-
Notifications
You must be signed in to change notification settings - Fork 4
/
FProfile.jl
527 lines (466 loc) · 19.8 KB
/
FProfile.jl
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
# Copyright notice: this file contains several functions derived from Julia's profile.jl
# ip = instruction pointer
# li = line-info
__precompile__()
module FProfile
export @fprofile, backtraces, tree, flat
export get_stackframe, get_method, get_specialization, get_file, get_function, get_module,
is_C_call, is_inlined, filter_bloodline, prune, details
using Base: Profile
using Base.Core: MethodInstance
using Base.Profile: ProfileFormat, LineInfoFlatDict, LineInfoDict, StackFrame,
tree_aggregate, flatten, purgeC, tree_format, UNKNOWN, show_spec_linfo,
rtruncto, ltruncto, tree_format_linewidth, count_flat, parse_flat, flatten
using DataFrames
using DataStructures: OrderedDict, Accumulator, counter
using Requires
const BackTraces = Vector{Tuple{Int64,Vector{StackFrame}}}
module MissingInfo # placeholders
missing_info() = nothing
missing_info() # call it to generate a specialization
const missing_info_method_instance = let res=nothing
Base.visit(spec->(res=spec;), methods(missing_info).ms[1].specializations)
res
end
end
using .MissingInfo: missing_info, missing_info_method_instance
struct ProfileData # a mere container for Base.Profile data
data::Vector
lidict::Profile.LineInfoDict
end
ProfileData() = ProfileData(Profile.retrieve()...)
Base.length(pd::ProfileData) = mapreduce(first, +, 0, backtraces(pd))
Base.show(io::IO, pd::ProfileData) =
write(io, "ProfileData($(length(pd)) backtraces)")
Profile.print(pd::ProfileData; kwargs...) = Profile.print(pd.data, pd.lidict; kwargs...)
""" `@fprofile(expr, delay=0.001, n=1000000)` profiles the execution of `expr`, taking a
snapshot (backtrace) every `delay` seconds (up to `n` backtraces). It returns the
profiling results as a `ProfileData` object. """
macro fprofile(expr, delay=0.001, n=1000000)
esc(quote
Profile.clear()
Profile.init(; n=$n, delay=$delay)
@profile $expr
res = $FProfile.ProfileData()
end)
end
""" `@fprofile(niter::Int, expr)` is shorthand for
```julia
@fprofile for _ in 1:niter
expr
end
```
"""
macro fprofile(niter::Int, expr, args...)
esc(quote
$FProfile.@fprofile(for _ in 1:$niter; $expr end, $(args...))
end)
end
################################################################################
# Accessor functions
get_line(obj)::Int = get_stackframe(obj).line
get_file(obj)::String = get_method(obj).file
get_specialization(obj)::MethodInstance =
get(get_stackframe(obj).linfo, missing_info_method_instance)
get_method(obj)::Method = get_specialization(obj).def
function get_function(sf)::Function
met = get_method(sf)
ftype = fieldtype(met.sig, 1)
return isdefined(ftype, :instance) ? ftype.instance : missing_info
end
get_module(obj)::Module = get_method(obj).module
get_stackframe(sf::StackFrame) = sf
get_specialization(mi::MethodInstance) = mi
get_method(met::Method) = met
get_function(fun::Function) = fun
get_line(line::Int) = line
get_file(file::Symbol) = file
get_module(m::Module) = m
is_C_call(sf::StackFrame) = sf.from_c
is_inlined(sf::StackFrame) = sf.inlined
const symbol2accessor_dict = OrderedDict(:stackframe=>get_stackframe,
# The line is already part of :stackframe, and
# grouping on :stackframe makes more sense
# anyway. :line=>get_line,
:specialization=>get_specialization,
:method=>get_method,
:file=>get_file,
:function=>get_function,
:module=>get_module)
const type2symbol_dict = Dict(StackFrame=>:stackframe,
MethodInstance=>:specialization,
Method=>:method,
String=>:file,
Function=>:function,
Module=>:module)::Any
function symbol2accessor(sym::Symbol)
@assert(haskey(symbol2accessor_dict, sym),
"Invalid accessor/combiner: $sym. Must be one of $(collect(Base.keys(symbol2accessor_dict)))")
symbol2accessor_dict[sym]
end
function type2symbol(T)
for (typ, sym) in type2symbol_dict
if T <: typ; return sym; end
end
error("Can only handle objects of types $(collect(keys(type2symbol)))")
end
type2accessor(x::Type) = symbol2accessor(type2symbol(x))
################################################################################
# backtraces
""" `backtraces(pd::ProfileData; flatten=true, C=false)` returns a vector of `(count,
backtrace)`, where `backtrace` is a `Vector{StackFrame}` which occurred `count`
times during the profiler run. If `C` is `false`, C function calls are excluded. """
function backtraces(pd::ProfileData; flatten=true, C=false)
data, lidict = pd.data, pd.lidict
if flatten
data, lidict = Profile.flatten(data, lidict)
end
data, counts = Profile.tree_aggregate(data)
out = BackTraces(0)
for (count, backtrace) in zip(counts, data)
new_trace = StackFrame[lidict[d] for d in backtrace if C || !is_C_call(lidict[d])]
if !isempty(new_trace)
push!(out, (count, new_trace))
end
end
return out
end
function select_backtrace_neighborhoods(btraces::BackTraces, pred::Function,
neighborhood::UnitRange)
