summaryrefslogtreecommitdiff
path: root/tcwg_bmk-build.sh
blob: 8f211f6f3f03e9cee9eb047b27fdd8361362c955 (plain)
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
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
#!/bin/bash

set -euf -o pipefail

scripts=$(dirname $0)
# shellcheck source=jenkins-helpers.sh
. $scripts/jenkins-helpers.sh
# shellcheck source=round-robin.sh
. $scripts/round-robin.sh

convert_args_to_variables "$@"

obligatory_variables rr[ci_project] rr[ci_config] ssh_host ssh_port

# Execution mode: baseline, bisect, jenkins-full
# shellcheck disable=SC2154
rr[mode]="${rr[mode]-baseline}"

# Set custom revision for one of the projects, and use baseline revisions
# for all other projects.
# shellcheck disable=SC2154
rr[baseline_branch]="${rr[baseline_branch]-linaro-local/ci/${rr[ci_project]}/${rr[ci_config]}}"
# shellcheck disable=SC2154
rr[update_baseline]="${rr[update_baseline]-update}"
# shellcheck disable=SC2154
rr[top_artifacts]="${rr[top_artifacts]-$(pwd)/artifacts}"

# Set metric to perf by default.
# shellcheck disable=SC2154
rr[metric]="${rr[metric]-perf}"

# {toolchain_name}-{toolchain_ver}-{target}-{bmk}-{cflags}
IFS=- read -a ci_config <<EOF
${rr[ci_config]}
EOF
# shellcheck disable=SC2154
rr[toolchain]=${rr[toolchain]-${ci_config[0]}}
# shellcheck disable=SC2154
rr[target]=${rr[target]-${ci_config[2]}}
benchmarks=("${benchmarks[@]-${ci_config[3]}}")
if [ x"${benchmarks[*]}" = x"default" ]; then
    benchmarks=("${ci_config[3]}")
fi
if ! test_array cflags; then
    ci_config=("${ci_config[@]:4}")
    # In ${ci_config[@]} we now have "-"-separated entries (due to IFS=- above).
    # We restore "-" in compiler flags when doing flags="$flags-$flag" below.
    # We use "_" to separate compiler options, and it is translated to " -"
    # in benchmark().
    cflags=()
    while [ ${#ci_config[@]} -ge 1 ]; do
	flags=""
	while [ ${#ci_config[@]} -ge 1 ]; do
	    flag="${ci_config[0]}"
	    ci_config=("${ci_config[@]:1}")
	    if [ x"$flag" = x"vs" ]; then
		break
	    fi
	    flags="$flags-$flag"
	done
	cflags+=("$flags")
    done
fi

gcc_mode=""
for i in $(seq 0 $(("${#cflags[@]}" - 1))); do
    cflags_mode=""
    if [[ x"${cflags[$i]}" == x*"VECT"* ]]; then
	rr[metric]="vect"
    fi

    case "${rr[target]}:${cflags[$i]}" in
	"arm:"*"mthumb"*) cflags_mode=thumb ;;
	"arm:"*"marm"*) cflags_mode=arm ;;
	"arm:-Os"*|"arm:-Oz"*)
	    cflags_mode=thumb
	    cflags[$i]="${cflags[$i]}_mthumb"
	    ;;
	"arm:"*)
	    cflags_mode=arm
	    cflags[$i]="${cflags[$i]}_marm"
	    ;;
	"arm_eabi:"*)
	    cflags[$i]="${cflags[$i]}_mthumb"
	    ;;
    esac
    if [ x"$gcc_mode" = x"" ]; then
	gcc_mode="$cflags_mode"
    elif [ x"$gcc_mode" != x"$cflags_mode" ]; then
	assert_with_msg "Unsupported arm/thumb configuration ${cflags[$(($i - 1))]} and ${cflags[$i]}" false
    fi
done

gcc_override_configure=()
# Set default ARM/Thumb mode for AArch32 compiler.  This ensures that libraries
# (Glibc, libgcc, libstdc++, libgfortran, etc.) are built in the "interesting"
# ISA.
case "$gcc_mode" in
    arm|thumb) gcc_override_configure+=("--set" "gcc_override_configure=--with-mode=$gcc_mode") ;;
esac
# Disable libsanitizer because it didn't build in April-June 2018.
# We need to test revisions during that time period for performance regressions.
gcc_override_configure+=("--set" "gcc_override_configure=--disable-libsanitizer")

