[clang] [llvm] [BPF] Make -mcpu=v3 as the default (PR #107008)

[Yuval Deutscher via llvm-commits]

yuvald-sweet-security wrote:

@yonghong-song thank you for taking the time to help with this issue and providing your suggestions, it’s greatly appreciated. I am also glad to hear that asm barriers are no longer necessary, as they caused quite some trouble for me in the past. However, I've encountered a few issues while attempting to apply this as a solution to the problem with moving to mpcu v3.

* while you are correct in that removing the barriers fixes the verifier failure in `trace_ret_vfs_writev_tail`, the original issue I've pointed out - the increased size of kernel verifier states and instructions when using v3 - still remains:

```
$ sudo /root/veristat /tmp/test-v1.o --filter=trace_ret_vfs_writev_tail
Processing 'test-v1.o'...
File       Program                    Verdict  Duration (us)  Insns  States  Program size  Jited size
---------  -------------------------  -------  -------------  -----  ------  ------------  ----------
test-v1.o  trace_ret_vfs_writev_tail  success          23158  25292    1874          7163       38486
---------  -------------------------  -------  -------------  -----  ------  ------------  ----------
Done. Processed 1 files, 0 programs. Skipped 1 files, 158 programs.

$ sudo /root/veristat /tmp/test-v3-nobarrier.o --filter=trace_ret_vfs_writev_tail
Processing 'test-v3-nobarrier.o'...
File                 Program                    Verdict  Duration (us)   Insns  States  Program size  Jited size
-------------------  -------------------------  -------  -------------  ------  ------  ------------  ----------
test-v3-nobarrier.o  trace_ret_vfs_writev_tail  success          69161  104971    7618          6999       37490
-------------------  -------------------------  -------  -------------  ------  ------  ------------  ----------
Done. Processed 1 files, 0 programs. Skipped 1 files, 158 programs.
```
As you can see the v3 codegen causes the verifier to consume about 4 times the number of instructions when verifying it, and while it's not such a big issue for this particular function, it can be an issue for larger functions which are already close to the verifier's 1 million instructions limit as this can cause them to go over it.

* Many verifier failures still remain after removing the barriers. e.g. the function `vfs_writev_magic_return` still fails even after I've removed all asm volatile/barrier vars (you can use [patch.txt](https://github.com/user-attachments/files/19407900/patch.txt) which removes all of them). Also, in some of the code that I have removing the barrier vars on v3 codegen actually introduces more verifier failures (compared to v3 with barrier vars), I'll see if I can make a minimal example of this.

thank you again for your thoughtful advice and support

https://github.com/llvm/llvm-project/pull/107008