# Returns a vector of traces, possibly longer than the input, that contains every
# neighborhood centered around where `pred(::StackFrame)` is true. Neighborhood
# that touch/overlap are merged.
out = BackTraces(0)
for (count, trace) in btraces
hits = find(pred, trace)
in_hood(i) = any(i in neighborhood+h for h in hits)
positions = 1:length(trace)
i = 1
while true
start = findnext(in_hood, positions, i)
if start != 0
stop = findnext(!in_hood, positions, start)
if stop == 0
push!(out, (count, trace[start:end]))
break
else
push!(out, (count, trace[start:stop-1]))
i = stop
end
else
break
end
end
end
return out
end
################################################################################
# tree view
""" `Node(object, count::Int, children::Vector{Node})` represents the `tree`
view of the profiling data. `object` is a `StackFrame`, unless combineby is specified. """
mutable struct Node
obj
count::Int
children::Vector{Node}
end
Node(node::Node, children::Vector{Node}) = Node(node.obj, node.count, children)
Base.getindex(node::Node, i::Int) = node.children[i]
Base.getindex(node::Node, i::Int, args...) = node[i][args...]
Base.length(node::Node) = length(node.children)
for acc in (:get_stackframe, :get_specialization, :get_method, :get_file, :get_function,
:get_module)
@eval $acc(node::Node) = $acc(node.obj)
end
# This filter code was taken from TraceCalls.jl
filter_descendents(f, node) = # helper
# Special casing because of #18852
isempty(node.children) ? Node[] : Node[n for child in node.children
for n in filter_(f, child)]
filter_(f, node) =
(f(node) ? [Node(node, filter_descendents(f, node))] :
filter_descendents(f, node))
Base.filter(f::Function, node::Node) = Node(node, filter_descendents(f, node))
Base.map(f::Function, node::Node) =
# Apply f(::Node), leaves first
f(Node(node, Node[map(f, child) for child in node.children]))
prune(node::Node, i=0) =
i<=0 ? Node(node, Node[]) : Node(node, Node[prune(n, i-1) for n in node.children])
const empty_node_dummy = Node(UNKNOWN, -1, [])
""" filter_bloodline(f::Function, node::Node; keep_descendents=true, keep_ancestors=true)
keeps all nodes in the tree for which `f(::Trace)` is true of some of its descendents OR
ancestors. """
function filter_bloodline(f::Function, node::Node; keep_descendents=true,
keep_ancestors=true)
if !keep_ancestors
return filter_bloodline(f, Node(UNKNOWN, -1, find_nodes(f, node));
keep_descendents=keep_descendents)
elseif f(get_stackframe(node))
return keep_descendents ? node : prune(node)
else
children0 = Node[filter_bloodline(f, sub_node; keep_descendents=keep_descendents)
for sub_node in node.children]
children = filter(c->c!==empty_node_dummy, children0)
return isempty(children) ? empty_node_dummy : Node(node, children)
end
end
""" `find_nodes(f, node)` returns a vector of all nodes satisfying `f` """
function find_nodes(f::Function, node::Node)
out = Node[]
function trav(n)
if f(n) push!(out, n) end
foreach(trav, n.children)
end
trav(node)
out
end
tree_line_string(obj, ntext::Int) = rtruncto(string(obj), ntext)
function tree_line_string(li::Union{StackFrame, Method}, ntext::Int)
widthfile = floor(Integer, 0.4ntext)
widthfunc = floor(Integer, 0.6ntext)
if li != UNKNOWN
if li isa StackFrame && li.line == li.pointer
return string("unknown function (pointer: 0x",
hex(li.pointer,2*sizeof(Ptr{Void})),
")")
else
fname = string(get_function(li))
if li isa Method
fname = sprint(Base.show_tuple_as_call, li.name, li.sig)
elseif !li.from_c && !isnull(li.linfo)
fname = sprint(show_spec_linfo, li)
end
return string(rtruncto(string(li.file), widthfile),
":",
li.line == -1 ? "?" : string(li.line),
"; ",
ltruncto(fname, widthfunc))
end
else
return nothing
end
end
function Profile.tree_format(obj, count::Int, level::Int, cols::Int,
ndigcounts::Int, ndigline::Int)
nindent = min(cols>>1, level)
ntext = cols - nindent - ndigcounts - ndigline - 5
base = " "^nindent
if level > nindent
nextra = level - nindent
nindent -= ndigits(nextra) + 2
base = string(base, "+", nextra, " ")
end
base = string(base, rpad(string(count), ndigcounts, " "), " ")
