[BOLT] [3.12] Python 3.12.7 --enable bolt option not working

[ptr1337]

Bug report

Bug description:

Hi together,

After updating llvm-bolt to 19.1.0 it is not possible to use the --enable-bolt function anymore.
Following can be found in the log:

BOLT-INFO: 0 out of 5950 functions in the binary (0.0%) have non-empty execution profile
BOLT-INSTRUMENTER: Number of indirect call site descriptors: 7119
BOLT-INSTRUMENTER: Number of indirect call target descriptors: 5874
BOLT-INSTRUMENTER: Number of function descriptors: 5874
BOLT-INSTRUMENTER: Number of branch counters: 78400
BOLT-INSTRUMENTER: Number of ST leaf node counters: 40936
BOLT-INSTRUMENTER: Number of direct call counters: 0
BOLT-INSTRUMENTER: Total number of counters: 119336
BOLT-INSTRUMENTER: Total size of counters: 954688 bytes (static alloc memory)
BOLT-INSTRUMENTER: Total size of string table emitted: 133862 bytes in file
BOLT-INSTRUMENTER: Total size of descriptors: 7642344 bytes in file
BOLT-INSTRUMENTER: Profile will be saved to file /tmp/pkg/src/Python-3.12.7/libpython3.12.so.1.0.bolt
BOLT-INFO: 65850 instructions were shortened
BOLT-INFO: removed 134 empty blocks
BOLT-INFO: UCE removed 1155 blocks and 71042 bytes of code
BOLT-INFO: padding code to 0xe00000 to accommodate hot text
BOLT-INFO: output linked against instrumentation runtime library, lib entry point is 0xf90930
BOLT-INFO: clear procedure is 0xf8c380
BOLT-INFO: patched build-id (flipped last bit)
BOLT-INFO: setting __bolt_runtime_start to 0xf908e0
BOLT-INFO: setting __bolt_runtime_fini to 0xf90930
BOLT-INFO: setting __hot_start to 0x800000
BOLT-INFO: setting __hot_end to 0xdf7438
BOLT-ERROR: unable to get new address corresponding to input address 0x1a6e11 in function sre_ucs1_match/1. Consider adding this function to --skip-funcs=...
make[1]: *** [Makefile:856: profile-bolt-stamp] Error 1
make[1]: Leaving directory '/tmp/pkg/src/Python-3.12.7'
make: *** [Makefile:885: bolt-opt] Error 2
==> ERROR: A failure occurred in build().

so, it appears to fail at the instrumentation and suggest to add --skip-funcs=sre_ucs1_match/*.

CPython versions tested on:

3.12

Operating systems tested on:

Linux

[Eclips4]

cc @corona10

[ZIZUN]

I will take a look

[ZIZUN]

 leesm@leesm-ubuntu  ~/Workspace/cpython   ./configure --enable-bolt
 leesm@leesm-ubuntu  ~/Workspace/cpython   make
->
BOLT-INFO: basic block reordering modified layout of 2699 functions (68.73% of profiled, 45.13% of total)
BOLT-INFO: UCE removed 36 blocks and 0 bytes of code
BOLT-INFO: splitting separates 1220980 hot bytes from 780031 cold bytes (61.02% of split functions is hot).
BOLT-INFO: 43 Functions were reordered by LoopInversionPass
BOLT-INFO: program-wide dynostats after all optimizations before SCTC and FOP:

         18740817903 : executed forward branches
          7826248652 : taken forward branches
          7349242935 : executed backward branches
          2852718872 : taken backward branches
          4004518103 : executed unconditional branches
          3353239032 : all function calls
          1262678534 : indirect calls
           137784052 : PLT calls
        259717127425 : executed instructions
         77083459173 : executed load instructions
         43910218984 : executed store instructions
          4541628185 : taken jump table branches
                   0 : taken unknown indirect branches
         30094578941 : total branches
         14683485627 : taken branches
         15411093314 : non-taken conditional branches
         10678967524 : taken conditional branches
         26090060838 : all conditional branches

         21694777552 : executed forward branches (+15.8%)
          1679657684 : taken forward branches (-78.5%)
          4863157010 : executed backward branches (-33.8%)
          2286864783 : taken backward branches (-19.8%)
          1866969944 : executed unconditional branches (-53.4%)
          2800009758 : all function calls (-16.5%)
           673537882 : indirect calls (-46.7%)
           137784052 : PLT calls (=)
        256484195370 : executed instructions (-1.2%)
         76738590906 : executed load instructions (-0.4%)
         43910090648 : executed store instructions (-0.0%)
          4541628185 : taken jump table branches (=)
                   0 : taken unknown indirect branches (=)
         28424904506 : total branches (-5.5%)
          5833492411 : taken branches (-60.3%)
         22591412095 : non-taken conditional branches (+46.6%)
          3966522467 : taken conditional branches (-62.9%)
         26557934562 : all conditional branches (+1.8%)

