Hello LLVM,
For some processors, the all or nothing behavior of the .align
directive is not optimal for improving fetch efficiency of branches.
For example, take an instruction fetch width of 16 bytes and a branch
target just 1 byte offset into a fetch block. Align injects 15 bytes
of padding rather than leaving well enough alone. Efficiency is
unlikely to improve for lack of 1 byte out of 16, and performance may
be hurt by all the padding not to mention code size.
Is there an existing way for a target to express a "near enough"
semantic for code alignment?
Regards,
-steve
Is there an existing way for a target to express a "near enough"
semantic for code alignment?
After more investigation:
The .align directive supports an optional MaxFillBytes parameter,
which would nicely prevent the over-padding problem. However,
MachineBlockPlacement cannot use this feature when aligning loops
since targets support getPrefLoopAlignment(), but not
getPrefLoopAlignmentMaxFill().
Does anyone dislike the idea of a new target lowering query
getPrefLoopAlignmentMaxFill()? This new query would allow
MachineBlockPlacement to specify the align directive's MaxFillBytes
parameter when aligning loops. Targets could then control
over-padding as they see fit. If there are no objections, I will
submit a patch.