ceil, FTZ_MustBeOn};
181     case Intrinsic::nvvm_fabs_d:
182       return {Intrinsic::fabs, FTZ_Any};
183     case Intrinsic::nvvm_floor_d:
184       return {Intrinsic::floor, FTZ_Any};
185     case Intrinsic::nvvm_floor_f:
186       return {Intrinsic::floor, FTZ_MustBeOff};
187     case Intrinsic::nvvm_floor_ftz_f:
188       return {Intrinsic::floor, FTZ_MustBeOn};
189     case Intrinsic::nvvm_fma_rn_d:
190       return {Intrinsic::fma, FTZ_Any};
191     case Intrinsic::nvvm_fma_rn_f:
192       return {Intrinsic::fma, FTZ_MustBeOff};
193     case Intrinsic::nvvm_fma_rn_ftz_f:
194       return {Intrinsic::fma, FTZ_MustBeOn};
195     case Intrinsic::nvvm_fma_rn_f16:
196       return {Intrinsic::fma, FTZ_MustBeOff, true};
197     case Intrinsic::nvvm_fma_rn_ftz_f16:
198       return {Intrinsic::fma, FTZ_MustBeOn, true};
199     case Intrinsic::nvvm_fma_rn_f16x2:
200       return {Intrinsic::fma, FTZ_MustBeOff, true};
201     case Intrinsic::nvvm_fma_rn_ftz_f16x2:
202       return {Intrinsic::fma, FTZ_MustBeOn, true};
203     case Intrinsic::nvvm_fma_rn_bf16:
204       return {Intrinsic::fma, FTZ_MustBeOff, true};
205     case Intrinsic::nvvm_fma_rn_ftz_bf16:
206       return {Intrinsic::fma, FTZ_MustBeOn, true};
207     case Intrinsic::nvvm_fma_rn_bf16x2:
208       return {Intrinsic::fma, FTZ_MustBeOff, true};
209     case Intrinsic::nvvm_fma_rn_ftz_bf16x2:
210       return {Intrinsic::fma, FTZ_MustBeOn, true};
211     case Intrinsic::nvvm_fmax_d:
212       return {Intrinsic::maxnum, FTZ_Any};
213     case Intrinsic::nvvm_fmax_f:
214       return {Intrinsic::maxnum, FTZ_MustBeOff};
215     case Intrinsic::nvvm_fmax_ftz_f:
216       return {Intrinsic::maxnum, FTZ_MustBeOn};
217     case Intrinsic::nvvm_fmax_nan_f:
218       return {Intrinsic::maximum, FTZ_MustBeOff};
219     case Intrinsic::nvvm_fmax_ftz_nan_f:
220       return {Intrinsic::maximum, FTZ_MustBeOn};
221     case Intrinsic::nvvm_fmax_f16:
222       return {Intrinsic::maxnum, FTZ_MustBeOff, true};
223     case Intrinsic::nvvm_fmax_ftz_f16:
224       return {Intrinsic::maxnum, FTZ_MustBeOn, true};
225     case Intrinsic::nvvm_fmax_f16x2:
226       return {Intrinsic::maxnum, FTZ_MustBeOff, true};
227     case Intrinsic::nvvm_fmax_ftz_f16x2:
228       return {Intrinsic::maxnum, FTZ_MustBeOn, true};
229     case Intrinsic::nvvm_fmax_nan_f16:
230       return {Intrinsic::maximum, FTZ_MustBeOff, true};
231     case Intrinsic::nvvm_fmax_ftz_nan_f16:
232       return {Intrinsic::maximum, FTZ_MustBeOn, true};
233     case Intrinsic::nvvm_fmax_nan_f16x2:
234       return {Intrinsic::maximum, FTZ_MustBeOff, true};
235     case Intrinsic::nvvm_fmax_ftz_nan_f16x2:
236       return {Intrinsic::maximum, FTZ_MustBeOn, true};
237     case Intrinsic::nvvm_fmin_d:
238       return {Intrinsic::minnum, FTZ_Any};
239     case Intrinsic::nvvm_fmin_f:
240       return {Intrinsic::minnum, FTZ_MustBeOff};
241     case Intrinsic::nvvm_fmin_ftz_f:
242       return {Intrinsic::minnum, FTZ_MustBeOn};
243     case Intrinsic::nvvm_fmin_nan_f:
244       return {Intrinsic::minimum, FTZ_MustBeOff};
245     case Intrinsic::nvvm_fmin_ftz_nan_f:
246       return {Intrinsic::minimum, FTZ_MustBeOn};
247     case Intrinsic::nvvm_fmin_f16:
248       return {Intrinsic::minnum, FTZ_MustBeOff, true};
249     case Intrinsic::nvvm_fmin_ftz_f16:
250       return {Intrinsic::minnum, FTZ_MustBeOn, true};
251     case Intrinsic::nvvm_fmin_f16x2:
252       return {Intrinsic::minnum, FTZ_MustBeOff, true};
253     case Intrinsic::nvvm_fmin_ftz_f16x2:
254       return {Intrinsic::minnum, FTZ_MustBeOn, true};