BOLT-INFO: SCTC: patched 60 tail calls (57 forward) tail calls (3 backward) from a total of 60 while removing 3 double jumps and removing 50 basic blocks totalling 250 bytes of code. CTCs total execution count is 36614663 and the number of times CTCs are taken is 25417844
BOLT-INFO: FOP optimized 1 redundant load(s) and 0 unused store(s)
BOLT-INFO: Frequency of redundant loads is 15661399 and frequency of unused stores is 0
BOLT-INFO: Frequency of loads changed to use a register is 15661399 and frequency of loads changed to use an immediate is 0
BOLT-INFO: FOP deleted 1 load(s) (dyn count: 15661399) and 0 store(s)
BOLT-INFO: FRAME ANALYSIS: 2248 function(s) were not optimized.
BOLT-INFO: FRAME ANALYSIS: 1999 function(s) (88.8% dyn cov) could not have its frame indices restored.
BOLT-INFO: Shrink wrapping moved 14 spills inserting load/stores and 0 spills inserting push/pops
BOLT-INFO: Shrink wrapping reduced 117436164 store executions (0.0% total instructions executed, 0.3% store instructions)
BOLT-INFO: Shrink wrapping failed at reducing 0 store executions (0.0% total instructions executed, 0.0% store instructions)
BOLT-INFO: Allocation combiner: 22 empty spaces coalesced (dyn count: 84363257).
BOLT-INFO: patched build-id (flipped last bit)
BOLT-INFO: setting _end to 0xcf093c
BOLT-INFO: setting _end to 0xcf093c
BOLT-INFO: setting __hot_start to 0xa00000
BOLT-INFO: setting __hot_end to 0xb7d512
touch profile-bolt-stamp
make[1]: Leaving directory '/home/leesm/Workspace/cpython'

 leesm@leesm-ubuntu  ~/Workspace/cpython  ./python
Python 3.12.7 (main, Nov  2 2024, 17:57:45) [GCC 9.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> exit()
 leesm@leesm-ubuntu  ~/Workspace/cpython  llvm-bolt --version
LLVM (http://llvm.org/):
  LLVM version 19.1.0
  Optimized build with assertions.
BOLT revision <unknown>

  Registered Targets:
    aarch64    - AArch64 (little endian)
    aarch64_32 - AArch64 (little endian ILP32)
    aarch64_be - AArch64 (big endian)
    arm64      - ARM64 (little endian)
    arm64_32   - ARM64 (little endian ILP32)
    x86        - 32-bit X86: Pentium-Pro and above
    x86-64     - 64-bit X86: EM64T and AMD64

I tested this issue on my setup with CPython 3.12 and LLVM-BOLT version 19.1.0 on a Ubuntu(20.04) environment, and I did not encounter any problems. The --enable-bolt function worked as expected. It appears that this issue might be other environment problems. @ptr1337

[ptr1337]

After Python 3.13 got pushed to archlinux stable, im still not able to bolt it with the same issues.
Since the llvm-bolt package is not in the archlinux repository, you can test it with following docker container:

git clone https://gitlab.archlinux.org/archlinux/packaging/packages/python.git
cd python
# enable --enable-bolt option
docker run --name dockerbuilder -e EXPORT_PKG=1 -e SYNC_DATABASE=1 -v $PWD:/pkg cachyos/docker-makepkg && docker rm dockerbuilder

[ms178]

I can confirm this issue using Clang-20git (c660b281b60085cbe40d73d692badd43d7708d20) on CachyOS with Python 3.13.1:

