llvm-project/bolt/test/link_fdata.py
Amir Ayupov 7085065c02
[BOLT] Support pre-aggregated returns (#143296)
Intel's Architectural LBR supports capturing branch type information
as part of LBR stack (SDM Vol 3B, part 2, October 2024):
```
20.1.3.2 Branch Types
The IA32_LBR_x_INFO.BR_TYPE and IA32_LER_INFO.BR_TYPE fields encode
the branch types as shown in Table 20-3.

Table 20-3. IA32_LBR_x_INFO and IA32_LER_INFO Branch Type Encodings

Encoding | Branch Type
  0000B  | COND
  0001B  | NEAR_IND_JMP
  0010B  | NEAR_REL_JMP
  0011B  | NEAR_IND_CALL
  0100B  | NEAR_REL_CALL
  0101B  | NEAR_RET
  011xB  | Reserved
  1xxxB  | OTHER_BRANCH

For a list of branch operations that fall into the categories above,
see Table 20-2. 

Table 20-2. Branch Type Filtering Details
Branch Type   | Operations Recorded
COND          | Jcc, J*CXZ, and LOOP*
NEAR_IND_JMP  | JMP r/m*
NEAR_REL_JMP  | JMP rel*
NEAR_IND_CALL | CALL r/m*
NEAR_REL_CALL | CALL rel* (excluding CALLs to the next sequential IP)
NEAR_RET      | RET (0C3H)
OTHER_BRANCH  | JMP/CALL ptr*, JMP/CALL m*, RET (0C8H), SYS*, 
interrupts, exceptions (other than debug exceptions), IRET, INT3, 
INTn, INTO, TSX Abort, EENTER, ERESUME, EEXIT, AEX, INIT, SIPI, RSM
```

Linux kernel can preserve branch type when `save_type` is enabled,
even if CPU does not support Architectural LBR:

f09079bd04/tools/perf/Documentation/perf-record.txt (L457-L460)

> - save_type: save branch type during sampling in case binary is not
available later.
For the platforms with Intel Arch LBR support (12th-Gen+ client or
4th-Gen Xeon+ server), the save branch type is unconditionally enabled
when the taken branch stack sampling is enabled.

Kernel-reported branch type values:

8c6bc74c7f/include/uapi/linux/perf_event.h (L251-L269)

This information is needed to disambiguate external returns (from
DSO/JIT) to an entry point or a landing pad, when BOLT can't
disassemble the branch source.

This patch adds new pre-aggregated types:
- return trace (R),
- external return fall-through (r).

For such types, the checks for fall-through start (not an entry or
a landing pad) are relaxed.

Depends on #143295.

