[llvm-dev] [RFC] Interprocedural MIR-level outlining pass

[Alex Bradbury via llvm-dev]

On 26 August 2016 at 22:26, Jessica Paquette via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> ================================
> Approach and Initial Results
> ================================
> To do this, we chose to implement an outliner. Outliners find sequences of
> instructions which would be better off as a function call, by some measure
> of "better". In this case, the measure of "better" is "makes code
> smaller".
>
>
> ================================
> Results
> ================================
> These results are from a fairly recent 64-bit Intel processor, using a
> version of Clang equipped with the outliner prototype versus an equivalent
> version of Clang without the outliner.
>
> CODE SIZE REDUCTION
> For tests >=4 Kb in non-outlined size, the outliner currently provides an
> average of 12.94% code size reduction on the LLVM test suite in comparison
> to a default Clang, up to 51.4% code size reduction. In comparison to a
> Clang with -Oz, the outliner provides an average of a 1.29% code size
> reduction, up to a 37% code size reduction. I believe that the -Oz numbers
> can be further improved by better tuning the outlining cost model.

Hi Jessica, as I said on Twitter many thanks for sharing such a
detailed write-up. One thing I was wondering is whether you'd looked
at your outlining scheme with reference to function prologues and
epilogues. There has been a patch proposed to use shared epilogues in
compiler-rt for AArch64 in order to reduce code size
(https://reviews.llvm.org/D15600) and work on the RISC-V compressed
instruction set has shown that outlining for prologues/epilogues can
reduce code size by ~5% and weaken the argument for supporting load
multiple and store multiple instructions
(https://riscv.org/wp-content/uploads/2015/06/riscv-compressed-workshop-june2015.pdf,
slide 15).

Having the compiler automatically detect such opportunities seems
attractive. Do you think that these this subset of outlining would be
better addressed by a specifically targeted pass, or within your
MachineOutliner?

Best,

Alex