[lld] [LLD][ELF] Change default flags for NOBITS sections with no inputs (PR #70447)

[Fangrui Song via llvm-commits]

https://github.com/MaskRay requested changes to this pull request.

Thank you for posting https://discourse.llvm.org/t/lld-output-section-flag-assignment-behavior/74359 but I don't think the change is desirable.

For `custom-section-type.s`, GNU ld's output does not set the `SHF_WRITE` flag for `nobits`.
```
Section Headers:
  [Nr] Name              Type            Address          Off    Size   ES Flg Lk Inf Al
  [ 0] trtab             GROUP           000000000000001e 00101e 000004 04   A  9   0  1
  [ 1] .text             PROGBITS        0000000000000000 001000 000002 00  AX  0   0  4
  [ 2] progbits          PROGBITS        0000000000000002 001002 000001 00   A  0   0  1
  [ 3] note              NOTE            0000000000000003 001003 000002 00   A  0   0  1
  [ 4] nobits            NOBITS          0000000000000005 001005 000001 00   A  0   0  1
...
```

We can change `nobits ( TYPE=SHT_NOBITS) : { BYTE(8) }` to reuse `(NOLOAD)` and still get `nobits` without the `SHF_WRITE` flag.

GNU ld has changed to not force `SHF_WRITE` since 2021: https://sourceware.org/bugzilla/show_bug.cgi?id=26378

Regarding
```
HEAP_SIZE = 1K;

SECTIONS {  }
SECTIONS
{
  .CPU0.heap (NOLOAD) : ALIGN(64) { . = ALIGN(64); . += HEAP_SIZE; }
  .CPU1.heap (NOLOAD) : ALIGN(64) { . = ALIGN(64); . += HEAP_SIZE; }
  .CPU2.heap (NOLOAD) : ALIGN(64) { . = ALIGN(64); . += HEAP_SIZE; }
  .code : {  *(.text*) }
}
```

I think GNU sets the `SHF_WRITE` for `. = ALIGN(64);` and `SHF_WRITE|SHF_ALLOC` for `. += HEAP_SIZE;`.
The behavior is inconsistent with data commands.
I believe LLD should not change anything regarding the current consistent behavior.
You probably need to adjust your linker script as Peter suggested.


https://github.com/llvm/llvm-project/pull/70447