| 1 | #!/usr/bin/env bash
|
| 2 | #
|
| 3 | # Quick test for a potential rewrite of mycpp.
|
| 4 | #
|
| 5 | # Usage:
|
| 6 | # pea/TEST.sh <function name>
|
| 7 |
|
| 8 | set -o nounset
|
| 9 | set -o pipefail
|
| 10 | set -o errexit
|
| 11 |
|
| 12 | source test/common.sh # run-test-funcs
|
| 13 | source devtools/common.sh
|
| 14 |
|
| 15 | source build/dev-shell.sh # find python3 in /wedge PATH component
|
| 16 |
|
| 17 | # This is just like the yapf problem in devtools/format.sh !
|
| 18 | # Pea needs a newer version of MyPy -- one that supports 'math'
|
| 19 | unset PYTHONPATH
|
| 20 | export PYTHONPATH=.
|
| 21 |
|
| 22 | readonly MYPY_VENV='_tmp/mypy-venv'
|
| 23 |
|
| 24 | install-mypy() {
|
| 25 | local venv=$MYPY_VENV
|
| 26 |
|
| 27 | rm -r -f -v $venv
|
| 28 |
|
| 29 | python3 -m venv $venv
|
| 30 |
|
| 31 | . $venv/bin/activate
|
| 32 |
|
| 33 | python3 -m pip install mypy
|
| 34 |
|
| 35 | # Says 1.5.1 (compiled: yes)
|
| 36 | mypy-version
|
| 37 | }
|
| 38 |
|
| 39 | mypy-version() {
|
| 40 | . $MYPY_VENV/bin/activate
|
| 41 | python3 -m mypy --version
|
| 42 | }
|
| 43 |
|
| 44 | #
|
| 45 | # Run Pea
|
| 46 | #
|
| 47 |
|
| 48 | pea-main() {
|
| 49 | pea/pea_main.py "$@"
|
| 50 | }
|
| 51 |
|
| 52 | parse-one() {
|
| 53 | pea-main parse "$@"
|
| 54 | }
|
| 55 |
|
| 56 | translate-cpp() {
|
| 57 | ### Used by mycpp/NINJA-steps.sh
|
| 58 |
|
| 59 | pea-main cpp "$@"
|
| 60 | }
|
| 61 |
|
| 62 | all-files() {
|
| 63 | # Can't run this on Soil because we only have build/py.sh py-source, not
|
| 64 | # 'minimal'
|
| 65 |
|
| 66 | # Update this file with build/dynamic-deps.sh pea-hack
|
| 67 |
|
| 68 | cat pea/oils-typecheck.txt
|
| 69 |
|
| 70 | for path in */*.pyi; do
|
| 71 | echo $path
|
| 72 | done
|
| 73 | }
|
| 74 |
|
| 75 | parse-all() {
|
| 76 | time all-files | xargs --verbose -- $0 pea-main parse
|
| 77 | }
|
| 78 |
|
| 79 | # Good illustration of "distributing your overhead"
|
| 80 | #
|
| 81 | # Total work goes up, while latency goes down. To a point. Then it goes back
|
| 82 | # up.
|
| 83 |
|
| 84 | # batch size 30
|
| 85 | #
|
| 86 | # real 0m0.342s
|
| 87 | # user 0m0.735s
|
| 88 | # sys 0m0.059s
|
| 89 | #
|
| 90 | # batch size 20
|
| 91 | #
|
| 92 | # real 0m0.305s
|
| 93 | # user 0m0.993s
|
| 94 | # sys 0m0.081s
|
| 95 | #
|
| 96 | # batch size 15
|
| 97 | #
|
| 98 | # real 0m0.299s
|
| 99 | # user 0m1.110s
|
| 100 | # sys 0m0.123s
|
| 101 | #
|
| 102 | # batch size 10
|
| 103 | #
|
| 104 | # real 0m0.272s
|
| 105 | # user 0m1.362s
|
| 106 | # sys 0m0.145s
|
| 107 |
|
| 108 | batch-size() {
|
| 109 | local num_files=$1
|
| 110 |
|
| 111 | local num_procs
|
| 112 | num_procs=$(nproc)
|
| 113 |
|
| 114 | # Use (p-1) as a fudge so we don't end up more batches than processors
|
| 115 | local files_per_process=$(( num_files / (num_procs - 1) ))
|
| 116 |
|
| 117 | echo "$num_procs $files_per_process"
|
| 118 | }
|
| 119 |
|
| 120 | demo-par() {
|
| 121 | ### Demo parallelism of Python processes
|
| 122 |
|
| 123 | local files
|
| 124 | num_files=$(all-files | wc -l)
|
| 125 |
|
| 126 | # 103 files
|
| 127 |
|
| 128 | shopt -s lastpipe
|
| 129 | batch-size $num_files | read num_procs optimal
