FreeCalypso > hg > tcs211-l1-reconst
diff README.old @ 290:95f308ba55c7
README for the almost fully reconstructed state
author | Mychaela Falconia <falcon@freecalypso.org> |
---|---|
date | Mon, 20 Mar 2017 00:37:25 +0000 |
parents | README@53d8cb470dc7 |
children |
line wrap: on
line diff
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/README.old Mon Mar 20 00:37:25 2017 +0000 @@ -0,0 +1,61 @@ +The purpose of this Hg tree (derived from leo2moko-debug) is to facilitate an +attempt at reconstructing an L1 source that would match the binary object +version we got with our TCS211 semi-src delivery from Sotovik. + +The idea is to deblob TCS211 L1 one module or small group of modules at a time. +Take individual C modules from the LoCosto L1 source and get them to compile +in the TCS211 environment with TCS211 versions of all include files, producing +new from-source objects that would serve as candidate replacements for the ones +we got as blobs. An absolutely essential aspect of this approach is +incrementality: we should be able to deblob just a few L1 objects (replace them +with build-from-source versions) while keeping the original blobs for all other +parts of L1 we haven't reached yet, and regression-test at each step to ensure +that the firmware still works. + +TI's build system for this fw does not provide very fine granularity as far as +compiling each individual module from source or taking it as a delivered blob. +Instead it operates on the granularity of libs, or sometimes (as in the case of +L1) groups of libs. In the case of L1, this tree has the build system set up +to rebuild l1_ext.lib and l1_int.lib from source, which necessarily includes +_everything_ that goes into these libs. We are nowhere close to having a +working reconstructed source for every module in these two libs, thus we have +dummies in the place of most of them to appease BuSyB. Therefore, if you +simply run ./winebuild.sh to build this fw version, the link step will fail. + +Instead this fw has to be built in a somewhat convoluted way: + +1. Run ./winebuild.sh initially to compile everything leading up to the link + step, and let it fail. + +2. Go into g23m/__out__/gsm*/obj/l1_ext, and copy in original blob objects for + all those modules which have not been source-reconstructed yet. Do likewise + for l1_int. + + [Update: the copyblobs.sh script does this job now.] + +3. Run ./winebuild.sh again. As long as the dummy C files have not been + touched and the copied-in objects have newer timestamps, the dummies won't + be recompiled and the copied-in objects won't get blown away. TI's make + will rerun the ar packaging step to produce l1_ext.lib and l1_int.lib from + the mix of objects (some compiled from source, others copied-in blobs) in + the obj directories, and then the final link should succeed. + +The development process is as follows: + +1. Choose an individual L1 C module to work on. + +2. Copy the source for the chosen module from the LoCosto source in the place + of the dummy in this tree. + +3. Run ./winebuild.sh in the g23m directory and examine the output in + report.txt. If the imported C module fails to compile, massage it until it + passes compilation. + +4. Once compilation passes, look at the generated COFF object in + g23m/__out__/gsm*/obj/l1_{ext,int}. Disassemble it with the tiobjd tool + from freecalypso-reveng, and compare against the disassembly of the + original ("official") binary blob version we got. Massage the reconstructed + source until it compiles into object code that matches the original. + + [Update: use the scripts under g23m/objdiff to facilitate the disassembly + and diffing.]