mirrors/cpython
1
0
Fork 0
mirror of https://github.com/python/cpython.git synced 2025-09-26 10:19:53 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 8

GH-109330: Dump and compare stats using opcode names, not numbers (GH-109335)

Browse source
This commit is contained in:
Michael Droettboom 2023-09-12 17:12:57 -04:00 • committed by GitHub
parent 90cf345ed4
commit 5dcbbd8861
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 35 additions and 53 deletions
Download patch file Download diff file Expand all files Collapse all files

16
Python/specialize.c
View file

@ -123,7 +123,7 @@ _Py_GetSpecializationStats(void) {
#define PRINT_STAT(i, field) \ #define PRINT_STAT(i, field) \
if (stats[i].field) { \ if (stats[i].field) { \
fprintf(out, " opcode[%d]." #field " : %" PRIu64 "\n", i, stats[i].field); \ fprintf(out, " opcode[%s]." #field " : %" PRIu64 "\n", _PyOpcode_OpName[i], stats[i].field); \
} }
static void static void
@ -131,11 +131,11 @@ print_spec_stats(FILE *out, OpcodeStats *stats)
{ {
/* Mark some opcodes as specializable for stats, /* Mark some opcodes as specializable for stats,
* even though we don't specialize them yet. */ * even though we don't specialize them yet. */
fprintf(out, "opcode[%d].specializable : 1\n", BINARY_SLICE); fprintf(out, "opcode[BINARY_SLICE].specializable : 1\n");
fprintf(out, "opcode[%d].specializable : 1\n", STORE_SLICE); fprintf(out, "opcode[STORE_SLICE].specializable : 1\n");
for (int i = 0; i < 256; i++) { for (int i = 0; i < 256; i++) {
if (_PyOpcode_Caches[i]) { if (_PyOpcode_Caches[i]) {
fprintf(out, "opcode[%d].specializable : 1\n", i); fprintf(out, "opcode[%s].specializable : 1\n", _PyOpcode_OpName[i]);
} }
PRINT_STAT(i, specialization.success); PRINT_STAT(i, specialization.success);
PRINT_STAT(i, specialization.failure); PRINT_STAT(i, specialization.failure);
@ -147,14 +147,14 @@ print_spec_stats(FILE *out, OpcodeStats *stats)
for (int j = 0; j < SPECIALIZATION_FAILURE_KINDS; j++) { for (int j = 0; j < SPECIALIZATION_FAILURE_KINDS; j++) {
uint64_t val = stats[i].specialization.failure_kinds[j]; uint64_t val = stats[i].specialization.failure_kinds[j];
if (val) { if (val) {
fprintf(out, " opcode[%d].specialization.failure_kinds[%d] : %" fprintf(out, " opcode[%s].specialization.failure_kinds[%d] : %"
PRIu64 "\n", i, j, val); PRIu64 "\n", _PyOpcode_OpName[i], j, val);
} }
} }
for (int j = 0; j < 256; j++) { for (int j = 0; j < 256; j++) {
if (stats[i].pair_count[j]) { if (stats[i].pair_count[j]) {
fprintf(out, "opcode[%d].pair_count[%d] : %" PRIu64 "\n", fprintf(out, "opcode[%s].pair_count[%s] : %" PRIu64 "\n",
i, j, stats[i].pair_count[j]); _PyOpcode_OpName[i], _PyOpcode_OpName[j], stats[i].pair_count[j]);
} }
} }
} }