|
| 130 |
|
| 131 | echo "Parsing $num_files files with $num_procs parallel processes"
|
| 132 | echo "Optimal batch size is $optimal"
|
| 133 |
|
| 134 | echo
|
| 135 |
|
| 136 | echo 'All at once:'
|
| 137 | time parse-all > /dev/null 2>&1
|
| 138 | echo
|
| 139 |
|
| 140 | # 5 is meant to be suboptimal
|
| 141 | for n in 50 30 20 10 5 $optimal; do
|
| 142 | echo "batch size $n"
|
| 143 | time all-files | xargs --verbose -P $num_procs -n $n -- \
|
| 144 | $0 parse-one > /dev/null 2>&1
|
| 145 | echo
|
| 146 | done
|
| 147 | }
|
| 148 |
|
| 149 | # - 0.40 secs to parse
|
| 150 | # - 0.56 secs pickle, so that's 160 ms
|
| 151 | # Then
|
| 152 | #
|
| 153 | # - 0.39 secs load pickle
|
| 154 | #
|
| 155 | # That's definitely slower than I want. It's 6.6 MB of data.
|
| 156 | #
|
| 157 | # So
|
| 158 | # - parallel parsing can be done in <300 ms
|
| 159 | # - parallel pickling
|
| 160 | # - serial unpickling (reduce) in 390 ms
|
| 161 | #
|
| 162 | # So now we're at ~700 ms or so. Can we type check in 300 ms in pure Python?
|
| 163 | #
|
| 164 | # What if we compress the generated ASDL? Those are very repetitive.
|
| 165 |
|
| 166 | # Problem statement:
|
| 167 |
|
| 168 | _serial-pickle() {
|
| 169 | mkdir -p _tmp
|
| 170 | local tmp=_tmp/serial
|
| 171 |
|
| 172 | time all-files | xargs --verbose -- $0 pea-main dump-pickles > $tmp
|
| 173 |
|
| 174 | ls -l -h $tmp
|
| 175 |
|
| 176 | echo 'loading'
|
| 177 | time pea-main load-pickles < $tmp
|
| 178 | }
|
| 179 |
|
| 180 | # 1.07 seconds
|
| 181 | serial-pickle() { time $0 _serial-pickle; }
|
| 182 |
|
| 183 | pickle-one() {
|
| 184 | pea-main dump-pickles "$@" > _tmp/p/$$
|
| 185 | }
|
| 186 |
|
| 187 | _par-pickle() {
|
| 188 | local files
|
| 189 | num_files=$(all-files | wc -l)
|
| 190 |
|
| 191 | shopt -s lastpipe
|
| 192 | batch-size $num_files | read num_procs optimal
|
| 193 |
|
| 194 | local dir=_tmp/p
|
| 195 | rm -r -f -v $dir
|
| 196 | mkdir -p $dir
|
| 197 |
|
| 198 | time all-files | xargs --verbose -P $num_procs -n $optimal -- $0 pickle-one
|
| 199 |
|
| 200 | ls -l -h $dir
|
| 201 |
|
| 202 | # This takes 410-430 ms? Wow that's slow.
|
| 203 | time cat $dir/* | pea-main load-pickles
|
| 204 | }
|
| 205 |
|
| 206 | # Can get this down to ~700 ms
|
| 207 | #
|
| 208 | # Note parsing serially in a single process is 410 ms !!! So this is NOT a win
|
| 209 | # unless we have more work besides parsing to parallelize.
|
| 210 | #
|
| 211 | # We can extract constants and forward declarations in parallel I suppose.
|
| 212 | #
|
| 213 | # BUT immutable string constants have to be de-duplciated! Though I guess that
|
| 214 | # is a natural 'reduce' step.
|
| 215 | #
|
| 216 | # And we can even do implementation and prototypes in parallel too?
|
| 217 | #
|
| 218 | # I think the entire algorithm can be OPTIMISTIC without serialized type
|
| 219 | # checking?
|
| 220 | #
|
| 221 | # I think
|
| 222 | #
|
| 223 | # a = 5
|
| 224 | # b = a # do not know the type without a global algorithm
|
| 225 | #
|
| 226 | # Or I guess you can do type checking within a function. Functions require
|
| 227 | # signatures. So yes let's do that in parallel.
