[llvm-dev] PGO GC inefficiency due to link.exe's IMAGE_COMDAT_SELECT_ASSOCIATIVE limitation

[Fāng-ruì Sòng via llvm-dev]

I have studied the GC state of Windows (IR/frontend) PGO a bit.
Below is my understanding. Hope folks can clarify it I make a mistake:)

For PGO, we need cnts/data/vals/names sections. cnts/data are main
metadata sections. vals is for value profiling.
For Windows, the cnts section is named .lprfc$M and the data section
is named .lprfd$M.
The garbage collection story is unfortunate.

If an IMAGE_COMDAT_SELECT_ASSOCIATIVE section defines an external
symbol, MSVC link.exe may report a spurious duplicate symbol error
(error LNK2005), even if the associative section would be discarded
after handling the leader symbol.
lld-link doesn't have this limitation.
However, a portable implementation needs to work around MSVC link.exe.

For a COMDAT .lprfd$M, its symbol must be external (linkonce_odr),
otherwise references to a non-prevailing symbol would cause an error.
Due to the limitation, .lprfd$M has to reside in its own COMDAT, no
sharing with .lprfc$M.

Different COMDAT groups mean that the liveness of one .lprfc$M does
not make its associative .lprfd$M live.
Since a .lprfd$M may be unreferenced, we have to conservatively assume
all COMDAT .lprfd$M live.
Since .lprfc$M input sections parallel .lprfd$M input sections, we
have to conservatively assume all COMDAT .lprfc$M live.
For an external symbol, we use a /INCLUDE: directive in .drectve to
mark it as a GC root.
As a result, .drectve may have many /INCLUDE: directives, just to work
around the link.exe limitation.
IIRC Chrome folks have found that the /INCLUDE: directives can cause
huge memory usage.

Note: for ELF we can use R_*_NONE to establish an artificial
dependency edge between two sections.
I don't think PE-COFF provides a similar feature.

So does link.exe accept feature requests? The external symbol
limitation in an IMAGE_COMDAT_SELECT_ASSOCIATIVE section looks quite
unfortunate...

It seems that /OPT:REF has another inefficiency - it cannot discard
non-COMDAT sections.
IIUC Fuchsia folks found that GC on non-COMDAT functions can save a
lot of space (40%? of PGO metadata sections). It is unfortunate that
link.exe doesn't discard non-COMDAT sections...



-- 
宋方睿