1 //===--------------------- ResourceManager.h --------------------*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 /// \file
10 ///
11 /// The classes here represent processor resource units and their management
12 /// strategy.  These classes are managed by the Scheduler.
13 ///
14 //===----------------------------------------------------------------------===//
15 
16 #ifndef LLVM_MCA_RESOURCE_MANAGER_H
17 #define LLVM_MCA_RESOURCE_MANAGER_H
18 
19 #include "llvm/ADT/ArrayRef.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/MC/MCSchedule.h"
23 #include "llvm/MCA/Instruction.h"
24 #include "llvm/MCA/Support.h"
25 
26 namespace llvm {
27 namespace mca {
28 
29 /// Used to notify the internal state of a processor resource.
30 ///
31 /// A processor resource is available if it is not reserved, and there are
32 /// available slots in the buffer.  A processor resource is unavailable if it
33 /// is either reserved, or the associated buffer is full. A processor resource
34 /// with a buffer size of -1 is always available if it is not reserved.
35 ///
36 /// Values of type ResourceStateEvent are returned by method
37 /// ResourceState::isBufferAvailable(), which is used to query the internal
38 /// state of a resource.
39 ///
40 /// The naming convention for resource state events is:
41 ///  * Event names start with prefix RS_
42 ///  * Prefix RS_ is followed by a string describing the actual resource state.
43 enum ResourceStateEvent {
44   RS_BUFFER_AVAILABLE,
45   RS_BUFFER_UNAVAILABLE,
46   RS_RESERVED
47 };
48 
49 /// Resource allocation strategy used by hardware scheduler resources.
50 class ResourceStrategy {
51   ResourceStrategy(const ResourceStrategy &) = delete;
52   ResourceStrategy &operator=(const ResourceStrategy &) = delete;
53 
54 public:
ResourceStrategy()55   ResourceStrategy() {}
56   virtual ~ResourceStrategy();
57 
58   /// Selects a processor resource unit from a ReadyMask.
59   virtual uint64_t select(uint64_t ReadyMask) = 0;
60 
61   /// Called by the ResourceManager when a processor resource group, or a
62   /// processor resource with multiple units has become unavailable.
63   ///
64   /// The default strategy uses this information to bias its selection logic.
used(uint64_t ResourceMask)65   virtual void used(uint64_t ResourceMask) {}
66 };
67 
68 /// Default resource allocation strategy used by processor resource groups and
69 /// processor resources with multiple units.
70 class DefaultResourceStrategy final : public ResourceStrategy {
71   /// A Mask of resource unit identifiers.
72   ///
73   /// There is one bit set for every available resource unit.
74   /// It defaults to the value of field ResourceSizeMask in ResourceState.
75   const uint64_t ResourceUnitMask;
76 
77   /// A simple round-robin selector for processor resource units.
78   /// Each bit of this mask identifies a sub resource within a group.
79   ///
80   /// As an example, lets assume that this is a default policy for a
81   /// processor resource group composed by the following three units:
82   ///   ResourceA -- 0b001
83   ///   ResourceB -- 0b010
84   ///   ResourceC -- 0b100
85   ///
86   /// Field NextInSequenceMask is used to select the next unit from the set of
87   /// resource units. It defaults to the value of field `ResourceUnitMasks` (in
88   /// this example, it defaults to mask '0b111').
89   ///
90   /// The round-robin selector would firstly select 'ResourceC', then
91   /// 'ResourceB', and eventually 'ResourceA'.  When a resource R is used, the
92   /// corresponding bit in NextInSequenceMask is cleared.  For example, if
93   /// 'ResourceC' is selected, then the new value of NextInSequenceMask becomes
94   /// 0xb011.
95   ///
96   /// When NextInSequenceMask becomes zero, it is automatically reset to the
97   /// default value (i.e. ResourceUnitMask).
98   uint64_t NextInSequenceMask;
99 
100   /// This field is used to track resource units that are used (i.e. selected)
101   /// by other groups other than the one associated with this strategy object.
102   ///
103   /// In LLVM processor resource groups are allowed to partially (or fully)
104   /// overlap. That means, a same unit may be visible to multiple groups.
105   /// This field keeps track of uses that have originated from outside of
106   /// this group. The idea is to bias the selection strategy, so that resources
107   /// that haven't been used by other groups get prioritized.
