Re: [RFC PATCH] pre-process: add __VA_OPT__ support

[Al Viro <viro@zeniv.linux.org.uk>]

On Wed, Mar 18, 2026 at 09:07:17PM -0700, Linus Torvalds wrote:

> I looked at clang at one point, and iirc it generated *much* better
> code, because I think it does tth esmart thing, which is to get rid of
> the notion of bitfields as quickly as possible, and then doing just
> regular integer optimizations.

FWIW, with the local tokenizer patches the single worst spot at the moment
(both for clang and for gcc build) is this:
        p = &hash_table[hash];
	while ((ident = *p) != NULL) {
		if (ident->len == (unsigned char) len) {
in create_hashed_ident().  This is from clang build, where
it's inlined into tokenize_stream():
   0.14 │       mov    (%rdx,%rcx,8),%rax                                                                                           
  12.34 │       mov    %r12d,%r15d                                                                                                  
   0.02 │       test   %rax,%rax                                                                                                    
   0.43 │     ↓ je     33e                                                                                                          
        │       mov    %rsp,%r14                                                                                                    
   0.29 │     ↓ jmp    319                                                                                                          
        │       nop                                                                                                                 
        │310:┌─→mov    0x0(%r13),%rax                                                                                               
   2.00 │    │  test   %rax,%rax                                                                                                    
   0.07 │    │↓ je     345                                                                                                          
   0.00 │319:│  mov    %rax,%r13                                                                                                    
        │    │if (ident->len == (unsigned char) len) {                                                                              
   0.19 │    ├──cmp    %r12b,0x10(%rax)                                                                                             
  16.86 │    └──jne    310                                                                                                          
and this is gcc build, where it's not inlined, so the percentages are
several times higher (out of 5.9% vs. out of 17.1% on the profiles I'm
looking at):
        │    p = &hash_table[hash];                                                                                                ▒
   0.77 │      mov    (%rax,%rdx,8),%rbx                                                                                           ▒
        │    while ((ident = *p) != NULL) {                                                                                        ▒
  27.77 │      test   %rbx,%rbx                                                                                                    ◆
   0.98 │    ↓ jne    3b                                                                                                           ▒
        │    ↓ jmp    c2                                                                                                           ▒
        │      nop                                                                                                                 ▒
        │    ident_hit++;                                                                                                          ▒
        │    return ident;                                                                                                         ▒
        │    }                                                                                                                     ▒
        │    next:                                                                                                                 ▒
        │    //misses++;                                                                                                           ▒
        │    p = &ident->next;                                                                                                     ▒
   0.00 │30:┌─→mov    (%rbx),%rax                                                                                                  ▒
        │   │while ((ident = *p) != NULL) {                                                                                        ▒
   6.11 │   │  test   %rax,%rax                                                                                                    ▒
   0.24 │   │↓ je     70                                                                                                           ▒
        │   │  mov    %rax,%rbx                                                                                                    ▒
        │   │if (ident->len == (unsigned char) len) {                                                                              ▒
   0.81 │3b:├──cmp    %bpl,0x10(%rbx)                                                                                              ▒
  50.82 │   └──jne    30                                                                                                           ▒

Most of the accesses are to single-element chain; it's not walking the
lists that hurts, it's the very first step.  The profiles are for userland
cycles; looking for stalled-cycles-frontend gives exact same hotspots.
The next one is lookup_symbol(); there we also walk linked lists.
The only difference is that lists are often longer than one entry...
Not sure what can be done about either.