#!/bin/bash
# Which decompression branch substr(v, 500000, 100) takes on 1 MB zstd values
# in a UTF8 and a SQL_ASCII database, and how many page faults the backend
# takes per value with the reused context and with a fresh one.  No timing.
#   branch_faults.sh REUSE_BINDIR FRESH_BINDIR WORKDIR
#     REUSE_BINDIR  bin/ of a build with the 0003 decode code (reused DCtx)
#     FRESH_BINDIR  bin/ of the same tree with the reuse undone
#                   (nocfbot-v9n-without-reuse.diff.txt: the v7 decode code)
#     WORKDIR       empty scratch directory; the cluster is ~30 MB
# One small cluster (shared_buffers 32MB), 20 values of 1 MB per database,
# the "random" corpus of bench.sh.  Queries run in single-user mode, under
# gdb (breakpoints at the entry of zstd_decompress_datum, the full branch,
# and zstd_decompress_datum_slice, printing slicelength) and under
# perf stat -e page-faults (one query vs six; faults per value =
# (six - one) / 5 / 20, which cancels startup).  Both builds must share the
# catalog version, as they do when only toast_compression.c differs.
set -eu
R=$1 F=$2 W=$3
D=$W/cl; P=$((56400 + RANDOM % 100))
"$R/initdb" -D $D -A trust --no-sync -U postgres -E SQL_ASCII --locale=C >/dev/null
cat >> $D/postgresql.conf <<CONF
port = $P
unix_socket_directories = '$W'
listen_addresses = ''
shared_buffers = 32MB
autovacuum = off
default_toast_compression = zstd
CONF
"$R/pg_ctl" -D $D -l $W/log -w start >/dev/null
q() { "$R/psql" -X -qAt -h $W -p $P -U postgres -v ON_ERROR_STOP=1 "$@"; }
q -d postgres -c "CREATE DATABASE u8 ENCODING 'UTF8' TEMPLATE template0 LC_COLLATE 'C' LC_CTYPE 'C'"
q -d postgres -c "CREATE DATABASE sqlascii ENCODING 'SQL_ASCII' TEMPLATE template0 LC_COLLATE 'C' LC_CTYPE 'C'"
for db in u8 sqlascii; do
  q -d $db -c "SELECT setseed(0.42)" -c "CREATE TABLE t (id int, v text)" \
    -c "INSERT INTO t SELECT g, left(string_agg(md5(random()::text) || encode(sha256(random()::text::bytea), 'base64'), ''), 1000000)
          FROM generate_series(1, 20) g, generate_series(1, 1000000 / 76 + 1) i GROUP BY g"
  q -d $db -c "SELECT format('db=%s encoding=%s eml=%s values=%s bytes=%s method=%s slice_md5=%s', current_database(),
                 current_setting('server_encoding'), pg_encoding_max_length(pg_char_to_encoding(current_setting('server_encoding'))),
                 count(*), min(octet_length(v)), string_agg(DISTINCT pg_column_compression(v), ','),
                 md5(string_agg(substr(v, 500000, 100), '' ORDER BY id))) FROM t"
done
"$R/pg_ctl" -D $D -m fast -w stop >/dev/null
cat > $W/br.gdb <<'GDB'
set pagination off
set confirm off
break *zstd_decompress_datum
commands
silent
continue
end
break *zstd_decompress_datum_slice
commands
silent
printf "slicelength=%d\n", $rsi
continue
end
run
info breakpoints
GDB
single() { "$1/postgres" --single -D $D -c shared_buffers=32MB $2; }
for db in u8 sqlascii; do
  for e in "substr(v, 500000, 100)" "substr(v, 1, 100)" "v || 'x'"; do
    out=$(echo "SELECT sum(length($e)) FROM t" | gdb -q -batch -x $W/br.gdb --args "$R/postgres" --single -D $D -c shared_buffers=32MB $db 2>&1)
    full=$(echo "$out" | sed -n '/^Num/,$p' | awk '/<zstd_decompress_datum>/{s=1;next} /<zstd_decompress_datum_slice>/{s=2;next} /already hit/{h[s]=$4} END{print (h[1]?h[1]:0)}')
    sl=$(echo "$out" | sed -n '/^Num/,$p' | awk '/<zstd_decompress_datum_slice>/{s=1;next} s&&/already hit/{print $4; exit}')
    lens=$(echo "$out" | grep -oE 'slicelength=[0-9]+' | sort -u | tr '\n' ' ')
    echo "branch db=$db expr=\"$e\" full_entries=$full slice_entries=${sl:-0} ${lens}"
  done
done
pf() {  # bindir db expr n
  for i in $(seq $4); do echo "SELECT sum(length($3)) FROM t;"; done |
    perf stat -x, -e page-faults "$1/postgres" --single -D $D -c shared_buffers=32MB $2 2>&1 >/dev/null |
    awk -F, '/page-faults/{print $1}'
}
for c in "sqlascii|substr(v, 500000, 100)" "sqlascii|substr(v, 1, 100)" "sqlascii|v || 'x'" "u8|substr(v, 500000, 100)"; do
  db=${c%%|*} e=${c#*|}
  for rep in 1 2 3; do
    for b in reuse fresh; do
      [ $b = reuse ] && bin=$R || bin=$F
      one=$(pf $bin $db "$e" 1); six=$(pf $bin $db "$e" 6)
      echo "faults db=$db expr=\"$e\" build=$b rep=$rep one=$one six=$six per_value=$(( (six - one) / 100 ))"
    done
  done
done
rm -rf $D $W/br.gdb $W/log