1 /*===- InstrProfilingValue.c - Support library for PGO instrumentation ----===*\
3 |* The LLVM Compiler Infrastructure
5 |* This file is distributed under the University of Illinois Open Source
6 |* License. See LICENSE.TXT for details.
8 \*===----------------------------------------------------------------------===*/
15 #include "InstrProfiling.h"
16 #include "InstrProfilingInternal.h"
17 #include "InstrProfilingUtil.h"
19 #define INSTR_PROF_VALUE_PROF_DATA
20 #define INSTR_PROF_COMMON_API_IMPL
21 #include "InstrProfData.inc"
23 static int hasStaticCounters = 1;
24 static int OutOfNodesWarnings = 0;
25 static int hasNonDefaultValsPerSite = 0;
26 #define INSTR_PROF_MAX_VP_WARNS 10
27 #define INSTR_PROF_DEFAULT_NUM_VAL_PER_SITE 16
28 #define INSTR_PROF_VNODE_POOL_SIZE 1024
31 /* A shared static pool in addition to the vnodes statically
32 * allocated by the compiler. */
33 COMPILER_RT_VISIBILITY ValueProfNode
34 lprofValueProfNodes[INSTR_PROF_VNODE_POOL_SIZE] COMPILER_RT_SECTION(
35 COMPILER_RT_SEG INSTR_PROF_VNODES_SECT_NAME_STR);
38 COMPILER_RT_VISIBILITY uint32_t VPMaxNumValsPerSite =
39 INSTR_PROF_DEFAULT_NUM_VAL_PER_SITE;
41 COMPILER_RT_VISIBILITY void lprofSetupValueProfiler() {
43 Str = getenv("LLVM_VP_MAX_NUM_VALS_PER_SITE");
45 VPMaxNumValsPerSite = atoi(Str);
46 hasNonDefaultValsPerSite = 1;
48 if (VPMaxNumValsPerSite > INSTR_PROF_MAX_NUM_VAL_PER_SITE)
49 VPMaxNumValsPerSite = INSTR_PROF_MAX_NUM_VAL_PER_SITE;
52 COMPILER_RT_VISIBILITY void lprofSetMaxValsPerSite(uint32_t MaxVals) {
53 VPMaxNumValsPerSite = MaxVals;
54 hasNonDefaultValsPerSite = 1;
57 /* This method is only used in value profiler mock testing. */
58 COMPILER_RT_VISIBILITY void
59 __llvm_profile_set_num_value_sites(__llvm_profile_data *Data,
60 uint32_t ValueKind, uint16_t NumValueSites) {
61 *((uint16_t *)&Data->NumValueSites[ValueKind]) = NumValueSites;
64 /* This method is only used in value profiler mock testing. */
65 COMPILER_RT_VISIBILITY const __llvm_profile_data *
66 __llvm_profile_iterate_data(const __llvm_profile_data *Data) {
70 /* This method is only used in value profiler mock testing. */
71 COMPILER_RT_VISIBILITY void *
72 __llvm_get_function_addr(const __llvm_profile_data *Data) {
73 return Data->FunctionPointer;
76 /* Allocate an array that holds the pointers to the linked lists of
77 * value profile counter nodes. The number of element of the array
78 * is the total number of value profile sites instrumented. Returns
79 * 0 if allocation fails.
