GitLab

remove some old cruft

Change-Id: Ie94b7dd352acec017979ce76fcc6d0f7fbd197b8

Prior to this cmake could only find "grub-mkrescue".

Change-Id: Id16ce4958d826953387c80865aafe47a5660eecd

In the FSM representation, interrupts are no longer modled as separate
tasks.

Change-Id: I3c9de5f629a5b36330e1339263700efabd34b822

Change-Id: If54efba52eed798bf95dda470f5b5bf714eb6f60

Change-Id: Ic2554caf81492cb93b73c423a28fa014c6288ce5

We explicitly enable the popcnt instruction in the code generation stage.

Change-Id: I799aabebede13da1dae3adc0e8cd864b4bc817f6

The priority of a task is caluclated from its dynamic-priority state
word and the event state words. Until now, we did not check this value
for being a correctly encoded value, which causes quite a number of SDCs
in the bench-coptermock-isorc numbers. Therefore, we check this value
for validity and hope for the best.

Change-Id: Ie95c1b64b56d79a6110455270fb099fee919269f

It is unwise to overwrite the set-event bitmask on a SetEvent system
call. ecc1_event1f will ensure the correct behavior in the future.

Change-Id: I98bf21664d381e6093c194ad0442ee168fc92764

Change-Id: I1d13300681efa07a74d2f82eef6a56f62e682fb1

Change-Id: I9a2790cf1dac9c068bd4ff98ae43d8bf9ba74613

Somes these testcases produced

  5{<><><>}33x

instead of

  5{<><><>}3x

This is not a application or OS bug, but a race condition with the
waiting times.

Change-Id: I1e847e13188b15d567ce4efd92f473a4f39a666b

The symbolic system execution is now able to enumerate all possible
system states with two different approaches. Either we can use the "old"
version which uses states with blocks to be executed in the next step,
like it is described in LCTES'15 (Dietrich et. al).

Or we can use the application finite state machines as basis. These
state machines are smaller (lesser nodes) than the application's
CFG. They are based on the task's ICFG.

Change-Id: I05e050ed3f39bd18ecf86b4d7bb423b4ce0825e0

The ApplicationFSM pass transforms the ICFG of a task to a finite state
machine. This construction is done by transforming CFG-edges to
FSM-states, and CFG-blocks to FSM-transitions.

In a second step, the CFG-computation blocks are handled as FSM-epsilon
transitions. Then a standard epsilon elimination is done for
subtask-FSM.

Change-Id: I874bc3b17ebaf16b58bb70226e5eef091e5cbbe9

Added a modifier for llvm bitcode files, which adds function attributes +
values to a specific function.

Change-Id: I57d71147c776686bfd18025d3090485b69da6bf8

Change-Id: Iebb4498dcb5c8a1ebb6bae8d13b75e4645e1cba6

Change-Id: I38d53965e1dfe1c337a946bc7d47239a62a5cd3e

Change-Id: I14a18026ef4e6320dd0c6da61fb8f6c4892ca16f

Not all bits in the cfg region word are used, therefore we can safely
reset them so a known state afterwards.

Change-Id: Ic059b80bf46f9a53742383025e495e653837cb68

Change-Id: I83cac7005d667bdfd2045f8e0e58c03bddf1a475

Inlining the scheduler reveals a high potential for specialization and
and SDC reduction. In the implementation of the zedboard benchmark, we
dropped the inlinability support. This is now possible again. With
inline-scheduler+specialization the scheduler is paritally instansiated
whereever possible.

Test-Config: --arch i386 --encoded=no --inline-scheduler=yes --specialize=yes
Test-Config: --arch i386 --encoded=no --inline-scheduler=yes --specialize=no
Test-Config: --arch i386 --encoded=no --inline-scheduler=no --specialize=yes
Test-Config: --arch i386 --encoded=no --inline-scheduler=no --specialize=no
Test-Config: --arch i386 --encoded=yes --inline-scheduler=yes --specialize=yes
Test-Config: --arch i386 --encoded=yes --inline-scheduler=yes --specialize=no
Test-Config: --arch i386 --encoded=yes --inline-scheduler=no --specialize=yes
Test-Config: --arch i386 --encoded=yes --inline-scheduler=no --specialize=no

Change-Id: I2749078a3cc4ebeeb19bde14eb769b5781ac409f

Inlining is hard. Convincing the compiler to actually inline the
systemcall bodies is even harder. Therefore we now use the always_inline
flag together with the -always-inline (llvm-opt) compiler
pass. Furthermore -gc-section drops all unused functions.

