07:22
What I will need is an object format for the generic namespace blobs.
07:26
I could use standard ELFs with relocations, however the blobs would not be directly ROM executable and such. So I suppose what I can do is have the code and data treads placed in the file and the table of contents placed at the end. There would be one for classes and another for resources. Although it may be possible to have relocations and not have them at the same time. Relocations could be useful for the end linking process into the SquirrelJME binary.
12:28
I received the Creator CI40 today, so this will be an interesting device to develop for.
# uname -a
Linux OpenWrt 4.1.13 #1 SMP Tue Jul 26 02:49:14 UTC 2016 mips GNU/Linux
# cat /proc/cpuinfo
system type : IMG Pistachio SoC (B0)
machine : IMG Marduk - Ci40 with ca8210
processor : 0
cpu model : MIPS interAptiv (multi) V2.0 FPU V0.0
BogoMIPS : 363.72
wait instruction : yes
microsecond timers : yes
tlb_entries : 64
extra interrupt vector : yes
hardware watchpoint : yes, count: 4, address/irw mask: [0x0ffc, 0x0ffc,
0x0ffb, 0x0ffb]
isa : mips1 mips2 mips32r1 mips32r2
ASEs implemented : mips16 dsp mt eva
shadow register sets : 1
kscratch registers : 0
package : 0
core : 0
VCED exceptions : not available
VCEI exceptions : not available
VPE : 0
processor : 1
cpu model : MIPS interAptiv (multi) V2.0 FPU V0.0
BogoMIPS : 272.79
wait instruction : yes
microsecond timers : yes
tlb_entries : 64
extra interrupt vector : yes
hardware watchpoint : yes, count: 4, address/irw mask: [0x0ffc, 0x0ffc,
0x0ffb, 0x0ffb]
isa : mips1 mips2 mips32r1 mips32r2
ASEs implemented : mips16 dsp mt eva
shadow register sets : 1
kscratch registers : 0
package : 0
core : 0
VCED exceptions : not available
VCEI exceptions : not available
VPE : 1
processor : 2
cpu model : MIPS interAptiv (multi) V2.0 FPU V0.0
BogoMIPS : 363.72
wait instruction : yes
microsecond timers : yes
tlb_entries : 64
extra interrupt vector : yes
hardware watchpoint : yes, count: 4, address/irw mask: [0x0ffc, 0x0ffc,
0x0ffb, 0x0ffb]
isa : mips1 mips2 mips32r1 mips32r2
ASEs implemented : mips16 dsp mt eva
shadow register sets : 1
kscratch registers : 0
package : 0
core : 1
VCED exceptions : not available
VCEI exceptions : not available
VPE : 0
processor : 3
cpu model : MIPS interAptiv (multi) V2.0 FPU V0.0
BogoMIPS : 363.72
wait instruction : yes
microsecond timers : yes
tlb_entries : 64
extra interrupt vector : yes
hardware watchpoint : yes, count: 4, address/irw mask: [0x0ffc, 0x0ffc,
0x0ffb, 0x0ffb]
isa : mips1 mips2 mips32r1 mips32r2
ASEs implemented : mips16 dsp mt eva
shadow register sets : 1
kscratch registers : 0
package : 0
core : 1
VCED exceptions : not available
VCEI exceptions : not available
VPE : 1
# free -m
total used free shared buffers cached
Mem: 185380 34200 151180 84 0 6888
-/+ buffers/cache: 27312 158068
Swap: 0 0 0
Going to install a Debian chroot since there currently is no Linux kernel available.
15:11 When
I should probably support reading the blob format before I write it, so I can have a clue as to how to write it out.
18:17
Appears the CPU on the device does not support denormal floating point operations, which is totally reasonable, and that the Linux kernel is not emulating these. I will need to make a note of this and have an illegal instruction handler for some CPUs to emulate operations on denormals so that even though they are not implemented by the CPU, the host code can still run.
19:56
I do wonder though if there is an extremely low memory way of writing that is not a complete mess and is quick to parse while also having some sense of security. I would like things done in a single pass. What if everything was a native method, including the class definitions. I really just need code and resources. However they should not be together at all, since if there is a breach then data could be executed as code (which would be very bad).
20:06
That actually could be something. Instead of having it where there is just
static data that is handled by an object format reader, instead each namespace
has an entry initialization. Basically this would be generated machine code
that when called will setup the namespace as it should be setup on the running
VM. Everything would essentially be executable for the most part. When a
namespace is desired to be used then its boot code is used which registers
everything. However this code will be shared among all namespaces for the most
part. When namespaces are read, there will need to be linear scan of
everything anyway. The only issue would be determining the size of entries
before they are written and actually known (unless ZipFile is used, but
every entry is treated as a stream of bytes anyway). When a classpath is
initialized, the namespace should be scanned and then code linked in as such
and classes registered as needed.
20:08
However, I am kind of going in circles. I want a really fast to output format that is sane and does not require lots of memory to use, but is easy and simple to parse. I will need an ending table listing classes and such. So perhaps I can just dump strings and code into the output, then build up an ending table. The class table will have references to the class header information along with fields and methods that are declared in the class (at known file addresses). Classes in the final table could be sorted by name so that they are easier to be found. Resources can be pretty much the same. I can use a single output stream format for this. There would be no string table at all. There would have to be an import table (for classes and field offsets), but just the information that is needed to initialize it. So essentially structures and code would be together (even though I dislike that) and the ending table would handle most of the things. So this would be similar to my original plan but a bit more sane as everything is in tables.