This follows the x86 xstate changes and implements a task_xstate slab
cache that is dynamically sized to match one of hard FP/soft FP/FPU-less.
This also tidies up and consolidates some of the SH-2A/SH-4 FPU
fragmentation. Now fpu state restorers are commonly defined, with the
init_fpu()/fpu_init() mess reworked to follow the x86 convention.
The fpu_init() register initialization has been replaced by xstate setup
followed by writing out to hardware via the standard restore path.
As init_fpu() now performs a slab allocation a secondary lighterweight
restorer is also introduced for the context switch.
In the future the DSP state will be rolled in here, too.
More work remains for math emulation and the SH-5 FPU, which presently
uses its own special (UP-only) interfaces.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
A number of small optimisations to FPU handling, in particular:
- move the task USEDFPU flag from the thread_info flags field (which
is accessed asynchronously to the thread) to a new status field,
which is only accessed by the thread itself. This allows locking to
be removed in most cases, or can be reduced to a preempt_lock().
This mimics the i386 behaviour.
- move the modification of regs->sr and thread_info->status flags out
of save_fpu() to __unlazy_fpu(). This gives the compiler a better
chance to optimise things, as well as making save_fpu() symmetrical
with restore_fpu() and init_fpu().
- implement prepare_to_copy(), so that when creating a thread, we can
unlazy the FPU prior to copying the thread data structures.
Also make sure that the FPU is disabled while in the kernel, in
particular while booting, and for newly created kernel threads,
In a very artificial benchmark, the execution time for 2500000
context switches was reduced from 50 to 45 seconds.
Signed-off-by: Stuart Menefy <stuart.menefy@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
sh port of the sLeAZY-fpu feature currently implemented for some architectures
such us i386.
Right now the SH kernel has a 100% lazy fpu behaviour.
This is of course great for applications that have very sporadic or no FPU use.
However for very frequent FPU users... you take an extra trap every context
switch.
The patch below adds a simple heuristic to this code: after 5 consecutive
context switches of FPU use, the lazy behavior is disabled and the context
gets restored every context switch.
After 256 switches, this is reset and the 100% lazy behavior is returned.
Tests with LMbench showed no regression.
I saw a little improvement due to the prefetching (~2%).
The tests below also show that, with this sLeazy patch, indeed,
the number of FPU exceptions is reduced.
To test this. I hacked the lat_ctx LMBench to use the FPU a little more.
sLeasy implementation
===========================================
switch_to calls | 79326
sleasy calls | 42577
do_fpu_state_restore calls| 59232
restore_fpu calls | 59032
Exceptions: 0x800 (FPU disabled ): 16604
100% Leazy (default implementation)
===========================================
switch_to calls | 79690
do_fpu_state_restore calls | 53299
restore_fpu calls | 53101
Exceptions: 0x800 (FPU disabled ): 53273
Signed-off-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: Stuart Menefy <stuart.menefy@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This fixes a bug in the FPU exception handler for the FCNVDS instruction.
To get the register number the instruction is shifted right by 9,
though it should be shifted right by 8.
More information at ST Linux bugzilla:
https://bugzilla.stlinux.com/show_bug.cgi?id=4892
Signed-off-by: Giuseppe Di Giore <giuseppe.di-giore@st.com>
Signed-off-by: Carmelo Amoroso <carmelo.amoroso@st.com>
Signed-off-by: Stuart Menefy <stuart.menefy@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This patch adds support for the SH4 to convert a subnormal double
into a float by catching the FPE and implementing the FCNVDS
instruction in software.
Signed-off-by: Carl Shaw <carl.shaw@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This follows the sparc changes a439fe51a1.
Most of the moving about was done with Sam's directions at:
http://marc.info/?l=linux-sh&m=121724823706062&w=2
with subsequent hacking and fixups entirely my fault.
Signed-off-by: Sam Ravnborg <sam@ravnborg.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Presently with preempt enabled there's the possibility to be preempted
after the TIF_USEDFPU test and the register save, leading to bogus
state post-__switch_to(). Use an explicit preempt_disable()/enable()
pair around unlazy_fpu()/clear_fpu() to avoid this. Follows the x86
change.
Reported-by: Takuo Koguchi <takuo.koguchi.sw@hitachi.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
There are a few different cases for figuring out how to
size the instruction. We read in the instruction located
at regs->pc - 4 when rewinding the opcode to figure out if
there's a 32-bit opcode before the faulting instruction, with
a default of a - 2 adjustment on a mismatch. In practice this
works for the cases where pc - 4 is just another 16-bit opcode,
or we happen to have a 32-bit and a 16-bit immediately
preceeding the pc value.
In the cases where we aren't rewinding, this is much less ugly..
We also don't bother fixing up the places where we're explicitly
dealing with 16-bit instructions, since this might lead to
confusion regarding the encoding size possibilities on other
CPU variants.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Remove the previous saving of fault codes into the thread_struct
as they are never used, and appeared to be inherited from x86.
Signed-off-by: Stuart Menefy <stuart.menefy@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This fixes up the kernel for gcc4. The existing exception handlers
needed some wrapping for pt_regs access, acessing the registers
via a RELOC_HIDE() pointer.
The strcpy() issues popped up here too, so add -ffreestanding and
kill off the symbol export.
Signed-off-by: Stuart Menefy <stuart.menefy@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
