[COFF] Limit PR39667 ICF partitioning changes to ARM64

[dmajor]

Summary

While the changes in rL347429 brought dramatic improvements to Firefox link time on ARM64, they also caused large regressions in link time on x86_64 PGO builds, severe enough that we had to back out our clang roll. I assume the slowdown has something to do with the sections created for profiling instrumentation. (It would be interesting to see what effect that patch had on ARM64 PGO builds, but we don't have such a thing yet.)

As a starting point, here's a patch to restrict rL347429 to ARM64 only. (And, if accepted, I'd suggest merging it to to 8.0.0.) Being unfamiliar with this code, I can't tell if this is a proper solution or a hack, so I'll rely on reviewers to let me know.

I don't know what to do about the corresponding ELF changes in D54773/rLLD347594. I'm not even sure if the regression occurs on ELF platforms.

[pcc]

D56955 should fix the underlying issue here.

[dmajor]

Even better -- thanks!

COFF/ICF.cpp

for (SectionChunk *SC : P.second->Sections)
       SC->Class[0] = NextId++;
 
   // Initially, we use hash values to partition sections.
   parallelForEach(Chunks, [&](SectionChunk *SC) {
     SC->Class[1] = xxHash64(SC->getContents());
   });
 
   // Combine the hashes of the sections referenced by each section into its
-  // hash.
+  // hash. This is only needed for ARM64, where the .pdata format makes those
+  // sections less likely to be unique.
   parallelForEach(Chunks, [&](SectionChunk *SC) {
     uint32_t Hash = SC->Class[1];
+    if (Config->Machine == ARM64)
       for (Symbol *B : SC->symbols())
         if (auto *Sym = dyn_cast_or_null<DefinedRegular>(B))
           Hash ^= Sym->getChunk()->Class[1];
     // Set MSB to 1 to avoid collisions with non-hash classs.
     SC->Class[0] = Hash | (1U << 31);
   });
 
   // From now on, sections in Chunks are ordered so that sections in
   // the same group are consecutive in the vector.
   std::stable_sort(Chunks.begin(), Chunks.end(),
                    [](SectionChunk *A, SectionChunk *B) {