Make everything 4% faster by skipping empty tasks [NFC]

[kripken]

When a visitor is the original

void visitFoo(Foo* curr) {}`

(that is, empty), and the doVisit is also unchanged,

static void doVisitFoo(Self* self, Foo* curr) { self->visitFoo(curr); }

(that is, it just calls the visitor), then we do not need to queue such
tasks for execution at all.

I had wanted to do this a while ago but just now figured out how to
make the double check work.

Measurements on -O3 on a large Kotlin testcase:

Before:

   166,045,403,096      cpu_core/instructions/u          #    1.42  insn per cycle              ( +-  0.01% )
    36,389,460,009      cpu_core/branches/u              #  512.421 M/sec                       ( +-  0.01% )
           26.0642 +- 0.0926 seconds time elapsed  ( +-  0.36% )

After:

   158,567,768,086      cpu_core/instructions/u          #    1.44  insn per cycle              ( +-  0.00% )
    34,644,951,424      cpu_core/branches/u              #  521.618 M/sec                       ( +-  0.01% )
            24.744 +- 0.163 seconds time elapsed  ( +-  0.66% )

Instructions and branches decrease by 5% with almost no noise. Wall time also
shrinks by about 5%, but the noise is around 1% there.

[tlively]

Nice!

[kripken]

Unfortunately gcc-11 errors on those function pointers not being constexpr. I tried to find a refactoring workaround among the users of the code, but couldn't. Instead I pushed a hack to avoid constexpr in gcc 11 and earlier.

Numbers from another machine, where I could test multiple compilers:

Before:

gcc-11
            9.3755 +- 0.0445 seconds time elapsed  ( +-  0.47% )
gcc-15
            8.7343 +- 0.0397 seconds time elapsed  ( +-  0.45% )
clang-19
            8.8103 +- 0.0211 seconds time elapsed  ( +-  0.24% )

After, with the opt:

gcc-11 (no constexpr)
            9.1343 +- 0.0263 seconds time elapsed  ( +-  0.29% )
gcc-15 (constexpr)
            8.5140 +- 0.0369 seconds time elapsed  ( +-  0.43% )
clang-19 (constexpr)
            8.5180 +- 0.0137 seconds time elapsed  ( +-  0.16% )

So with and without constexpr, there is a 2.5% speedup with this patch on gcc, and 3.5% with clang.

[kripken]

Testing some C++, Dart, and Java, the average speedup is around 4%.