72
Tools/scripts/summarize_stats.py
View file

@ -16,22 +16,6 @@ if os.name == "nt":
else: else:
DEFAULT_DIR = "/tmp/py_stats/" DEFAULT_DIR = "/tmp/py_stats/"
#Create list of all instruction names
specialized = iter(opcode._specialized_opmap.keys())
opname = ["<0>"]
for name in opcode.opname[1:]:
if name.startswith("<"):
try:
name = next(specialized)
except StopIteration:
pass
opname.append(name)
# opcode_name --> opcode
# Sort alphabetically.
opmap = {name: i for i, name in enumerate(opname)}
opmap = dict(sorted(opmap.items()))
TOTAL = "specialization.hit", "specialization.miss", "execution_count" TOTAL = "specialization.hit", "specialization.miss", "execution_count"
def format_ratio(num, den): def format_ratio(num, den):
@ -200,12 +184,12 @@ def gather_stats(input):
raise ValueError(f"{input:r} is not a file or directory path") raise ValueError(f"{input:r} is not a file or directory path")
def extract_opcode_stats(stats): def extract_opcode_stats(stats):
opcode_stats = [ {} for _ in range(256) ] opcode_stats = collections.defaultdict(dict)
for key, value in stats.items(): for key, value in stats.items():
if not key.startswith("opcode"): if not key.startswith("opcode"):
continue continue
n, _, rest = key[7:].partition("]") name, _, rest = key[7:].partition("]")
opcode_stats[int(n)][rest.strip(".")] = value opcode_stats[name][rest.strip(".")] = value
return opcode_stats return opcode_stats
def parse_kinds(spec_src, prefix="SPEC_FAIL"): def parse_kinds(spec_src, prefix="SPEC_FAIL"):
@ -246,11 +230,10 @@ def categorized_counts(opcode_stats):
specialized_instructions = { specialized_instructions = {
op for op in opcode._specialized_opmap.keys() op for op in opcode._specialized_opmap.keys()
if "__" not in op} if "__" not in op}
for i, opcode_stat in enumerate(opcode_stats): for name, opcode_stat in opcode_stats.items():
if "execution_count" not in opcode_stat: if "execution_count" not in opcode_stat:
continue continue
count = opcode_stat['execution_count'] count = opcode_stat['execution_count']
name = opname[i]
if "specializable" in opcode_stat: if "specializable" in opcode_stat:
not_specialized += count not_specialized += count
elif name in specialized_instructions: elif name in specialized_instructions:
@ -314,13 +297,13 @@ def emit_table(header, rows):
def calculate_execution_counts(opcode_stats, total): def calculate_execution_counts(opcode_stats, total):
counts = [] counts = []
for i, opcode_stat in enumerate(opcode_stats): for name, opcode_stat in opcode_stats.items():
if "execution_count" in opcode_stat: if "execution_count" in opcode_stat:
count = opcode_stat['execution_count'] count = opcode_stat['execution_count']
miss = 0 miss = 0
if "specializable" not in opcode_stat: if "specializable" not in opcode_stat:
miss = opcode_stat.get("specialization.miss") miss = opcode_stat.get("specialization.miss")
counts.append((count, opname[i], miss)) counts.append((count, name, miss))
counts.sort(reverse=True) counts.sort(reverse=True)
cumulative = 0 cumulative = 0
rows = [] rows = []
@ -381,16 +364,17 @@ def get_defines():
def emit_specialization_stats(opcode_stats): def emit_specialization_stats(opcode_stats):
defines = get_defines() defines = get_defines()
with Section("Specialization stats", summary="specialization stats by family"): with Section("Specialization stats", summary="specialization stats by family"):
for i, opcode_stat in enumerate(opcode_stats): for name, opcode_stat in opcode_stats.items():
name = opname[i]
print_specialization_stats(name, opcode_stat, defines) print_specialization_stats(name, opcode_stat, defines)
def emit_comparative_specialization_stats(base_opcode_stats, head_opcode_stats): def emit_comparative_specialization_stats(base_opcode_stats, head_opcode_stats):
defines = get_defines() defines = get_defines()
with Section("Specialization stats", summary="specialization stats by family"): with Section("Specialization stats", summary="specialization stats by family"):
for i, (base_opcode_stat, head_opcode_stat) in enumerate(zip(base_opcode_stats, head_opcode_stats)): opcodes = set(base_opcode_stats.keys()) & set(head_opcode_stats.keys())
name = opname[i] for opcode in opcodes:
print_comparative_specialization_stats(name, base_opcode_stat, head_opcode_stat, defines) print_comparative_specialization_stats(