255     case Intrinsic::nvvm_fmin_nan_f16:
256       return {Intrinsic::minimum, FTZ_MustBeOff, true};
257     case Intrinsic::nvvm_fmin_ftz_nan_f16:
258       return {Intrinsic::minimum, FTZ_MustBeOn, true};
259     case Intrinsic::nvvm_fmin_nan_f16x2:
260       return {Intrinsic::minimum, FTZ_MustBeOff, true};
261     case Intrinsic::nvvm_fmin_ftz_nan_f16x2:
262       return {Intrinsic::minimum, FTZ_MustBeOn, true};
263     case Intrinsic::nvvm_sqrt_rn_d:
264       return {Intrinsic::sqrt, FTZ_Any};
265     case Intrinsic::nvvm_sqrt_f:
266       // nvvm_sqrt_f is a special case.  For  most intrinsics, foo_ftz_f is the
267       // ftz version, and foo_f is the non-ftz version.  But nvvm_sqrt_f adopts
268       // the ftz-ness of the surrounding code.  sqrt_rn_f and sqrt_rn_ftz_f are
269       // the versions with explicit ftz-ness.
270       return {Intrinsic::sqrt, FTZ_Any};
271     case Intrinsic::nvvm_trunc_d:
272       return {Intrinsic::trunc, FTZ_Any};
273     case Intrinsic::nvvm_trunc_f:
274       return {Intrinsic::trunc, FTZ_MustBeOff};
275     case Intrinsic::nvvm_trunc_ftz_f:
276       return {Intrinsic::trunc, FTZ_MustBeOn};
277 
278     // NVVM intrinsics that map to LLVM cast operations.
279     //
280     // Note that llvm's target-generic conversion operators correspond to the rz
281     // (round to zero) versions of the nvvm conversion intrinsics, even though
282     // most everything else here uses the rn (round to nearest even) nvvm ops.
283     case Intrinsic::nvvm_d2i_rz:
284     case Intrinsic::nvvm_f2i_rz:
285     case Intrinsic::nvvm_d2ll_rz:
286     case Intrinsic::nvvm_f2ll_rz:
287       return {Instruction::FPToSI};
288     case Intrinsic::nvvm_d2ui_rz:
289     case Intrinsic::nvvm_f2ui_rz:
290     case Intrinsic::nvvm_d2ull_rz:
291     case Intrinsic::nvvm_f2ull_rz:
292       return {Instruction::FPToUI};
293     case Intrinsic::nvvm_i2d_rz:
294     case Intrinsic::nvvm_i2f_rz:
295     case Intrinsic::nvvm_ll2d_rz:
296     case Intrinsic::nvvm_ll2f_rz:
297       return {Instruction::SIToFP};
298     case Intrinsic::nvvm_ui2d_rz:
299     case Intrinsic::nvvm_ui2f_rz:
300     case Intrinsic::nvvm_ull2d_rz:
301     case Intrinsic::nvvm_ull2f_rz:
302       return {Instruction::UIToFP};
303 
304     // NVVM intrinsics that map to LLVM binary ops.
305     case Intrinsic::nvvm_div_rn_d:
306       return {Instruction::FDiv, FTZ_Any};
307 
308     // The remainder of cases are NVVM intrinsics that map to LLVM idioms, but
309     // need special handling.
310     //
311     // We seem to be missing intrinsics for rcp.approx.{ftz.}f32, which is just
312     // as well.
313     case Intrinsic::nvvm_rcp_rn_d:
314       return {SPC_Reciprocal, FTZ_Any};
315 
316       // We do not currently simplify intrinsics that give an approximate
317       // answer. These include:
318       //
319       //   - nvvm_cos_approx_{f,ftz_f}
320       //   - nvvm_ex2_approx_{d,f,ftz_f}
321       //   - nvvm_lg2_approx_{d,f,ftz_f}
322       //   - nvvm_sin_approx_{f,ftz_f}
323       //   - nvvm_sqrt_approx_{f,ftz_f}
324       //   - nvvm_rsqrt_approx_{d,f,ftz_f}
325       //   - nvvm_div_approx_{ftz_d,ftz_f,f}
326       //   - nvvm_rcp_approx_ftz_d
327       //
328       // Ideally we'd encode them as e.g. "fast call @llvm.cos", where "fast"
329       // means that fastmath is enabled in the intrinsic.  Unfortunately only
330       // binary operators (currently) have a fastmath bit in SelectionDAG, so
331       // this information gets lost and we can't select on it.