line = tree_line_string(obj, ntext)
return line === nothing ? nothing : base*line
end
# tree_format_linewidth is just the width of the line number, which is used to center
# the printing and provide the most useful information. 0 is a cop-out; FIXME
Base.Profile.tree_format_linewidth(x) = 0
function Base.show(io::IO, node::Node)
cols::Int = Base.displaysize(io)[2]
level = get(io, :profile_tree_level, 0)
str = tree_format(node.obj, node.count, level, cols,
get(io, :profile_ndigcounts, ndigits(node.count)),
get(io, :profile_ndigline, tree_format_linewidth(node.obj)))
if str !== nothing println(io, str) end
if !isempty(node.children)
io2 = IOContext(io,
profile_tree_level=level+1,
profile_ndigcounts=maximum(ndigits(child.count)
for child in node.children),
profile_ndigline=maximum(tree_format_linewidth(child.obj)
for child in node.children))
for c in node.children
show(io2, c)
end
end
end
""" `tree(pd::ProfileData; C = false, mincount::Int = 0, maxdepth=-1)` displays a tree
of function calls, along with the number of backtraces going through each call.
- `C` -- If `true`, backtraces from C and Fortran code are shown (normally they are excluded).
- `maxdepth` -- Limits the depth higher than `maxdepth` in the `:tree` format.
- `mincount` -- Limits the printout to only those lines with at least `mincount` occurrences.
- `combineby` -- Aggregate by `:specialization, :method, :file, :function`, or `:module`
"""
tree(pd::ProfileData; C = false, mincount::Int = 0, maxdepth=-1, combineby=:stackframe) =
tree(backtraces(pd; C=C); mincount=mincount, maxdepth=maxdepth, combineby=combineby)
function tree(bt::BackTraces; mincount::Int = 0, maxdepth=-1, combineby=:stackframe)
# We start with an empty Node tree, then iterate over every trace, adding counts and
# new branches.
combiner = symbol2accessor(combineby)
root = Node(UNKNOWN, -1, [])
for (count, trace) in bt
node = root
prev_obj = nothing
for sf in trace
let obj=combiner(sf) # for speed - see #15276
if !(obj isa StackFrame) && prev_obj == obj
continue
end
prev_obj = obj
# Speed note: now that Node.obj is untyped, this line might be a
# bottleneck.
i = findfirst(n->n.obj==obj, node.children)
if i == 0
# Make a new branch
next_node = Node(obj, 0, Node[])
push!(node.children, next_node)
else
next_node = node.children[i]
end
next_node.count += count
node = next_node
end
end
end
# Sort the children in each node alphabetically. See Profile.liperm.
# root = map(n->Node(n, n.children[Profile.liperm(map(n->n.obj, n.children))]),
# root)
if maxdepth != -1
root = prune(root, maxdepth)
end
return filter(node->node.count >= mincount, root)
end
function tree(pd::ProfileData, object::T, neighborhood::UnitRange=-1:1; kwargs...) where T
acc = type2accessor(T)
tree(select_backtrace_neighborhoods(backtraces(pd),
sf->acc(sf)==object, neighborhood);
kwargs...)
end
include("tree_base.jl")
################################################################################
# flat view
function counts_from_traces(backtraces::Vector, key::Function,
encountered_key::Function=key)
counts = Dict()
encountered = Set()
for (trace_count, trace) in backtraces
empty!(encountered)
for sf in trace
ek = encountered_key(sf)
if !(ek in encountered)
push!(encountered, ek)
k = key(sf)
counts[k] = get(counts, k, 0) + trace_count
end
end
end
return counts
end
function self_counts_from_traces(backtraces::Vector, key::Function, applicable::Function)
counts = Dict()
for (trace_count, trace) in backtraces
for sf in @view trace[end:-1:1]
if applicable(sf)
k = key(sf)
counts[k] = get(counts, k, 0) + trace_count
break
end
end
end
return counts
end
# -----------------------------------------------------------------------------
function is_applicable(f::Function, object)
try
f(object)
catch e
if e isa MethodError; return false else rethrow() end
end
true
end
""" flat(pd::ProfileData; C=false, combineby=:stackframe, percent=true, inlined=true)
Returns aggregated profiling results as a `DataFrame`. Arguments:
- `C=false`: whether to exclude C calls
- `combineby`: one of `[:stackframe, :specialization, :method, :file, :function, :module]`.