108   ///
109   /// The end goal is to (try to) keep the resource distribution as much uniform
110   /// as possible. By construction, this mask only tracks one-level of resource
111   /// usage. Therefore, this strategy is expected to be less accurate when same
112   /// units are used multiple times by other groups within a single round of
113   /// select.
114   ///
115   /// Note: an LRU selector would have a better accuracy at the cost of being
116   /// slightly more expensive (mostly in terms of runtime cost). Methods
117   /// 'select' and 'used', are always in the hot execution path of llvm-mca.
118   /// Therefore, a slow implementation of 'select' would have a negative impact
119   /// on the overall performance of the tool.
120   uint64_t RemovedFromNextInSequence;
121 
122 public:
DefaultResourceStrategy(uint64_t UnitMask)123   DefaultResourceStrategy(uint64_t UnitMask)
124       : ResourceStrategy(), ResourceUnitMask(UnitMask),
125         NextInSequenceMask(UnitMask), RemovedFromNextInSequence(0) {}
126   virtual ~DefaultResourceStrategy() = default;
127 
128   uint64_t select(uint64_t ReadyMask) override;
129   void used(uint64_t Mask) override;
130 };
131 
132 /// A processor resource descriptor.
133 ///
134 /// There is an instance of this class for every processor resource defined by
135 /// the machine scheduling model.
136 /// Objects of class ResourceState dynamically track the usage of processor
137 /// resource units.
138 class ResourceState {
139   /// An index to the MCProcResourceDesc entry in the processor model.
140   const unsigned ProcResourceDescIndex;
141   /// A resource mask. This is generated by the tool with the help of
142   /// function `mca::computeProcResourceMasks' (see Support.h).
143   ///
144   /// Field ResourceMask only has one bit set if this resource state describes a
145   /// processor resource unit (i.e. this is not a group). That means, we can
146   /// quickly check if a resource is a group by simply counting the number of
147   /// bits that are set in the mask.
148   ///
149   /// The most significant bit of a mask (MSB) uniquely identifies a resource.
150   /// Remaining bits are used to describe the composition of a group (Group).
151   ///
152   /// Example (little endian):
153   ///            Resource |  Mask      |  MSB       |  Group
154   ///            ---------+------------+------------+------------
155   ///            A        |  0b000001  |  0b000001  |  0b000000
156   ///                     |            |            |
157   ///            B        |  0b000010  |  0b000010  |  0b000000
158   ///                     |            |            |
159   ///            C        |  0b010000  |  0b010000  |  0b000000
160   ///                     |            |            |
161   ///            D        |  0b110010  |  0b100000  |  0b010010
162   ///
163   /// In this example, resources A, B and C are processor resource units.
164   /// Only resource D is a group resource, and it contains resources B and C.
165   /// That is because MSB(B) and MSB(C) are both contained within Group(D).
166   const uint64_t ResourceMask;
167 
168   /// A ProcResource can have multiple units.
169   ///
170   /// For processor resource groups this field is a mask of contained resource
171   /// units. It is obtained from ResourceMask by clearing the highest set bit.
172   /// The number of resource units in a group can be simply computed as the
173   /// population count of this field.
174   ///
175   /// For normal (i.e. non-group) resources, the number of bits set in this mask
176   /// is equivalent to the number of units declared by the processor model (see
177   /// field 'NumUnits' in 'ProcResourceUnits').
178   uint64_t ResourceSizeMask;
179 
180   /// A mask of ready units.
181   uint64_t ReadyMask;
182 
183   /// Buffered resources will have this field set to a positive number different
184   /// than zero. A buffered resource behaves like a reservation station
185   /// implementing its own buffer for out-of-order execution.
186   ///
187   /// A BufferSize of 1 is used by scheduler resources that force in-order
188   /// execution.
189   ///
190   /// A BufferSize of 0 is used to model in-order issue/dispatch resources.
191   /// Since in-order issue/dispatch resources don't implement buffers, dispatch
192   /// events coincide with issue events.
193   /// Also, no other instruction ca be dispatched/issue while this resource is
194   /// in use. Only when all the "resource cycles" are consumed (after the issue
195   /// event), a new instruction ca be dispatched.
196   const int BufferSize;
197 
198   /// Available slots in the buffer (zero, if this is not a buffered resource).
199   unsigned AvailableSlots;
200 
201   /// This field is set if this resource is currently reserved.
202   ///
203   /// Resources can be reserved for a number of cycles.