# Build the right libs depending on the target. Hardcode the
# cpu/float-abit/mode to speed up toolchain builds: the alternative of
# building rmprofile multilibs takes a very long time. Setting the
# values here has the drawback that we have to build one toolchain per
# board type.
case "${rr[target]}" in
    "arm_eabi") gcc_override_configure+=("--set" "gcc_override_configure=--disable-multilib"
					 "--set" "gcc_override_configure=--with-cpu=cortex-m4"
					 "--set" "gcc_override_configure=--with-mode=thumb"
					 "--set" "gcc_override_configure=--with-float=hard"
					) ;;
esac

case "${rr[toolchain]}" in
    llvm)
      # shellcheck disable=SC2154
      rr[components]="binutils gcc glibc llvm" ;;
    gnu)
      rr[components]="binutils gcc glibc" ;;
    gnu_eabi)
      rr[components]="binutils gcc newlib" ;;
    *) assert_with_msg "Unknown toolchain \"${rr[toolchain]}\"" false ;;
esac

# Use baseline branches by default.
for c in ${rr[components]}; do
    rr[${c}_git]=${rr[${c}_git]-baseline}
done

start_at="${start_at-default}"
finish_at="${finish_at-default}"
verbose="${verbose-true}"
verbose2="${verbose2-false}"

if $verbose2; then set -x; fi

trap print_traceback EXIT

# Set start and finish steps for different modes.
default_start_at=""
default_finish_at=""
case "${rr[mode]}" in
    "baseline")
	default_finish_at="update_baseline"
	;;
    "bisect")
	single_updated_component="$(print_single_updated_component)"
	case $single_updated_component in
	    binutils) default_start_at="build_abe-binutils" ;;
	    gcc) default_start_at="build_abe-stage1" ;;
	    glibc) default_start_at="clean_sysroot" ;;
	    llvm) default_start_at="build_llvm-true" ;;
	    newlib) default_start_at="build_abe-newlib" ;;
	    *) assert_with_msg \
        "Invalid single updated component \"$single_updated_component\"" false
      ;;
	esac
	default_finish_at="check_regression"
	;;
    "jenkins-full") ;;
esac
if [ x"$start_at" = x"default" ]; then
    start_at="$default_start_at"
fi
if [ x"$finish_at" = x"default" ]; then
    finish_at="$default_finish_at"
fi

run_step_init "$start_at" "$finish_at" "${rr[top_artifacts]}" "$verbose"

# If we bisect a regression between different major versions of Glibc,
# then we might get a mixed sysroot with several versions of ld-M.N.so and
# other binaries installed side-by-side.  Such a sysroot will break
# benchmarking, which requires a single ld-*.so binary to be present.
# Forcefully delete sysroot before building C library.
clean_sysroot ()
{
    (
    set -euf -o pipefail

    local gnu_target sysroot
    gnu_target=$(print_gnu_target ${rr[target]})
    sysroot="$(pwd)/abe/builds/destdir/x86_64-pc-linux-gnu/$gnu_target/libc"

    rm -rf "$sysroot"
    )
}

benchmark ()
{
    (
    set -euf -o pipefail

    local bmk_cflags="$1"
    local results_id_file="$3"

    sanity_check_pwd

    # shellcheck disable=SC2115
    rm -rf "$(pwd)"/bin
    mkdir "$(pwd)"/bin

    local bmk_flags bmk_ldflags reboot run_profile
    bmk_flags="$(echo $bmk_cflags | sed -e "s/_/ -/g" -e "s/LTO/flto/g")"
    bmk_flags="$(echo $bmk_cflags | sed -e "s/_/ -/g" -e "s/VECT/fmetric-vect/g")"
    case "$bmk_cflags" in
	"-Os"*|"-Oz"*)
	    reboot=false
	    run_profile="parallel"
	    ;;
	*)
	    reboot=true
	    run_profile="serial"
	    ;;
    esac