|
| 228 | #
|
| 229 | # --
|
| 230 | #
|
| 231 | # The ideal way to do this would be to split Oils up into MODULES, like
|
| 232 | #
|
| 233 | # _debuild/
|
| 234 | # builtin/
|
| 235 | # core/
|
| 236 | # data_lang/
|
| 237 | # frontend/
|
| 238 | # osh/
|
| 239 | # ysh/
|
| 240 | # Smaller: pgen2/ pylib/ tools/
|
| 241 | #
|
| 242 | # And modules are acyclic, and can compile on their own with dependencies. If
|
| 243 | # you pick random .py files and spit out header files, I think they won't compile.
|
| 244 | # The forward declarations and constants will work, but the prototype won't.
|
| 245 |
|
| 246 | par-pickle() { time $0 _par-pickle; }
|
| 247 |
|
| 248 | sum1() {
|
| 249 | awk '{ sum += $1 } END { print sum }'
|
| 250 | }
|
| 251 |
|
| 252 | sum-sizes() {
|
| 253 | xargs -I {} -- find {} -printf '%s %p\n' | sum1
|
| 254 | }
|
| 255 |
|
| 256 | size-ratio() {
|
| 257 | # all-files
|
| 258 | # echo _tmp/p/*
|
| 259 |
|
| 260 | # 1.96 MB of source code
|
| 261 | all-files | sum-sizes
|
| 262 |
|
| 263 | # 7.13 MB of pickle files
|
| 264 | # Weirdly echo _tmp/p/* doesn't work here
|
| 265 | for f in _tmp/p/*; do echo $f; done | sum-sizes
|
| 266 | }
|
| 267 |
|
| 268 | # Only 47 ms!
|
| 269 | # I want the overhead to be less than 1 second:
|
| 270 | # 1. parallel parsing + pickle
|
| 271 | # 2. serial unpickle + type check
|
| 272 | # 3. starting the process
|
| 273 | #
|
| 274 | # So unpickling is slow.
|
| 275 |
|
| 276 | osh-overhead() {
|
| 277 | time bin/osh -c 'echo hi'
|
| 278 | }
|
| 279 |
|
| 280 |
|
| 281 | # MyPy dev version takes 10.2 seconds the first time (without their mypyc
|
| 282 | # speedups)
|
| 283 | #
|
| 284 | # 0.150 seconds the second time, WITHOUT code changes
|
| 285 | # 0.136 seconds
|
| 286 |
|
| 287 | # 4.1 seconds: whitespace change
|
| 288 | # 3.9 seconds: again, and this is on my fast hoover machine
|
| 289 |
|
| 290 | # 5.0 seconds - Invalid type!
|
| 291 | # 4.9 seconds - again invalid
|
| 292 |
|
| 293 |
|
| 294 | mypy-compare() {
|
| 295 | devtools/types.sh check-oils
|
| 296 | }
|
| 297 |
|
| 298 | check-types() {
|
| 299 |
|
| 300 | # install-mypy creates this. May not be present in CI machine.
|
| 301 | local activate=$MYPY_VENV/bin/activate
|
| 302 | if test -f $activate; then
|
| 303 | . $activate
|
| 304 | fi
|
| 305 |
|
| 306 | time python3 -m mypy --strict pea/pea_main.py
|
| 307 | }
|
| 308 |
|
| 309 | test-translate() {
|
| 310 | translate-cpp bin/oils_for_unix.py
|
| 311 | }
|
| 312 |
|
| 313 | test-syntax-error() {
|
| 314 | set +o errexit
|
| 315 |
|
| 316 | # error in Python syntax
|
| 317 | parse-one pea/testdata/py_err.py
|
| 318 | assert $? -eq 1
|
| 319 |
|
| 320 | # error in signature
|
| 321 | parse-one pea/testdata/sig_err.py
|
| 322 | assert $? -eq 1
|
| 323 |
|
| 324 | # error in assignment
|
| 325 | parse-one pea/testdata/assign_err.py
|
| 326 | assert $? -eq 1
|
| 327 | }
|
| 328 |
|
| 329 | run-tests() {
|
| 330 | # Making this separate for soil/worker.sh
|
| 331 |
|
| 332 | echo 'Running test functions'
|
| 333 | run-test-funcs
|
| 334 | }
|
| 335 |
|
| 336 | "$@"
|