[libre-riscv-dev] LD/ST address matcher

[Luke Kenneth Casson Leighton]

On Monday, June 3, 2019, Samuel Falvo II <sam.falvo at gmail.com> wrote:

> Since you need to generate byte write enables for the AXI write
> transactions,


You do?  That's interesting to know.  Ok so the logic should be split into
a module, for reuse, at least. Perhaps even outputted (or inputted) to save
gates.



>
> that same logic can potentially be reused to generate
> your byte-level bitmap for use elsewhere.


Good insight, thx Samuel.


>
> On Mon, Jun 3, 2019 at 7:38 AM Luke Kenneth Casson Leighton
> <lkcl at lkcl.net> wrote:
> >
> > http://git.libre-riscv.org/?p=soc.git;a=blob;f=src/
> scoreboard/addr_match.py;h=657f95b68851b4df7a40eec031694e
> 2246c3ea49;hb=HEAD
> >
> > Am making progress understanding the LD/ST Dependency Matrix, after
> staring
> > at it for nearly four hours. More on this later.
> >
> > Another sub-task is to detect overlapping addresses. Or, at least to work
> > out if the addresses MIGHT overlap. If the algorithm is overzealous, all
> it
> > does is affect performance as the engine has to drop down near to single
> > issue.
> >
> > Jacob, what sort of LD/ST profile can we expect from Kazan? The reason
> for
> > asking is because the simplest algorithm is to check bits 4 through 15 or
> > so and ignore the bottom 4 bits and everything 16 or above.
> >
> > The result of doing that would be that a byte LD/ST would hit against a
> > DWORD 64 bit LD/ST, ie be detected as "needs to be treated as
> single-issue,
> > no parallism allowed".
> >
> > A better algorithm would be to additionally take the bottom 4 bits, and
> > expand them out into a bitmap.  A byte LD/ST would have 1 bit set, a
> > misaligned WORD LD/ST would have 4 bits set somewhere on a nonaligned
> > boundary, and so on.
> >
> > I believe it would be as simple as:
> >
> > 1/3/0xf/0xff << (offset)
> >
> > Where offset is:
> >
> > * bit 3 of the address if the size is 8
> > * bits 2 to 3 if it is 4
> > * bits 1 to 3 if size is 2
> > * bits 0 to 3 if size is 1 byte.
> >
> > That would create the bitmap, which would work for aligned as well as
> > misaligned hit/matches.
> >
> > The alternative is to do an NxN set of range compares, checking the
> *range*
> > of address plus size against all other addresses plus sizes.
> >
> > Questions
> >
> > 1 is this even necessary
> > 2 are there any alternative algorithms
> >
> > L.
> >
> >
> >
> > --
> > ---
> > crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68
> > _______________________________________________
> > libre-riscv-dev mailing list
> > libre-riscv-dev at lists.libre-riscv.org
> > http://lists.libre-riscv.org/mailman/listinfo/libre-riscv-dev
>
>
>
> --
> Samuel A. Falvo II
>
> _______________________________________________
> libre-riscv-dev mailing list
> libre-riscv-dev at lists.libre-riscv.org
> http://lists.libre-riscv.org/mailman/listinfo/libre-riscv-dev
>


-- 
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68