    local bench_list bin cc gnu_target sysroot toolchain
    gnu_target=$(print_gnu_target ${rr[target]})
    sysroot="$(pwd)/abe/builds/destdir/x86_64-pc-linux-gnu/$gnu_target/libc"
    case "${rr[toolchain]}" in
	llvm)
	    local llvm_target
	    llvm_target=$(echo "$gnu_target" | sed -e "s/^arm-/armv7a-/")
	    bmk_flags="$bmk_flags --target=$llvm_target --sysroot=$sysroot"
	    bmk_ldflags="$bmk_flags"
	    # Use LLD for LLVM configurations.
	    # Also, BFD linker crashes for AArch32 LTO builds,
	    # see https://projects.linaro.org/browse/LLVM-562 .
	    case "$bmk_ldflags" in
		*"-fuse-ld="*) ;;
		*) bmk_ldflags="$bmk_ldflags -fuse-ld=lld" ;;
	    esac
	    bin="$(pwd)/llvm-install/bin"
	    cc="$bin/"
	    toolchain="llvm"
	    ;;
	gnu|gnu_eabi)
	    bmk_ldflags="$bmk_flags"
	    bin="$(pwd)/abe/builds/destdir/x86_64-pc-linux-gnu/bin"
	    cc="$bin/$gnu_target-"
	    toolchain="gnu"
	    ;;
    esac
    case "${rr[toolchain]}:${benchmarks[@]}" in
	llvm:spec2k6) bench_list="c_and_cxx" ;;
	gnu:spec2k6) bench_list="all" ;;
	llvm:spec2017) bench_list="spec2017_speed_nofortran" ;;
	gnu:spec2017) bench_list="spec2017_speed" ;;
	*) bench_list="${benchmarks[*]}" ;;
    esac
    # shellcheck disable=SC2154
    sysroot="ssh://$ssh_host:$ssh_port:$sysroot"

    local hw_tag
    case "${rr[ci_project]}:${rr[target]}" in
	*_apm_32*:*) hw_tag=apm_32 ;;
	*_apm_64*:*) hw_tag=apm_64 ;;
	*_apm*:arm*) hw_tag=apm_32 ;;
	*_apm*:aarch64) hw_tag=apm_64 ;;
	*_sq_32*:*) hw_tag=sq_32 ;;
	*_sq_64*:*) hw_tag=sq_64 ;;
	*_sq*:arm*) hw_tag=sq_32 ;;
	*_sq*:aarch64) hw_tag=sq_64 ;;
	*_tk1_32*:*) hw_tag=tk1_32 ;;
	*_tk1*:arm*) hw_tag=tk1_32 ;;
	*_tx1_32*:*) hw_tag=tx1_32 ;;
	*_tx1_64*:*) hw_tag=tx1_64 ;;
	*_tx1*:arm*) hw_tag=tx1_32 ;;
	*_tx1*:aarch64) hw_tag=tx1_64 ;;
	*_stm32*:arm*) hw_tag=stm32 ;;
	*) echo "ERROR: Unknown hw_tag for ${rr[ci_project]}:${rr[target]}"; exit 1 ;;
    esac

    local hw image_arch toolchain_proto
    toolchain_proto=ssh
    case "$hw_tag" in
	apm_32) hw=apm; image_arch=armhf ;;
	apm_64) hw=apm; image_arch=arm64 ;;
	sq_32) hw=sq; image_arch=armhf ;;
	sq_64) hw=sq; image_arch=arm64 ;;
	tk1_32) hw=tk1; image_arch=armhf ;;
	tx1_32) hw=tx1; image_arch=armhf ;;
	tx1_64) hw=tx1; image_arch=arm64 ;;
	stm32)
	    hw=stm32; image_arch=amd64
	    # When running benchmarks on stm32, we prefer to rsync the
	    # toolchain to the board's host machine -- dev-02.tcwglab.
	    toolchain_proto=rsync
	    ;;
	*) echo "ERROR: Unknown hw_tag $hw_tag"; exit 1 ;;
    esac

    local results_id="$hw_tag/${rr[ci_project]}/${rr[mode]}-${rr[ci_config]}/@build_num@"

