Use clang-20

[thevar1able]

Resolves: #78765

Changelog category (leave one):

• Performance Improvement

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

Use clang from llvm-20

[clickhouse-gh]

Workflow [PR], commit [227c67f]

[azat]

Build (amd_tidy) — Failed: Build ClickHouse

$ grep 'error:' build_clickhouse.log | grep -o '\[.*\]' | sort | uniq -cd
     10 [bugprone-nondeterministic-pointer-iteration-order,-warnings-as-errors]
     16 [cert-arr39-c,-warnings-as-errors]
      3 [clang-diagnostic-error]
      3 [clang-diagnostic-uninitialized,-warnings-as-errors]
   2595 [modernize-use-integer-sign-comparison,-warnings-as-errors]
 153682 [portability-template-virtual-member-function,-warnings-as-errors]
    403 [readability-container-contains,-warnings-as-errors]

[azat]

No wonder it does not work - clang simply assume that the stack is 8MB - https://github.com/llvm/llvm-project/blob/788b50a4384985f2c221cfd8d290cabc6f59e646/clang/include/clang/Basic/Stack.h#L26

As for why it happens only for arm64 it is likely due to it has more registers

And here is an example of a stacktrace - https://pastila.nl/?0000e83f/bcc54268487047ba960a50304e756557#7k0S6lGYXncNor/j2uTBqw==

[azat]

Introduced in c978f0f7

$ git tag --contains c978f0f7
llvmorg-21-init
llvmorg-20.1.3
llvmorg-20.1.2
llvmorg-20.1.1
llvmorg-20.1.0-rc3
llvmorg-20.1.0-rc2
llvmorg-20.1.0-rc1
llvmorg-20.1.0

[azat]

This is how our recursion for Settings.cpp looks like :)

Clang AST:

P.S. Yes, it is better to show as screenshot since I have to zoom out as much as possible

UPD: in text - https://pastila.nl/?0010852e/d728efbefd772d270ccc6d00a5c8685a#M7N24U/ZDFhfF0/ZzJko1A==

[azat]

Except for comments LGMT, also worth checking perf tests (though they are unstable right now)

[azat]

column_array_filter query 7 looks related

[rschu1ze]

L. 63-64 say that LSE is part of ARM v.1, why do we add it again here?

[thevar1able]

Something failed without it, but I don't remember what exactly. Let's try removing them when other stuff is fixed, and see if it is really necessary.

[rschu1ze]

21

[thevar1able]

Yay, AVX-related regression is gone. But there are new ones, like sum() and if().

I'm going to close this PR and reopen a new one for clang-21.