`combineby=stackframe` provides the most detailed report, with method and line number.
The other options combine rows that have the same `combineby`.
- `percent=true`: show percentages
- `inlined=true`: whether to show inlined function calls
"""
flat(pd::ProfileData; C=false, combineby=:stackframe, percent=true, inlined=true,
# internal parameter
_module=nothing) =
flat(backtraces(pd; flatten=true, C=C); combineby=combineby,
percent=percent, inlined=inlined, _module=_module)
function flat(btraces::BackTraces;
C=false,
combineby=:stackframe,
percent=true,
inlined=true,
# internal parameter
_module=nothing)
if _module!==nothing && combineby in (:function, :file)
# Because a function/file isn't uniquely associated to a module
error("Cannot combineby $combineby if a module is provided; try `combineby=:method`")
end
count_dict = counts_from_traces(btraces, symbol2accessor(combineby))
keys = collect(Base.keys(count_dict))
@assert !isempty(keys) "ProfileData contains no applicable traces"
ntrace = sum(first, btraces)
perc(var::Symbol, counts) =
(percent ? Symbol(var, "_pct") => round.(counts ./ ntrace * 100, 2) :
var => counts)
count_cols = [perc(:count, [count_dict[sf] for sf in keys])]
if _module !== nothing
self_count_dict = self_counts_from_traces(btraces, symbol2accessor(combineby),
sf->get_module(sf) in _module)
push!(count_cols, perc(:self, [get(self_count_dict, sf, 0) for sf in keys]))
end
df = DataFrame(OrderedDict(count_cols..., combineby=>keys))
# Use this code to add all the remaining columns
# [col=>map(f, keys) for (col, f) in symbol2accessor_dict
# if is_applicable(f, first(keys))]...))
if _module !== nothing; df = df[[get_module(obj) in _module for obj in df[combineby]], :] end
if !inlined; df = df[!is_inlined.(df[:stackframe]), :] end
return sort(df, cols=percent ? :count_pct : :count, rev=true)
end
flat(pd::ProfileData, _module::Tuple; kwargs...) =
flat(pd; kwargs..., _module=_module)
flat(pd::ProfileData, _module::Module; kwargs...) =
flat(pd; kwargs..., _module=(_module,))
""" `df_combineby(df::DataFrame)` returns by what this `df` was combined. """
df_combineby(df::DataFrame) =
names(df)[findfirst(n->haskey(symbol2accessor_dict, n), names(df))]
tree(pd::ProfileData, df::DataFrame, nrow::Int, neighborhood::UnitRange=-1:1) =
tree(pd, df[nrow, df_combineby(df)], neighborhood)
################################################################################
# Comparisons
function my_outer_join(df1, df2, on::Vector)
# See DataFrames.jl#1270 for why this is necessary
kept = setdiff(names(df1), on)
row_vals(df, i) = map(last, DataFrameRow(df, i))
mk_dict(df) = Dict(row_vals(df, i)=>i for i in 1:size(df, 1))
dict1 = mk_dict(df1[:, on])
dict2 = mk_dict(df2[:, on])
df_keys = unique(vcat(df1, df2)[on])
build_col(df, dict, col) = [haskey(dict, row_vals(df_keys, i)) ?
df[dict[row_vals(df_keys, i)], col] : 0
for i in 1:size(df_keys, 1)]
df_kept = DataFrame()
for col in kept
col1, col2 = Symbol(col, :_1), Symbol(col, :_2)
df_kept[col1] = build_col(df1, dict1, col)
df_kept[col2] = build_col(df2, dict2, col)
df_kept[Symbol(col, :_diff)] = round.(df_kept[col2] .- df_kept[col1], 2)
end
return hcat(df_kept, df_keys)
end
function flat(pd1::ProfileData, pd2::ProfileData; _module=nothing, kwargs...)
df1 = flat(pd1; _module=_module, kwargs...)
df2 = flat(pd2; _module=_module, kwargs...)
combineby_ind = _module === nothing ? 2 : 3
df = FProfile.my_outer_join(df1, df2, names(df1)[combineby_ind:end])
return sort(df, cols=names(df)[3], rev=true)
end
flat(pd1::ProfileData, pd2::ProfileData, _module::Tuple; kwargs...) =
flat(pd1, pd2; kwargs..., _module=_module)
flat(pd1::ProfileData, pd2::ProfileData, _module::Module; kwargs...) =
flat(pd1, pd2; kwargs..., _module=(_module,))
################################################################################
# ProfileView
@require ProfileView begin
ProfileView.view(pd::ProfileData; kwargs...) =
ProfileView.view(pd.data; lidict=pd.lidict, kwargs...)
end
end # module