[Mesa-dev] [PATCH 00/21] nir: Functions!

[Jason Ekstrand]

On Sat, Feb 13, 2016 at 10:01 PM, Jordan Justen <jordan.l.justen at intel.com>
wrote:

> On 2016-02-13 18:14:17, Jason Ekstrand wrote:
> > This patch series adds actual function support to NIR.  Previously all
> NIR
> > usage relied on GLSL to have lowered functions away entirely.  As a
> result,
> > any support for functions that was there was dead code and not really
> > tested.  With SPIR-V on the horizon, NIR will need real function support.
> >
> > The series starts by deleting some dead code in glsl_to_nir and then
> > reworking the way functions and parameters are handled.  Here's the new
> > (i.e., actually well-defined) scheme:
> >
> >  - The nir_function_impl struct has a params array that contains pointers
> >    to the variables that are the functions arguments and a return_var
> array
> >    that, if not NULL points to the return variable.
> >  - Each of these variables has the "param" mode and does not exist in any
> >    variable list but instead is reachable only through the params array
> or
> >    the return_var pointer.
> >  - Each variable with the "param" mode has a location.  For parameters,
> >    that location is the index into the param array; for return values,
> the
> >    location is defined to be -1.
> >  - Parameters and return values are not printed with the list of locals
> but
> >    are, instead, only printed in the argument list and are now printed
> with
> >    their type.
> >
> > Since glsl_to_nir, prog_to_nir, and tgsi_to_nir never produce any
> functions
> > other than main(), it's safe to do these reworks in basically any order.
> >
> > The next part of the series builds up to doing function inlining.
> However,
> > function inlining requires return lowering which pass depends on being
> able
> > to repair SSA form when it's done so we need an SSA repair pass.  This,
> in
> > turn requires a phi builder and, while we're at it, we might as well
> rework
> > the into-SSA pass.  Yeah, I know, it's kind of the long way there, but
> what
> > can you do?
> >
> > Once we have an SSA repair pass, return lowering and function inlining
> come
> > fairly quickly.  However, in order to make return lowering possible, we
> > have to patch up control-flow handling so it doesn't break when we try
> > remove stuff from the top level of the function.
> >
> > Jason Ekstrand (21):
> >   nir/glsl: Remove dead function parameter handling code
> >   nir: Add a new "param" variable mode for parameters and return
> >     variables
> >   nir: Add a helper for creating a "bare" nir_function_impl
>
> Yeah, I suppose you could move the param init bits to the next patch.
> It seems a little better, but also not too important to me.
>

Done!


>
> >   nir: Create function parameters in function_impl_create
> >   nir/print: Factor variable name lookup into a helper
> >   nir/print: Better function argument printing
> >   nir/validate: Better function validation
> >   nir/clone: Add support for cloning a single function_impl
>
> 1-8 Reviewed-by: Jordan Justen <jordan.l.justen at intel.com>
>

Thanks!


>
> >   nir: Add a phi node placement helper
> >   nir/dominance: Handle unreachable blocks
> >   nir/vars_to_ssa: Use the new nir_phi_builder helper
> >   util/bitset: Allow iterating over const bitsets
> >   nir: Add a pass to repair SSA form
> >   nir/cf: Handle relinking top-level blocks
> >   nir: Add a function for comparing cursors
> >   nir/cf: Make extracting or re-inserting nothing a no-op
> >   nir/builder: Add a helper for inserting jump instructions
> >   nir: Add a cursor helper for getting a cursor after any phi nodes
> >   nir: Add return lowering pass
> >   nir/builder: Add helpers for easily inserting copy_var intrinsics
> >   nir: Add a pass to inline functions
> >
> >  src/compiler/Makefile.sources            |   5 +
> >  src/compiler/nir/Makefile.sources        |   5 +
> >  src/compiler/nir/glsl_to_nir.cpp         |  51 +---
> >  src/compiler/nir/nir.c                   | 115 +++++++-
> >  src/compiler/nir/nir.h                   |  38 ++-
> >  src/compiler/nir/nir_builder.h           |  30 ++
> >  src/compiler/nir/nir_clone.c             | 112 +++++--
> >  src/compiler/nir/nir_control_flow.c      |  16 +-
> >  src/compiler/nir/nir_dominance.c         |   6 +-
> >  src/compiler/nir/nir_inline_functions.c  | 270 +++++++++++++++++
> >  src/compiler/nir/nir_lower_returns.c     | 246 ++++++++++++++++
> >  src/compiler/nir/nir_lower_vars_to_ssa.c | 484
> +++++++++----------------------
> >  src/compiler/nir/nir_phi_builder.c       | 254 ++++++++++++++++
> >  src/compiler/nir/nir_phi_builder.h       |  84 ++++++
> >  src/compiler/nir/nir_print.c             |  76 +++--
> >  src/compiler/nir/nir_repair_ssa.c        | 157 ++++++++++
> >  src/compiler/nir/nir_validate.c          |  24 +-
> >  src/util/bitset.h                        |   2 +-
> >  18 files changed, 1490 insertions(+), 485 deletions(-)
> >  create mode 100644 src/compiler/nir/nir_inline_functions.c
> >  create mode 100644 src/compiler/nir/nir_lower_returns.c
> >  create mode 100644 src/compiler/nir/nir_phi_builder.c
> >  create mode 100644 src/compiler/nir/nir_phi_builder.h
> >  create mode 100644 src/compiler/nir/nir_repair_ssa.c
> >
> > --
> > 2.5.0.400.gff86faf
> >
> > _______________________________________________
> > mesa-dev mailing list
> > mesa-dev at lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20160214/9be303a3/attachment.html>