Attempt to detect incorrect/improveable incorrect linker flags

[fredemmott]

Is your feature request related to a problem? Please describe.

I use CMake RelWithDebInfo builds - I don't use Release as I'm never going to distribute something if I don't have debug symbols for it :)

It seems that this disables some of the usual release mode linker flags, in particular:

• /INCREMENTAL:NO - SizeBench does warn about this
• /OPT:ICF
• /OPT:REF

Describe the solution you'd like

It seems like SizeBench could make a reasonable guess that /OPT:ICF or /OPT:REF were not passed, and suggest I enable them.

Describe alternatives you've considered

doing it with sizebench without 'magic' :)

I eventually found all of these flags through looking through SizeBench's reports; for example, the lack of /OPT:REF was obvious from this:

I'm only using one of those WinRT classes, so the majority of them should not be in my executable. After everything else, finding this and turning on /OPT:REF cut my executable from 3MB to 1.5MB

Additional context

Thank you so much for this incredible tool. I read the 2021 blog post last night, and this morning I easily cut my executable from 12.5MB to 1.5MB.

[dmachaj]

SizeBench uses the DIA APIs to examine the PDB files, which I believe include the full command-line for every translation unit as well as the linker command-line. Therefore I think it would be possible to detect this without having to rely on any sort of heuristic or guess.

That said, one tool that may also be of note is BinSkim. That is mainly intended to detect security-related issues with compiler and linker arguments. However, a handful of performance related issues will output warnings. See: https://github.com/microsoft/binskim/blob/main/docs/BinSkimRules.md#rule-ba6004enablecomdatfolding for one of them.

If I remember correctly the logic to reason over the command lines for cl.exe and link.exe was... nontrivial.