llvm dev - Sep 2016 - LLVM v3.9.0 and '.p2align'

While investigating a new bug in some code using our LLVM v3.9.0 derived
compiler, it turns out that it now emits '.p2align' unconditionally if
the
alignment value is a power of 2.  Previously this was dependent on
'MAI->getAlignmentIsInBytes()'.

 

We don't support non-power-of 2 alignment anyway, but our assembler does not
support '.p2align' and there appears to be no way of changing this in a
target configurable way.

 

There have been emails about the fact that the 'MCAsmStreamer'
implementation is hidden, and the argument is that if the target needs
something different it can implement its own specialisation of
'MCStreamer'.
While this is technically true, but 'MCAsmStreamer' already does more
than
90% of what I need, and to implement my own specialisation really means
cloning 'MCAsmStreamer', and editing those few places where my assembly
requirements are different to it.  And then the target that does this clone
is required to carefully track future revisions of 'MCAsmStreamer' to
ensure
that their clone is still a clone.

 

Generally it is good practice to reuse existing code, and specialise only
for the differences, so it really make a lot more sense to publish it in a
header, and allow the target to sub-class it with overrides for each of the
methods it wants to handle differently.

 

As it is, I have had to add yet another tweak to this code, all of which are
relatively trivial:

 

*         'MCAsmStreamer::EmitAssignment' changed to implement aliases
as
'.alias symA symB' instead of 'symA - symB'

*         'MCAsmStreamer::EmitBytes' to break long '.byte'
sequences into
multiple '.byte' sequences on multiple lines (actually, I could retire
this
change as our assembler now handles the long sequences)

*         'MCAsmStreamer::EmitFill' because we need an additional
argument
in the directive

*         'MCAsmStreamer::EmitValueToAlignment' to use '.align'
instead of
'.p2align'

 

These are such small changes that it really does not warrant me cloning
1,600 lines of code so that I can tweak 8 of them, and being able to
sub-class the otherwise very useful 'MCAsmStreamer' class would greatly
simplify adapting LLVM for non-gas compliant assemblers.  It seems to me
that splitting this file into a source and a header is not going to create
maintenance problems for LLVM, while at the same time it would greatly
simplify the implementation of compilers for non-mainstream assemblers.

 

Thanks,

 

            MartinO

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160923/12168ee2/attachment.html>