BOLT-INFO: Target architecture: x86_64
BOLT-INFO: BOLT version: c660b281b60085cbe40d73d692badd43d7708d20
BOLT-INFO: first alloc address is 0x200000
BOLT-INFO: creating new program header table at address 0xa00000, offset 0x800000
BOLT-WARNING: debug info will be stripped from the binary. Use -update-debug-sections to keep it.
BOLT-INFO: enabling relocation mode
BOLT-INFO: forcing -jump-tables=move for instrumentation
BOLT-INFO: enabling lite mode
BOLT-INFO: 0 out of 6 functions in the binary (0.0%) have non-empty execution profile
BOLT-INSTRUMENTER: Number of indirect call site descriptors: 2
BOLT-INSTRUMENTER: Number of indirect call target descriptors: 4
BOLT-INSTRUMENTER: Number of function descriptors: 4
BOLT-INSTRUMENTER: Number of branch counters: 1
BOLT-INSTRUMENTER: Number of ST leaf node counters: 4
BOLT-INSTRUMENTER: Number of direct call counters: 0
BOLT-INSTRUMENTER: Total number of counters: 5
BOLT-INSTRUMENTER: Total size of counters: 40 bytes (static alloc memory)
BOLT-INSTRUMENTER: Total size of string table emitted: 47 bytes in file
BOLT-INSTRUMENTER: Total size of descriptors: 356 bytes in file
BOLT-INSTRUMENTER: Profile will be saved to file /tmp/makepkg/python/src/Python-3.13.1/python.bolt
BOLT-INFO: padding code to 0xe00000 to accommodate hot text
BOLT-INFO: output linked against instrumentation runtime library, lib entry point is 0xe06890
BOLT-INFO: clear procedure is 0xe02360
BOLT-INFO: setting __bolt_runtime_start to 0xe06850
BOLT-INFO: setting __bolt_runtime_fini to 0xe06890
BOLT-INFO: setting __hot_start to 0xc00000
BOLT-INFO: setting __hot_end to 0xc00134
BOLT-INFO: shared object or position-independent executable detected
BOLT-INFO: Target architecture: x86_64
BOLT-INFO: BOLT version: c660b281b60085cbe40d73d692badd43d7708d20
BOLT-INFO: first alloc address is 0x0
BOLT-INFO: creating new program header table at address 0xc00000, offset 0xc00000
BOLT-WARNING: debug info will be stripped from the binary. Use -update-debug-sections to keep it.
BOLT-INFO: enabling relocation mode
BOLT-INFO: forcing -jump-tables=move for instrumentation
BOLT-INFO: enabling lite mode
BOLT-WARNING: Failed to analyze 1081 relocations
BOLT-WARNING: 26 collisions detected while hashing binary objects. Use -v=1 to see the list.
BOLT-INFO: 0 out of 7729 functions in the binary (0.0%) have non-empty execution profile
BOLT-INSTRUMENTER: Number of indirect call site descriptors: 1611
BOLT-INSTRUMENTER: Number of indirect call target descriptors: 7663
BOLT-INSTRUMENTER: Number of function descriptors: 7663
BOLT-INSTRUMENTER: Number of branch counters: 87823
BOLT-INSTRUMENTER: Number of ST leaf node counters: 43203
BOLT-INSTRUMENTER: Number of direct call counters: 0
BOLT-INSTRUMENTER: Total number of counters: 131026
BOLT-INSTRUMENTER: Total size of counters: 1048208 bytes (static alloc memory)
BOLT-INSTRUMENTER: Total size of string table emitted: 166433 bytes in file
BOLT-INSTRUMENTER: Total size of descriptors: 8686344 bytes in file
BOLT-INSTRUMENTER: Profile will be saved to file /tmp/makepkg/python/src/Python-3.13.1/libpython3.13.so.1.0.bolt
BOLT-INFO: 67316 instructions were shortened
BOLT-INFO: removed 59 empty blocks
BOLT-INFO: UCE removed 491 blocks and 30006 bytes of code
BOLT-INFO: padding code to 0x1600000 to accommodate hot text
BOLT-INFO: output linked against instrumentation runtime library, lib entry point is 0x17d38a0
BOLT-INFO: clear procedure is 0x17cf370
BOLT-INFO: setting __bolt_runtime_start to 0x17d3860
BOLT-INFO: setting __bolt_runtime_fini to 0x17d38a0
BOLT-INFO: setting __hot_start to 0xe00000
BOLT-INFO: setting __hot_end to 0x148ed34
BOLT-ERROR: unable to get new address corresponding to input address 0x476f7a in function _PyEval_EvalFrameDefault. Consider adding this function to --skip-funcs=...

And if I do as advised by BOLT, I see another function that errors out instead:

BOLT-ERROR: unable to get new address corresponding to input address 0x4fa143 in function sre_ucs1_match/1(*2). Consider adding this function to --skip-funcs=...

[ptr1337]

Also, another idea:
Instead of relaying on instrumentation (which works on most enviroments) we could also add an option.
@lseman provided an option a while ago at following patch: https://termbin.com/zksh

[ptr1337]

Adding to the bolt commands:
--skip-funcs=sre_ucs1_match/1,_PyEval_EvalFrameDefault.localalias/1
appears to fix it.

[indygreg]

The unable to get new address corresponding to input address error was added in LLVM 19.1 by llvm/llvm-project#89681. It can trigger when using computed gotos in PIC compiled code. Not sure about the original report, but _PyEval_EvalFrameDefault uses computed gotos, making it BOLT incompatible with LLVM 19.1. (And behavior on older LLVM versions may be incorrect, leading to buggy behavior.)

I think a proper fix here is to add any functions with computed gotos to the BOLT exclusion list. Presumably a future LLVM release will gain the ability to perform these dynamic relocations, so we'll [eventually] want some form of LLVM version sniffing to control the behavior.

[ms178]

@indygreg This might be the relevant LLVM-MR for gaining that functionality: llvm/llvm-project#120267