opcode, base_opcode_stats[opcode], head_opcode_stats[opcode], defines
)
def calculate_specialization_effectiveness(opcode_stats, total): def calculate_specialization_effectiveness(opcode_stats, total):
basic, not_specialized, specialized = categorized_counts(opcode_stats) basic, not_specialized, specialized = categorized_counts(opcode_stats)
@ -407,12 +391,12 @@ def emit_specialization_overview(opcode_stats, total):
for title, field in (("Deferred", "specialization.deferred"), ("Misses", "specialization.miss")): for title, field in (("Deferred", "specialization.deferred"), ("Misses", "specialization.miss")):
total = 0 total = 0
counts = [] counts = []
for i, opcode_stat in enumerate(opcode_stats): for name, opcode_stat in opcode_stats.items():
# Avoid double counting misses # Avoid double counting misses
if title == "Misses" and "specializable" in opcode_stat: if title == "Misses" and "specializable" in opcode_stat:
continue continue
value = opcode_stat.get(field, 0) value = opcode_stat.get(field, 0)
counts.append((value, opname[i])) counts.append((value, name))
total += value total += value
counts.sort(reverse=True) counts.sort(reverse=True)
if total: if total:
@ -539,29 +523,27 @@ def emit_comparative_gc_stats(base_stats, head_stats):
def get_total(opcode_stats): def get_total(opcode_stats):
total = 0 total = 0
for opcode_stat in opcode_stats: for opcode_stat in opcode_stats.values():
if "execution_count" in opcode_stat: if "execution_count" in opcode_stat:
total += opcode_stat['execution_count'] total += opcode_stat['execution_count']
return total return total
def emit_pair_counts(opcode_stats, total): def emit_pair_counts(opcode_stats, total):
pair_counts = [] pair_counts = []
for i, opcode_stat in enumerate(opcode_stats): for name_i, opcode_stat in opcode_stats.items():
if i == 0:
continue
for key, value in opcode_stat.items(): for key, value in opcode_stat.items():
if key.startswith("pair_count"): if key.startswith("pair_count"):
x, _, _ = key[11:].partition("]") name_j, _, _ = key[11:].partition("]")
if value: if value:
pair_counts.append((value, (i, int(x)))) pair_counts.append((value, (name_i, name_j)))
with Section("Pair counts", summary="Pair counts for top 100 pairs"): with Section("Pair counts", summary="Pair counts for top 100 pairs"):
pair_counts.sort(reverse=True) pair_counts.sort(reverse=True)
cumulative = 0 cumulative = 0
rows = [] rows = []
for (count, pair) in itertools.islice(pair_counts, 100): for (count, pair) in itertools.islice(pair_counts, 100):
i, j = pair name_i, name_j = pair
cumulative += count cumulative += count
rows.append((opname[i] + " " + opname[j], count, format_ratio(count, total), rows.append((f"{name_i} {name_j}", count, format_ratio(count, total),
format_ratio(cumulative, total))) format_ratio(cumulative, total)))
emit_table(("Pair", "Count:", "Self:", "Cumulative:"), emit_table(("Pair", "Count:", "Self:", "Cumulative:"),
rows rows
@ -577,18 +559,18 @@ def emit_pair_counts(opcode_stats, total):
successors[first][second] = count successors[first][second] = count
total_predecessors[second] += count total_predecessors[second] += count
total_successors[first] += count total_successors[first] += count
for name, i in opmap.items(): for name in opcode_stats.keys():
total1 = total_predecessors[i] total1 = total_predecessors[name]
total2 = total_successors[i] total2 = total_successors[name]
if total1 == 0 and total2 == 0: if total1 == 0 and total2 == 0:
continue continue
pred_rows = succ_rows = () pred_rows = succ_rows = ()
if total1: if total1:
pred_rows = [(opname[pred], count, f"{count/total1:.1%}") pred_rows = [(pred, count, f"{count/total1:.1%}")
for (pred, count) in predecessors[i].most_common(5)] for (pred, count) in predecessors[name].most_common(5)]
if total2: if total2:
succ_rows = [(opname[succ], count, f"{count/total2:.1%}") succ_rows = [(succ, count, f"{count/total2:.1%}")
for (succ, count) in successors[i].most_common(5)] for (succ, count) in successors[name].most_common(5)]
with Section(name, 3, f"Successors and predecessors for {name}"): with Section(name, 3, f"Successors and predecessors for {name}"):
emit_table(("Predecessors", "Count:", "Percentage:"), emit_table(("Predecessors", "Count:", "Percentage:"),
pred_rows pred_rows