Test Plan: updated callcont-fallthru.s
2025-06-20 03:17:08 -07:00

157 lines
5.9 KiB
Python
Executable File

#!/usr/bin/env python3
"""
This script reads the input from stdin, extracts all lines starting with
"# FDATA: " (or a given prefix instead of "FDATA"), parses the directives,
replaces symbol names ("#name#") with either symbol values or with offsets from
respective anchor symbols, and prints the resulting file to stdout.
"""
import argparse
import os
import shutil
import subprocess
import sys
import re
parser = argparse.ArgumentParser()
parser.add_argument("input")
parser.add_argument("objfile", help="Object file to extract symbol values from")
parser.add_argument("output")
parser.add_argument("prefix", nargs="?", default="FDATA", help="Custom FDATA prefix")
parser.add_argument("--nmtool", default="nm", help="Path to nm tool")
parser.add_argument("--no-lbr", action="store_true")
parser.add_argument("--no-redefine", action="store_true")
args = parser.parse_args()
# Regexes to extract FDATA lines from input and parse FDATA and pre-aggregated
# profile data
prefix_pat = re.compile(f"^# {args.prefix}: (.*)")
# FDATA records:
# <is symbol?> <closest elf symbol or DSO name> <relative FROM address>
# <is symbol?> <closest elf symbol or DSO name> <relative TO address>
# <number of mispredictions> <number of branches>
fdata_pat = re.compile(r"([01].*) (?P<mispred>\d+) (?P<exec>\d+)")
# Pre-aggregated profile:
# {T|R|S|E|B|F|f|r} <start> [<end>] [<ft_end>] <count> [<mispred_count>]
# <loc>: [<id>:]<offset>
preagg_pat = re.compile(r"(?P<type>[TRSBFfr]) (?P<offsets_count>.*)")
# No-LBR profile:
# <is symbol?> <closest elf symbol or DSO name> <relative address> <count>
nolbr_pat = re.compile(r"([01].*) (?P<count>\d+)")
# Replacement symbol: #symname#
replace_pat = re.compile(r"#(?P<symname>[^#]+)#")
# Read input and construct the representation of fdata expressions
# as (src_tuple, dst_tuple, mispred_count, exec_count) tuples, where src and dst
# are represented as (is_sym, anchor, offset) tuples
exprs = []
with open(args.input, "r") as f:
for line in f.readlines():
prefix_match = prefix_pat.match(line)
if not prefix_match:
continue
profile_line = prefix_match.group(1)
fdata_match = fdata_pat.match(profile_line)
preagg_match = preagg_pat.match(profile_line)
nolbr_match = nolbr_pat.match(profile_line)
if fdata_match:
src_dst, mispred, execnt = fdata_match.groups()
# Split by whitespaces not preceded by a backslash (negative lookbehind)
chunks = re.split(r"(?<!\\) +", src_dst)
# Check if the number of records separated by non-escaped whitespace
# exactly matches the format.
assert (
len(chunks) == 6
), f"ERROR: wrong format/whitespaces must be escaped:\n{line}"
exprs.append(("FDATA", (*chunks, mispred, execnt)))
elif nolbr_match:
loc, count = nolbr_match.groups()
# Split by whitespaces not preceded by a backslash (negative lookbehind)
chunks = re.split(r"(?<!\\) +", loc)
# Check if the number of records separated by non-escaped whitespace
# exactly matches the format.
assert (
len(chunks) == 3
), f"ERROR: wrong format/whitespaces must be escaped:\n{line}"
exprs.append(("NOLBR", (*chunks, count)))
elif preagg_match:
exprs.append(("PREAGG", preagg_match.groups()))
else:
exit("ERROR: unexpected input:\n%s" % line)
# Read nm output: <symbol value> <symbol type> <symbol name>
is_llvm_nm = os.path.basename(os.path.realpath(shutil.which(args.nmtool))) == "llvm-nm"
nm_output = subprocess.run(
[
args.nmtool,
"--defined-only",
"--special-syms" if is_llvm_nm else "--synthetic",
args.objfile,
],
text=True,
capture_output=True,
).stdout
# Populate symbol map
symbols = {}
for symline in nm_output.splitlines():
symval, _, symname = symline.split(maxsplit=2)
if symname in symbols and args.no_redefine:
continue
symbols[symname] = symval
def evaluate_symbol(issym, anchor, offsym):
sym_match = replace_pat.match(offsym)
if not sym_match:
# No need to evaluate symbol value, return as is
return f"{issym} {anchor} {offsym}"
symname = sym_match.group("symname")
assert symname in symbols, f"ERROR: symbol {symname} is not defined in binary"
# Evaluate to an absolute offset if issym is false
if issym == "0":
return f"{issym} {anchor} {symbols[symname]}"
# Evaluate symbol against its anchor if issym is true
assert anchor in symbols, f"ERROR: symbol {anchor} is not defined in binary"
anchor_value = int(symbols[anchor], 16)
symbol_value = int(symbols[symname], 16)
sym_offset = symbol_value - anchor_value
return f'{issym} {anchor} {format(sym_offset, "x")}'
def replace_symbol(matchobj):
"""
Expects matchobj to only capture one group which contains the symbol name.
"""
symname = matchobj.group("symname")
assert symname in symbols, f"ERROR: symbol {symname} is not defined in binary"
return symbols[symname]
with open(args.output, "w", newline="\n") as f:
if args.no_lbr:
print("no_lbr", file=f)
for etype, expr in exprs:
if etype == "FDATA":
issym1, anchor1, offsym1, issym2, anchor2, offsym2, execnt, mispred = expr
print(
evaluate_symbol(issym1, anchor1, offsym1),
evaluate_symbol(issym2, anchor2, offsym2),
execnt,
mispred,
file=f,
)
elif etype == "NOLBR":
issym, anchor, offsym, count = expr
print(evaluate_symbol(issym, anchor, offsym), count, file=f)
elif etype == "PREAGG":
# Replace all symbols enclosed in ##
print(expr[0], re.sub(replace_pat, replace_symbol, expr[1]), file=f)
else:
exit("ERROR: unhandled expression type:\n%s" % etype)