82 static int allocateValueProfileCounters(__llvm_profile_data *Data) {
83 uint64_t NumVSites = 0;
86 /* This function will never be called when value site array is allocated
87 statically at compile time. */
88 hasStaticCounters = 0;
89 /* When dynamic allocation is enabled, allow tracking the max number of
91 if (!hasNonDefaultValsPerSite)
92 VPMaxNumValsPerSite = INSTR_PROF_MAX_NUM_VAL_PER_SITE;
94 for (VKI = IPVK_First; VKI <= IPVK_Last; ++VKI)
95 NumVSites += Data->NumValueSites[VKI];
98 (ValueProfNode **)calloc(NumVSites, sizeof(ValueProfNode *));
101 if (!COMPILER_RT_BOOL_CMPXCHG(&Data->Values, 0, Mem)) {
108 static ValueProfNode *allocateOneNode(__llvm_profile_data *Data, uint32_t Index,
112 if (!hasStaticCounters)
113 return (ValueProfNode *)calloc(1, sizeof(ValueProfNode));
115 /* Early check to avoid value wrapping around. */
116 if (CurrentVNode + 1 > EndVNode) {
117 if (OutOfNodesWarnings++ < INSTR_PROF_MAX_VP_WARNS) {
118 PROF_WARN("Unable to track new values: %s. "
119 " Consider using option -mllvm -vp-counters-per-site=<n> to "
121 " value profile counters at compile time. \n",
122 "Running out of static counters");
126 Node = COMPILER_RT_PTR_FETCH_ADD(ValueProfNode, CurrentVNode, 1);
127 /* Due to section padding, EndVNode point to a byte which is one pass
128 * an incomplete VNode, so we need to skip the last incomplete node. */
129 if (Node + 1 > EndVNode)
135 COMPILER_RT_VISIBILITY void
136 __llvm_profile_instrument_target(uint64_t TargetValue, void *Data,
137 uint32_t CounterIndex) {
138 __llvm_profile_data *PData = (__llvm_profile_data *)Data;
142 if (!PData->Values) {
143 if (!allocateValueProfileCounters(PData))
147 ValueProfNode **ValueCounters = (ValueProfNode **)PData->Values;
148 ValueProfNode *PrevVNode = NULL;
149 ValueProfNode *MinCountVNode = NULL;
150 ValueProfNode *CurVNode = ValueCounters[CounterIndex];
151 uint64_t MinCount = UINT64_MAX;
153 uint8_t VDataCount = 0;
155 if (TargetValue == CurVNode->Value) {
159 if (CurVNode->Count < MinCount) {
160 MinCount = CurVNode->Count;
161 MinCountVNode = CurVNode;
163 PrevVNode = CurVNode;
164 CurVNode = CurVNode->Next;
168 if (VDataCount >= VPMaxNumValsPerSite) {
169 /* Bump down the min count node's count. If it reaches 0,
170 * evict it. This eviction/replacement policy makes hot
171 * targets more sticky while cold targets less so. In other
172 * words, it makes it less likely for the hot targets to be
173 * prematurally evicted during warmup/establishment period,
174 * when their counts are still low. In a special case when
175 * the number of values tracked is reduced to only one, this
176 * policy will guarantee that the dominating target with >50%
177 * total count will survive in the end. Note that this scheme
178 * allows the runtime to track the min count node in an adaptive
179 * manner. It can correct previous mistakes and eventually
180 * lock on a cold target that is alread in stable state.
182 * In very rare cases, this replacement scheme may still lead
183 * to target loss. For instance, out of \c N value slots, \c N-1
184 * slots are occupied by luke warm targets during the warmup
185 * period and the remaining one slot is competed by two or more
186 * very hot targets. If those hot targets occur in an interleaved
187 * way, none of them will survive (gain enough weight to throw out
188 * other established entries) due to the ping-pong effect.
189 * To handle this situation, user can choose to increase the max
190 * number of tracked values per value site. Alternatively, a more
191 * expensive eviction mechanism can be implemented. It requires
192 * the runtime to track the total number of evictions per-site.
193 * When the total number of evictions reaches certain threshold,
194 * the runtime can wipe out more than one lowest count entries
195 * to give space for hot targets.
197 if (!MinCountVNode->Count || !(--MinCountVNode->Count)) {
198 CurVNode = MinCountVNode;
199 CurVNode->Value = TargetValue;
205 CurVNode = allocateOneNode(PData, CounterIndex, TargetValue);
208 CurVNode->Value = TargetValue;
211 uint32_t Success = 0;
212 if (!ValueCounters[CounterIndex])
214 COMPILER_RT_BOOL_CMPXCHG(&ValueCounters[CounterIndex], 0, CurVNode);
215 else if (PrevVNode && !PrevVNode->Next)
216 Success = COMPILER_RT_BOOL_CMPXCHG(&(PrevVNode->Next), 0, CurVNode);
218 if (!Success && !hasStaticCounters) {
225 * The target values are partitioned into multiple regions/ranges. There is one
226 * contiguous region which is precise -- every value in the range is tracked
227 * individually. A value outside the precise region will be collapsed into one
228 * value depending on the region it falls in.
230 * There are three regions:
231 * 1. (-inf, PreciseRangeStart) and (PreciseRangeLast, LargeRangeValue) belong
232 * to one region -- all values here should be mapped to one value of
233 * "PreciseRangeLast + 1".
234 * 2. [PreciseRangeStart, PreciseRangeLast]
235 * 3. Large values: [LargeValue, +inf) maps to one value of LargeValue.
237 * The range for large values is optional. The default value of INT64_MIN
238 * indicates it is not specified.