    # shellcheck disable=SC2154
    remote_exec "ci.linaro.org:2222::-l $USER@linaro.org" \
		build tcwg-benchmark -w \
		-p bmk_hw=$hw \
		-p bench_list="$bench_list" \
		-p cflags="$bmk_flags" \
		-p ldflags="$bmk_ldflags" \
		-p testmode=benchmark \
		-p displaytag="${rr[ci_project]}/${rr[mode]}-${rr[ci_config]}" \
		-p ignore_errors=true \
		-p toolchain_url=$toolchain_proto://$ssh_host:$ssh_port:$cc \
		-p toolchain_type=$toolchain \
		-p sysroot="$sysroot" \
		-p results_id="$results_id" \
		-p reboot="$reboot" \
		-p run_profile="$run_profile" \
		-p image_arch="$image_arch" \
		${scripts_branch+-p scripts_branch="$scripts_branch"} \
		${bmk_branch+-p bmk_branch="$bmk_branch"} \
	| tee $run_step_artifacts/benchmark-start.log

    local build_num
    build_num=$(cat $run_step_artifacts/benchmark-start.log \
		    | sed -e "s/.*#\([0-9]\+\).*/\1/")
    assert_with_msg "Benchmark build number should not be 0!" \
      [ "$build_num" -gt "0" ]

    local build_status
    local build_ret
    while true; do
	(remote_exec "ci.linaro.org:2222::-l $USER@linaro.org" \
		     console tcwg-benchmark -n 1 -f $build_num || true) \
	    | tee -a $run_step_artifacts/benchmark.log

	# Ssh connection to ci.linaro.org occasionally drops.  We need
	# to check whether benchmarking has finished, and, if not, continue
	# to watch its output.
	build_status=$(tail -n 1 $run_step_artifacts/benchmark.log)
	case "$build_status" in
	    "Finished: SUCCESS")
	        build_ret=0
	        break
	        ;;
	    "Finished: "*)
		echo "# Benchmarking infra is offline:" >> ${rr[top_artifacts]}/results
		echo "-$EXTERNAL_FAIL" >> ${rr[top_artifacts]}/results
	        build_ret=1
	        break
	        ;;
	esac

	# Sleep a little to avoid flooding ci.linaro.org on transient ssh
	# failures.
	sleep 5
    done

    echo "$results_id" | sed -e "s/@build_num@/$build_num/g" \
			     > "$results_id_file"
    return $build_ret
    )
}

# Compare results obtained from perf data between $1 and $2
# and generate results-compare.csv
compare_results_perf ()
{
    (
    set -euf -o pipefail

    case "${cflags[0]}" in
	"-Os"*|"-Oz"*)
	    # We use 1% tolerance for binary size
	    # and 10% tolerance for symbol size.
	    exe_threshold=101
	    symbol_threshold=110
	    ;;
	*)
	    # We use 3% tolerance for binary speed
	    # and 15% tolerance for symbol speed.
	    exe_threshold=103
	    symbol_threshold=115
	    # Reduce thresholds when bisecting to avoid considering borderline
	    # regressions as spurious.  This should break cycles of build and
	    # bisect jobs triggering each other on borderline regressions.
	    if [ x"${rr[mode]}" = x"bisect" ]; then
		exe_threshold=102
		symbol_threshold=110
	    fi
	    ;;
    esac

    local -a arr
    local bmk symbol time size result prev_bmk
    echo "bmk,symbol,result" > $run_step_artifacts/results-compare.csv
    printf "extra_build_params=" > $run_step_artifacts/extra-bisect-params