204   /// Instructions can still be dispatched to reserved resources. However,
205   /// istructions dispatched to a reserved resource cannot be issued to the
206   /// underlying units (i.e. pipelines) until the resource is released.
207   bool Unavailable;
208 
209   const bool IsAGroup;
210 
211   /// Checks for the availability of unit 'SubResMask' in the group.
isSubResourceReady(uint64_t SubResMask)212   bool isSubResourceReady(uint64_t SubResMask) const {
213     return ReadyMask & SubResMask;
214   }
215 
216 public:
217   ResourceState(const MCProcResourceDesc &Desc, unsigned Index, uint64_t Mask);
218 
getProcResourceID()219   unsigned getProcResourceID() const { return ProcResourceDescIndex; }
getResourceMask()220   uint64_t getResourceMask() const { return ResourceMask; }
getReadyMask()221   uint64_t getReadyMask() const { return ReadyMask; }
getBufferSize()222   int getBufferSize() const { return BufferSize; }
223 
isBuffered()224   bool isBuffered() const { return BufferSize > 0; }
isInOrder()225   bool isInOrder() const { return BufferSize == 1; }
226 
227   /// Returns true if this is an in-order dispatch/issue resource.
isADispatchHazard()228   bool isADispatchHazard() const { return BufferSize == 0; }
isReserved()229   bool isReserved() const { return Unavailable; }
230 
setReserved()231   void setReserved() { Unavailable = true; }
clearReserved()232   void clearReserved() { Unavailable = false; }
233 
234   /// Returs true if this resource is not reserved, and if there are at least
235   /// `NumUnits` available units.
236   bool isReady(unsigned NumUnits = 1) const;
237 
isAResourceGroup()238   bool isAResourceGroup() const { return IsAGroup; }
239 
containsResource(uint64_t ID)240   bool containsResource(uint64_t ID) const { return ResourceMask & ID; }
241 
markSubResourceAsUsed(uint64_t ID)242   void markSubResourceAsUsed(uint64_t ID) {
243     assert(isSubResourceReady(ID));
244     ReadyMask ^= ID;
245   }
246 
releaseSubResource(uint64_t ID)247   void releaseSubResource(uint64_t ID) {
248     assert(!isSubResourceReady(ID));
249     ReadyMask ^= ID;
250   }
251 
getNumUnits()252   unsigned getNumUnits() const {
253     return isAResourceGroup() ? 1U : countPopulation(ResourceSizeMask);
254   }
255 
256   /// Checks if there is an available slot in the resource buffer.
257   ///
258   /// Returns RS_BUFFER_AVAILABLE if this is not a buffered resource, or if
259   /// there is a slot available.
260   ///
261   /// Returns RS_RESERVED if this buffered resource is a dispatch hazard, and it
262   /// is reserved.
263   ///
264   /// Returns RS_BUFFER_UNAVAILABLE if there are no available slots.
265   ResourceStateEvent isBufferAvailable() const;
266 
267   /// Reserve a slot in the buffer.
reserveBuffer()268   void reserveBuffer() {
269     if (AvailableSlots)
270       AvailableSlots--;
271   }
272 
273   /// Release a slot in the buffer.
releaseBuffer()274   void releaseBuffer() {
275     if (BufferSize > 0)
276       AvailableSlots++;
277     assert(AvailableSlots <= static_cast<unsigned>(BufferSize));
278   }
279 
280 #ifndef NDEBUG
281   void dump() const;
282 #endif
283 };
284 
285 /// A resource unit identifier.
286 ///
287 /// This is used to identify a specific processor resource unit using a pair
288 /// of indices where the 'first' index is a processor resource mask, and the
289 /// 'second' index is an index for a "sub-resource" (i.e. unit).
290 typedef std::pair<uint64_t, uint64_t> ResourceRef;
291 
292 // First: a MCProcResourceDesc index identifying a buffered resource.
293 // Second: max number of buffer entries used in this resource.
294 typedef std::pair<unsigned, unsigned> BufferUsageEntry;
295 
296 /// A resource manager for processor resource units and groups.
297 ///
298 /// This class owns all the ResourceState objects, and it is responsible for
299 /// acting on requests from a Scheduler by updating the internal state of
300 /// ResourceState objects.
301 /// This class doesn't know about instruction itineraries and functional units.
302 /// In future, it can be extended to support itineraries too through the same
303 /// public interface.
304 class ResourceManager {
305   // Set of resources available on the subtarget.
306   //
307   // There is an instance of ResourceState for every resource declared by the
308   // target scheduling model.