240 COMPILER_RT_VISIBILITY void __llvm_profile_instrument_range(
241 uint64_t TargetValue, void *Data, uint32_t CounterIndex,
242 int64_t PreciseRangeStart, int64_t PreciseRangeLast, int64_t LargeValue) {
244 if (LargeValue != INT64_MIN && (int64_t)TargetValue >= LargeValue)
245 TargetValue = LargeValue;
246 else if ((int64_t)TargetValue < PreciseRangeStart ||
247 (int64_t)TargetValue > PreciseRangeLast)
248 TargetValue = PreciseRangeLast + 1;
250 __llvm_profile_instrument_target(TargetValue, Data, CounterIndex);
254 * A wrapper struct that represents value profile runtime data.
255 * Like InstrProfRecord class which is used by profiling host tools,
256 * ValueProfRuntimeRecord also implements the abstract intefaces defined in
257 * ValueProfRecordClosure so that the runtime data can be serialized using
258 * shared C implementation.
260 typedef struct ValueProfRuntimeRecord {
261 const __llvm_profile_data *Data;
262 ValueProfNode **NodesKind[IPVK_Last + 1];
263 uint8_t **SiteCountArray;
264 } ValueProfRuntimeRecord;
266 /* ValueProfRecordClosure Interface implementation. */
268 static uint32_t getNumValueSitesRT(const void *R, uint32_t VK) {
269 return ((const ValueProfRuntimeRecord *)R)->Data->NumValueSites[VK];
272 static uint32_t getNumValueDataRT(const void *R, uint32_t VK) {
274 const ValueProfRuntimeRecord *Record = (const ValueProfRuntimeRecord *)R;
275 if (Record->SiteCountArray[VK] == INSTR_PROF_NULLPTR)
277 for (I = 0; I < Record->Data->NumValueSites[VK]; I++)
278 S += Record->SiteCountArray[VK][I];
282 static uint32_t getNumValueDataForSiteRT(const void *R, uint32_t VK,
284 const ValueProfRuntimeRecord *Record = (const ValueProfRuntimeRecord *)R;
285 return Record->SiteCountArray[VK][S];
288 static ValueProfRuntimeRecord RTRecord;
289 static ValueProfRecordClosure RTRecordClosure = {
290 &RTRecord, INSTR_PROF_NULLPTR, /* GetNumValueKinds */
291 getNumValueSitesRT, getNumValueDataRT, getNumValueDataForSiteRT,
292 INSTR_PROF_NULLPTR, /* RemapValueData */
293 INSTR_PROF_NULLPTR, /* GetValueForSite, */
294 INSTR_PROF_NULLPTR /* AllocValueProfData */
298 initializeValueProfRuntimeRecord(const __llvm_profile_data *Data,
299 uint8_t *SiteCountArray[]) {
300 unsigned I, J, S = 0, NumValueKinds = 0;
301 ValueProfNode **Nodes = (ValueProfNode **)Data->Values;
302 RTRecord.Data = Data;
303 RTRecord.SiteCountArray = SiteCountArray;
304 for (I = 0; I <= IPVK_Last; I++) {
305 uint16_t N = Data->NumValueSites[I];
311 RTRecord.NodesKind[I] = Nodes ? &Nodes[S] : INSTR_PROF_NULLPTR;
312 for (J = 0; J < N; J++) {
313 /* Compute value count for each site. */
315 ValueProfNode *Site =
316 Nodes ? RTRecord.NodesKind[I][J] : INSTR_PROF_NULLPTR;
323 RTRecord.SiteCountArray[I][J] = C;
327 return NumValueKinds;
330 static ValueProfNode *getNextNValueData(uint32_t VK, uint32_t Site,
331 InstrProfValueData *Dst,
332 ValueProfNode *StartNode, uint32_t N) {
334 ValueProfNode *VNode = StartNode ? StartNode : RTRecord.NodesKind[VK][Site];
335 for (I = 0; I < N; I++) {
336 Dst[I].Value = VNode->Value;
337 Dst[I].Count = VNode->Count;
343 static uint32_t getValueProfDataSizeWrapper(void) {
344 return getValueProfDataSize(&RTRecordClosure);
347 static uint32_t getNumValueDataForSiteWrapper(uint32_t VK, uint32_t S) {
348 return getNumValueDataForSiteRT(&RTRecord, VK, S);
351 static VPDataReaderType TheVPDataReader = {
352 initializeValueProfRuntimeRecord, getValueProfRecordHeaderSize,
353 getFirstValueProfRecord, getNumValueDataForSiteWrapper,
354 getValueProfDataSizeWrapper, getNextNValueData};
356 COMPILER_RT_VISIBILITY VPDataReaderType *lprofGetVPDataReader() {
357 return &TheVPDataReader;