    # Read result lines from <(tail -n +2 ...) below.
    # "-n +2" is to skip the header line.
    prev_bmk=""
    while IFS=, read -a arr; do
	bmk=${arr[0]}
	symbol=${arr[1]}
	time=${arr[2]}
	size=${arr[3]}
	case "${cflags[0]}" in
	    "-Os"*|"-Oz"*) metric="$size" ;;
	    *) metric="$time" ;;
	esac

	# Skip case where we have no info ("n/a")
	if [ "$metric" != "n/a" ]; then
	    # Remove padding from the tail of $symbol (padding is added by
	    # csvs2table.py for better formatting).
	    local short_symbol="${symbol%%[ ]*}"
	    case "$short_symbol" in
		"["*) threshold=$symbol_threshold ;;
		*"_base.default") threshold=$exe_threshold ;;
		*) threshold=$symbol_threshold ;;
	    esac
	    if ! [ "$metric" -le "$threshold" ]; then
		result=100
		echo "# $bmk,$symbol regressed by $metric" >> $run_step_artifacts/results.regressions
		if [ x"$bmk" != x"$prev_bmk" ]; then
		    printf "++benchmarks %s " $bmk >> $run_step_artifacts/extra-bisect-params
		    prev_bmk="$bmk"
		fi
	    else
		result=1
	    fi
	    echo "$bmk,$symbol,$result" >> $run_step_artifacts/results-compare.csv
	fi
    done < <(tail -n +2 $run_step_artifacts/results.csv)
    printf "\n" >> $run_step_artifacts/extra-bisect-params
    )
}

compare_results_vect ()
{
    (
    set -euf -o pipefail
    echo "bmk,symbol,result" > $run_step_artifacts/results-compare.csv

    while IFS=, read -a arr; do
        bmk=${arr[0]}
        # hack to trim padding
        symbol=$(echo ${arr[1]} | xargs)
        base_num_vect_loops=${arr[3]}
        target_num_vect_loops=${arr[4]}
        if (( base_num_vect_loops > target_num_vect_loops )); then
	    echo "$bmk, $symbol, $base_num_vect_loops, $target_num_vect_loops" \
		>> $run_step_artifacts/results-compare.csv
        fi
    done < <(tail -n +2 $run_step_artifacts/results.csv)
    )
}

compare_results ()
{
    (
    set -euf -o pipefail

    local metric=$1
    local ref_results_id="$2"
    local new_results_id="$3"
    local cmp_options="$4"

    local results_ref results_new
    results_ref=$(cat $ref_results_id)
    results_new=$(cat $new_results_id)

    case "${rr[target]}" in
        "arm_eabi")
            cmp_options="$cmp_options --has_perf_logs no"
            ;;
    esac

    $scripts/tcwg-benchmark-results.sh \
        --results_ref $results_ref ++results $results_new \
        --top_artifacts "$run_step_artifacts" --verbose $verbose \
        --metric "$metric" $cmp_options \
        > $run_step_artifacts/results.log 2>&1

    case $metric in
	"perf")
	    compare_results_perf
	    ;;
	"vect")
	    compare_results_vect
	    ;;
	*)
	    echo "Invalid metric: $metric";
	    exit 1
	    ;;
    esac
    )
}

# Exit with code 0 if no new regressions between results_id-1 and -2 compared to
# regression between results_id-1 and -2 in base-artifacts/.
no_regression_vs_p ()
{
    (
    set -euf -o pipefail