309   //
310   // Elements of this vector are ordered by resource kind. In particular,
311   // resource units take precedence over resource groups.
312   //
313   // The index of a processor resource in this vector depends on the value of
314   // its mask (see the description of field ResourceState::ResourceMask).  In
315   // particular, it is computed as the position of the most significant bit set
316   // (MSB) in the mask plus one (since we want to ignore the invalid resource
317   // descriptor at index zero).
318   //
319   // Example (little endian):
320   //
321   //             Resource | Mask    |  MSB    | Index
322   //             ---------+---------+---------+-------
323   //                 A    | 0b00001 | 0b00001 |   1
324   //                      |         |         |
325   //                 B    | 0b00100 | 0b00100 |   3
326   //                      |         |         |
327   //                 C    | 0b10010 | 0b10000 |   5
328   //
329   //
330   // The same index is also used to address elements within vector `Strategies`
331   // and vector `Resource2Groups`.
332   std::vector<std::unique_ptr<ResourceState>> Resources;
333   std::vector<std::unique_ptr<ResourceStrategy>> Strategies;
334 
335   // Used to quickly identify groups that own a particular resource unit.
336   std::vector<uint64_t> Resource2Groups;
337 
338   // A table to map processor resource IDs to processor resource masks.
339   SmallVector<uint64_t, 8> ProcResID2Mask;
340 
341   // Keeps track of which resources are busy, and how many cycles are left
342   // before those become usable again.
343   SmallDenseMap<ResourceRef, unsigned> BusyResources;
344 
345   // Returns the actual resource unit that will be used.
346   ResourceRef selectPipe(uint64_t ResourceID);
347 
348   void use(const ResourceRef &RR);
349   void release(const ResourceRef &RR);
350 
351   unsigned getNumUnits(uint64_t ResourceID) const;
352 
353   // Overrides the selection strategy for the processor resource with the given
354   // mask.
355   void setCustomStrategyImpl(std::unique_ptr<ResourceStrategy> S,
356                              uint64_t ResourceMask);
357 
358 public:
359   ResourceManager(const MCSchedModel &SM);
360   virtual ~ResourceManager() = default;
361 
362   // Overrides the selection strategy for the resource at index ResourceID in
363   // the MCProcResourceDesc table.
setCustomStrategy(std::unique_ptr<ResourceStrategy> S,unsigned ResourceID)364   void setCustomStrategy(std::unique_ptr<ResourceStrategy> S,
365                          unsigned ResourceID) {
366     assert(ResourceID < ProcResID2Mask.size() &&
367            "Invalid resource index in input!");
368     return setCustomStrategyImpl(std::move(S), ProcResID2Mask[ResourceID]);
369   }
370 
371   // Returns RS_BUFFER_AVAILABLE if buffered resources are not reserved, and if
372   // there are enough available slots in the buffers.
373   ResourceStateEvent canBeDispatched(ArrayRef<uint64_t> Buffers) const;
374 
375   // Return the processor resource identifier associated to this Mask.
376   unsigned resolveResourceMask(uint64_t Mask) const;
377 
378   // Consume a slot in every buffered resource from array 'Buffers'. Resource
379   // units that are dispatch hazards (i.e. BufferSize=0) are marked as reserved.
380   void reserveBuffers(ArrayRef<uint64_t> Buffers);
381 
382   // Release buffer entries previously allocated by method reserveBuffers.
383   void releaseBuffers(ArrayRef<uint64_t> Buffers);
384 
385   // Reserve a processor resource. A reserved resource is not available for
386   // instruction issue until it is released.
387   void reserveResource(uint64_t ResourceID);
388 
389   // Release a previously reserved processor resource.
390   void releaseResource(uint64_t ResourceID);
391 
392   bool canBeIssued(const InstrDesc &Desc) const;
393 
394   void issueInstruction(
395       const InstrDesc &Desc,
396       SmallVectorImpl<std::pair<ResourceRef, ResourceCycles>> &Pipes);
397 
398   void cycleEvent(SmallVectorImpl<ResourceRef> &ResourcesFreed);
399 
400 #ifndef NDEBUG
dump()401   void dump() const {
402     for (const std::unique_ptr<ResourceState> &Resource : Resources)
403       Resource->dump();
404   }
405 #endif
406 };
407 } // namespace mca
408 } // namespace llvm
409 
410 #endif // LLVM_MCA_RESOURCE_MANAGER_H
411