332       //
333       // TODO: div and rcp are lowered to a binary op, so these we could in
334       // theory lower them to "fast fdiv".
335 
336     default:
337       return {};
338     }
339   }();
340 
341   // If Action.FtzRequirementTy is not satisfied by the module's ftz state, we
342   // can bail out now.  (Notice that in the case that IID is not an NVVM
343   // intrinsic, we don't have to look up any module metadata, as
344   // FtzRequirementTy will be FTZ_Any.)
345   if (Action.FtzRequirement != FTZ_Any) {
346     // FIXME: Broken for f64
347     DenormalMode Mode = II->getFunction()->getDenormalMode(
348         Action.IsHalfTy ? APFloat::IEEEhalf() : APFloat::IEEEsingle());
349     bool FtzEnabled = Mode.Output == DenormalMode::PreserveSign;
350 
351     if (FtzEnabled != (Action.FtzRequirement == FTZ_MustBeOn))
352       return nullptr;
353   }
354 
355   // Simplify to target-generic intrinsic.
356   if (Action.IID) {
357     SmallVector<Value *, 4> Args(II->args());
358     // All the target-generic intrinsics currently of interest to us have one
359     // type argument, equal to that of the nvvm intrinsic's argument.
360     Type *Tys[] = {II->getArgOperand(0)->getType()};
361     return CallInst::Create(
362         Intrinsic::getDeclaration(II->getModule(), *Action.IID, Tys), Args);
363   }
364 
365   // Simplify to target-generic binary op.
366   if (Action.BinaryOp)
367     return BinaryOperator::Create(*Action.BinaryOp, II->getArgOperand(0),
368                                   II->getArgOperand(1), II->getName());
369 
370   // Simplify to target-generic cast op.
371   if (Action.CastOp)
372     return CastInst::Create(*Action.CastOp, II->getArgOperand(0), II->getType(),
373                             II->getName());
374 
375   // All that's left are the special cases.
376   if (!Action.Special)
377     return nullptr;
378 
379   switch (*Action.Special) {
380   case SPC_Reciprocal:
381     // Simplify reciprocal.
382     return BinaryOperator::Create(
383         Instruction::FDiv, ConstantFP::get(II->getArgOperand(0)->getType(), 1),
384         II->getArgOperand(0), II->getName());
385   }
386   llvm_unreachable("All SpecialCase enumerators should be handled in switch.");
387 }
388 
389 std::optional<Instruction *>
390 NVPTXTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
391   if (Instruction *I = simplifyNvvmIntrinsic(&II, IC)) {
392     return I;
393   }
394   return std::nullopt;
395 }
396 
397 InstructionCost NVPTXTTIImpl::getArithmeticInstrCost(
398     unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
399     TTI::OperandValueInfo Op1Info, TTI::OperandValueInfo Op2Info,
400     ArrayRef<const Value *> Args,
401     const Instruction *CxtI) {
402   // Legalize the type.
403   std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
404 
405   int ISD = TLI->InstructionOpcodeToISD(Opcode);
406 
407   switch (ISD) {
408   default:
409     return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
410                                          Op2Info);
411   case ISD::ADD:
412   case ISD::MUL:
413   case ISD::XOR:
414   case ISD::OR:
415   case ISD::AND:
416     // The machine code (SASS) simulates an i64 with two i32. Therefore, we
417     // estimate that arithmetic operations on i64 are twice as expensive as
418     // those on types that can fit into one machine register.
419     if (LT.second.SimpleTy == MVT::i64)
420       return 2 * LT.first;
421     // Delegate other cases to the basic TTI.
422     return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
423                                          Op2Info);
424   }
425 }
426 
427 void NVPTXTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
428                                            TTI::UnrollingPreferences &UP,
429                                            OptimizationRemarkEmitter *ORE) {
430   BaseT::getUnrollingPreferences(L, SE, UP, ORE);
431 
432   // Enable partial unrolling and runtime unrolling, but reduce the
433   // threshold.  This partially unrolls small loops which are often
434   // unrolled by the PTX to SASS compiler and unrolling earlier can be
435   // beneficial.
436   UP.Partial = UP.Runtime = true;
437   UP.PartialThreshold = UP.Threshold / 4;
438 }
439 
440 void NVPTXTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
441                                          TTI::PeelingPreferences &PP) {
442   BaseT::getPeelingPreferences(L, SE, PP);
443 }
444