    local ref_artifacts=$1
    local new_artifacts=$2

    # Check for build and correctness regressions.
    no_build_regression_p "$@"

    # Generate ref-results-compare.csv.  The value of "1" means that the result
    # in the 2nd run is no worse than the result in the 1st run (as expected).
    # The value of "100" means that the result in the 2nd run is worse than
    # the result in the 1st run (unexpected).
    # Note that we can grab previously-generated ref-results-compares.csv from
    # base-artifacts/, but it could have been generated with an older version
    # of scripts, so it's safer and more resilient to re-generate it from original
    # perf data.
    if [ ! -f "$ref_artifacts/results_id-1" ] || [ ! -f "$ref_artifacts/results_id-2" ]; then
	return 0
    fi
    # <Workaround> missing reference results, which we have listed in
    # tcwg-benchmark-results.broken-list.  Once all entries referencing missing
    # results are discarded, we'll remove this workaround.
    # Otherwise compare_results will fail while fetching baseline results,
    # and we'll consider this failure as a regression.
    if cat "$scripts/tcwg-benchmark-results.broken-list" \
	    | grep -q "^$(cat $ref_artifacts/results_id-1)\$\|^$(cat $ref_artifacts/results_id-2)\$"; then
	return 0
    fi
    # </Workaround>
    compare_results "${rr[metric]}" "$ref_artifacts/results_id-1" "$ref_artifacts/results_id-2" \
		    "--num_dsos 1 --num_symbols 0"

    while IFS= read -r -d '' i
    do
        mv $i "$(dirname $i)"/ref-"$(basename $i)"
    done < <(find $run_step_artifacts/ -type f -name "results*" -print0)

    # Similarly, generate new-results-compare.csv.
    if [ ! -f "$new_artifacts/results_id-1" ] || [ ! -f "$new_artifacts/results_id-2" ]; then
	return 1
    fi
    compare_results "${rr[metric]}" "$new_artifacts/results_id-1" "$new_artifacts/results_id-2" \
		    "--num_dsos 1 --num_symbols 0"
    while IFS= read -r -d '' i
    do
        mv $i "$(dirname $i)"/new-"$(basename $i)"
    done < <(find $run_step_artifacts/ -type f -name "results*" -print0)

    # Now compare the two reports.
    # If "ref" has value of "100" (bad state), and "new" has value of "100"
    # (also bad state), then we get no change, no regression, and final value
    # of 100% * 100/100 == 100.
    #
    # If "ref" has value of "1" (good state), and "new" has value of "1"
    # (also good state), then we get no change, no regression, and final value
    # of 100% * 1/1 == 100.
    #
    # If "ref" has value of "100" (bad state), and "new" has value of "1"
    # (good state), then we get a progression, and final value
    # of 100% * 1/100 == 1.
    #
    # If "ref" has value of "1" (good state), and "new" has value of "100"
    # (bad state), then we get a regression, and final value
    # of 100% * 100/1 == 10000.  We detect this below by comparing vs "5000".
    $scripts/../bmk-scripts/csvs2table.py -p 0 --relative $run_step_artifacts/ref-results-compare.csv $run_step_artifacts/new-results-compare.csv > $run_step_artifacts/results-compare.csv

    local -a arr
    local bmk symbol result status prev_bmk
    local -a bisect_bmks

    # Read result lines from <(tail -n +2 ...) below.
    # "-n +2" is to skip the header line.  Set $status to "1" if there is
    # a regression.
    status=0
    prev_bmk=""
    # Delete results.regressions generated by compare_results() calls above.
    rm -f $run_step_artifacts/results.regressions
    while IFS=, read -a arr; do
	bmk=${arr[0]}
	symbol=${arr[1]}
	result=${arr[2]}
	if ! [ "$result" -le "5000" ]; then
	    echo "# $bmk,$symbol regressed" >> $run_step_artifacts/results.regressions
	    status=1
	    if [ x"$bmk" != x"$prev_bmk" ]; then
		bisect_bmks+=("++benchmarks" "$bmk")
		prev_bmk="$bmk"
	    fi
	fi
    done < <(tail -n +2 $run_step_artifacts/results-compare.csv)
    echo "extra_build_params=${bisect_bmks[*]}" > $run_step_artifacts/extra-bisect-params
    return $status
    )
}

# Exit with code 0 if no regression compared to base-artifacts/.
# Inspect build results ./results and performance results in ./results_id.
no_regression_to_base_p ()
{
    (
    set -euf -o pipefail

    no_build_regression_p "$@"

    local ref_artifacts=$1
    local new_artifacts=$2

    if ! [ -f "$ref_artifacts/results_id" ]; then
	return 0
    fi
    # <Workaround> missing reference results, which we have listed in
    # tcwg-benchmark-results.broken-list.  Once all entries referencing missing
    # results are discarded, we'll remove this workaround.
    # Otherwise compare_results will fail while fetching baseline results,
    # and we'll consider this failure as a regression.
    if cat "$scripts/tcwg-benchmark-results.broken-list" \
	    | grep -q "^$(cat $ref_artifacts/results_id)\$"; then
	return 0
    fi
    # </Workaround>
    if ! [ -f "$new_artifacts/results_id" ]; then
	return 1
    fi