Change-Id: Ifd429e28ebb330944230076bca3886f4ef1171e3

Until know there where two different versions of scheduler.h.in, one for
encoded and one for the unencoded variant. In essence they were nearly
identical. Therefore, they could be unified without changing the
execution sequence.

Change-Id: Ifff4cd4a18bee135e286f330212aa14247cc2277

Like the last change, this one does not change the SDC rate, but does only make
them detectable by application/FAIL*. Until now, MPU faults were routed to isr_undetected

Change-Id: Iac81f8611753cafbddb6d1f9c4bbe505a788890e

Until now, unhandled interrupt resulted in a hanging machine. This did *not*
cause the SDC rates to change, but made the TIMEOUT numbers for fault injection
campaigns slightly too high.

Change-Id: I4616c5ddad66659bab604cd3ed5b5405ceea44ce

This includes using nm again, instead of llvm-nm.

Change-Id: Ic8b1b6054427ddaf028ece5e7758309ee0d4b3be

Change-Id: Ia516171592bc51c9fcef37fa9d2f2c9e947dd148

Change-Id: I5303f2cd78122dd07921a97c5a598254bc3915fd

Change-Id: I35c6ef1c68bbcc016602b7ee70dd41a48a8a4087

unreachable should not, and does not return.
This also silences a compiler warning.
Uses new C++11 attribute syntax [[noreturn]].

Change-Id: Ifb4438bffb6d6df641003c53d3a2bf4a825b403c

Since privilege isolation is a very costly feature, and we're not sure
wheter it is good or bad for dependability, we make it optional. This
allows us to evaluate it separately.

Test-Config: --arch i386 --privilege-isolation f
Test-Config: --arch i386 --privilege-isolation f --mpu t
Test-Config: --arch i386 --privilege-isolation f --encoded t
Test-Config: --arch i386 --privilege-isolation f --basic-tasks t
Change-Id: Ia22a41eb610b090aeaac486d006a4b4ea2ab762a

The push %edi was totally unecessary.

Change-Id: I4fc8806c722a1690f6a83678c35aa508ee64c3cf

All MMU related datastructres are now removed from the image, if dOSEK
is build without MPU.

Change-Id: If01c87027df73b92f3df979d0955098c4b5c91c3

The disable interrupts at that place, caused the bench-coptermock to break.

Change-Id: Iedea2a6c259e276c3101dc8970ff29a2e1b57e9a

We implement the PreciseSystemState with a more efficient variant. It
uses the __slot__ feature of python (therefore I removed it from the
inheritance chain, which causes some copying of code).

With this change the memory consumption for bench-coptermock drops from
~60Mb to ~40Mb for the SymbolicSystemExecution pass. With pypy3, the
memory consumption drops to 23Mbyte.

Change-Id: Ic87be8afa97d6861dfe49c9cffd40d3bdc6464a5

In order drop our dependency on llvmpy, which seems to be abandoned at
the moment, I implemented the llvm-extractor tool. It:

- splits basic blocks before and after each CallInst/InvokeInst
- rename the system calls
- extracts the system call arguments
- dumps a python dict with the strucutre

Change-Id: I2adac08c14ceb2e49ffe1bbcb43216b6769b06c1

In order to port the "old" isorc benchmark, I had to adapt the code only
a little bit to work around the constant-referencing of system objects
pollicy of dOSEK. Facts about the benchmark are:

- If run at 50Mhz
  - Takes 1.3 seconds (The original ISOR benchmark took 1.29 seconds)
  - Emits 110 checkpoints
    - Original CiAO: 107 (almost no diff to the current trace)
    - Original eCOS: 110 (two execution blocks are shuffled)

Change-Id: I6ab96fae29e1c8376da95ec4f622dc1d92843f4d

The system call was already implemented in the scheduler API, but was
not exposed to the applications.

Change-Id: I08206e0a5e83011b63efb8257c1a57c00a93a8ba

trace-analyze.py did crash on the current traces due to a API change in
stats_binary. Additionally syscall_end markers are now also detected as
end of system-call--regions markers.

Change-Id: I2d9c09d235a6c2afda34d2b6aace86665b39d2b2

Change-Id: Ic573cdf0d03f55fd523007127c15a0480ef436c8