[PATCH v3 0/7] perf diff: diff cycles at basic block level

From: Jin Yao
Date: Mon Jun 17 2019 - 01:06:28 EST


In some cases small changes in hot loops can show big differences.
But it's difficult to identify these differences.

perf diff currently can only diff symbols (functions). We can also expand
it to diff cycles of individual programs blocks as reported by timed LBR.
This would allow to identify changes in specific code accurately.

With this patch set, for example,

# perf record -b ./div
# perf record -b ./div
# perf diff -c cycles

# Event 'cycles'
#
# Baseline Block cycles diff [start:end] Shared Object Symbol
# ........ .................................... ................ ....................................
#
49.03% -9 [ 4ef: 520] div [.] main
49.03% 0 [ 4e8: 4ea] div [.] main
49.03% 0 [ 4ef: 500] div [.] main
49.03% 0 [ 4ef: 51c] div [.] main
49.03% 0 [ 4ef: 535] div [.] main
18.82% 0 [ 3ac40: 3ac4d] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac40: 3ac5c] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac40: 3ac76] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac40: 3ac88] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac90: 3ac9c] libc-2.23.so [.] __random_r
16.29% -8 [ 3aac0: 3aac0] libc-2.23.so [.] __random
16.29% 0 [ 3aac0: 3aad2] libc-2.23.so [.] __random
16.29% 0 [ 3aae0: 3aae7] libc-2.23.so [.] __random
16.29% 0 [ 3ab03: 3ab0f] libc-2.23.so [.] __random
16.29% 0 [ 3ab14: 3ab1b] libc-2.23.so [.] __random
16.29% 0 [ 3ab28: 3ab2e] libc-2.23.so [.] __random
16.29% 0 [ 3ab4a: 3ab53] libc-2.23.so [.] __random
8.11% 0 [ 640: 644] div [.] compute_flag
8.11% 0 [ 649: 659] div [.] compute_flag
5.46% 0 [ 3af60: 3af60] libc-2.23.so [.] rand
5.46% 0 [ 3af60: 3af64] libc-2.23.so [.] rand
2.25% 0 [ 490: 490] div [.] rand@plt
0.01% 26 [ c00a27: c00a27] [kernel.vmlinux] [k] native_irq_return_iret
0.00% -157 [ 2bf9f2: 2bfa63] [kernel.vmlinux] [k] update_blocked_averages
0.00% -56 [ 2bf980: 2bf9d3] [kernel.vmlinux] [k] update_blocked_averages
0.00% 48 [ 2bf934: 2bf942] [kernel.vmlinux] [k] update_blocked_averages
0.00% 3 [ 2bfb38: 2bfb67] [kernel.vmlinux] [k] update_blocked_averages

The 'cycles' is a new perf-diff computation selection, which enables
the displaying of cycles difference of same program basic block amongst
two perf.data. The program basic block is the code block between two
branches in a function.

v3:
---
In v3, the major change is to move most of block stuffs from
'struct hist_entry' to new structure 'struct block_hist' and
update the code accordingly. But we still have to keep the
block_info in 'struct hist_entry' since we need to compare by
block info when inserting new entry to hists.

Others are minor changes, such as abs() -> labs(), removing
duplicated ops and etc.

Changed patches:
perf diff: Use hists to manage basic blocks per symbol
perf diff: Link same basic blocks among different data
perf diff: Print the basic block cycles diff

v2:
---
Keep standard perf diff format.

Following is the v1 output.

# perf diff --basic-block

# Cycles diff Basic block (start:end)
# ........... .......................
#
-208 hrtimer_interrupt (30b9e0:30ba42)
-157 update_blocked_averages (2bf9f2:2bfa63)
-126 interrupt_entry (c00880:c0093a)
-86 hrtimer_interrupt (30bb29:30bb32)
-74 hrtimer_interrupt (30ba65:30bac4)
-56 update_blocked_averages (2bf980:2bf9d3)
48 update_blocked_averages (2bf934:2bf942)
-35 native_write_msr (267900:26790b)
26 native_irq_return_iret (c00a27:c00a27)
22 rcu_check_callbacks (2febb6:2febdc)
-21 __hrtimer_run_queues (30b220:30b2a3)
19 pvclock_gtod_notify (14ba0:14c1b)
-18 task_tick_fair (2c5d29:2c5d41)

Jin Yao (7):
perf util: Create block_info structure
perf util: Add block_info in hist_entry
perf diff: Check if all data files with branch stacks
perf diff: Use hists to manage basic blocks per symbol
perf diff: Link same basic blocks among different data
perf diff: Print the basic block cycles diff
perf diff: Documentation -c cycles option

tools/perf/Documentation/perf-diff.txt | 14 +-
tools/perf/builtin-diff.c | 383 ++++++++++++++++++++++++++++++++-
tools/perf/ui/stdio/hist.c | 27 +++
tools/perf/util/hist.c | 40 +++-
tools/perf/util/hist.h | 9 +
tools/perf/util/sort.h | 13 ++
tools/perf/util/symbol.c | 22 ++
tools/perf/util/symbol.h | 23 ++
tools/perf/util/symbol_conf.h | 1 +
9 files changed, 522 insertions(+), 10 deletions(-)

--
2.7.4