    # Make sure there is no stray results.regression file, which we use
    # as failure marker.
    assert ! [ -f $run_step_artifacts/results.regressions ]

    local compare_opts=""
    case "${cflags[0]}" in
	*"_LTO"*) compare_opts="--num_symbols 0 --entry_threshold 10" ;;
    esac
    compare_results "${rr[metric]}" "$ref_artifacts/results_id" "$new_artifacts/results_id" "$compare_opts"
    if [ -f $run_step_artifacts/results.regressions ]; then
	return 1
    fi
    return 0
    )
}

# Implement rr[breakup_updated_components] hook.
tcwg_bmk_breakup_updated_components ()
{
    (
    set -euf -o pipefail

    # Compiler changes tend to cause the most regressions.
    # Breakup updated components into compiler and the rest of components
    # to reduce the number of builds.
    local cc
    case "${rr[toolchain]}" in
	llvm) cc="llvm" ;;
	gnu|gnu_eabi) cc="gcc" ;;
	*) assert false ;;
    esac

    if print_updated_components "\n" | grep -q "^$cc\$"; then
	echo "$cc"
	print_updated_components "\n" | grep -v "^$cc\$" | tr '\n' ' ' | sed -e "s/ \$//g"
	echo
    else
	print_updated_components "\n"
    fi
    )
}
# shellcheck disable=SC2154
rr[breakup_updated_components]=tcwg_bmk_breakup_updated_components

run_step stop_on_fail -10 reset_artifacts
run_step stop_on_fail x prepare_abe
run_step skip_on_fail -9 build_abe binutils
run_step skip_on_fail -8 build_abe stage1 -- "${gcc_override_configure[@]}"
run_step skip_on_fail x clean_sysroot
case "${rr[components]}" in
    *glibc*)
	run_step skip_on_fail -7 build_abe linux
	run_step skip_on_fail -6 build_abe glibc
	;;
    *newlib*)
	run_step skip_on_fail -6 build_abe newlib
	;;
esac

patch_branch=""
if [ x"${rr[metric]}" = x"vect" ]; then
    patch_branch="--patch linaro-local/vect-metric-branch"
fi

run_step skip_on_fail -5 build_abe stage2 -- $patch_branch "${gcc_override_configure[@]}"

case "${rr[toolchain]}" in
    llvm) run_step skip_on_fail -3 build_llvm true ;;
esac
case "${#cflags[@]}" in
    2)
	# Don't bisect benchmark build/run failures in *-vs-* configurations.
	# Bisections happen only for regressions with build scores >=0,
	# which will happen if benchmark "${cflags[1]}" succeeds.
	run_step skip_on_fail -1 benchmark "${cflags[0]}" -- ${rr[top_artifacts]}/results_id-1
	run_step skip_on_fail 0 benchmark "${cflags[1]}" -- ${rr[top_artifacts]}/results_id-2
	# Set final "build" score to "1" for compatibility with older results
	run_step skip_on_fail 1 true
	# shellcheck disable=SC2154
	rr[no_regression_p]=no_regression_vs_p
	run_step reset_on_fail x check_regression
	;;
    1)
	# Bisect benchmark build/run failures in non-vs configurations.
	# Set score to "0" with "true".
	run_step skip_on_fail 0 true
	run_step skip_on_fail 1 benchmark "${cflags[0]}" -- ${rr[top_artifacts]}/results_id
	rr[no_regression_p]=no_regression_to_base_p
	run_step reset_on_fail x check_regression
	;;
esac
run_step stop_on_fail x update_baseline
run_step stop_on_fail x push_baseline

trap "" EXIT