2020-10-16 07:40 skrev Sebastian Huber:
On 29/06/2020 13:27, Daniel Hellstrom wrote:
From: Daniel Cederman <ceder...@gaisler.com>
---
cpukit/score/cpu/sparc/cpu_asm.S | 6 ++++--
cpukit/score/cpu/sparc/syscall.S | 3 ++-
2 files changed, 6 insertions(+), 3 deletions(-)
diff --git a/cpukit/score/cpu/sparc/cpu_asm.S
b/cpukit/score/cpu/sparc/cpu_asm.S
index d5afd5f..bfad3fb 100644
--- a/cpukit/score/cpu/sparc/cpu_asm.S
+++ b/cpukit/score/cpu/sparc/cpu_asm.S
@@ -683,11 +683,13 @@ isr_dispatch:
cmp %l6, %l7
bne,a .Ldisable_fp
andn %l0, %l5, %l0
+ st %g0, [%g6 + SPARC_PER_CPU_FP_OWNER_OFFSET]
ba .Lthread_dispatch_done
- st %g0, [%g6 + SPARC_PER_CPU_FP_OWNER_OFFSET]
+ nop
.Ldisable_fp:
+ st %l0, [%fp + ISF_PSR_OFFSET]
ba .Lthread_dispatch_done
- st %l0, [%fp + ISF_PSR_OFFSET]
+ nop
.Lnon_fp_thread_dispatch:
#elif defined(SPARC_USE_SYNCHRONOUS_FP_SWITCH)
/* Test if we interrupted a floating point thread (PSR[EF]
== 1) */
diff --git a/cpukit/score/cpu/sparc/syscall.S
b/cpukit/score/cpu/sparc/syscall.S
index 4f4ef32..737a501 100644
--- a/cpukit/score/cpu/sparc/syscall.S
+++ b/cpukit/score/cpu/sparc/syscall.S
@@ -218,9 +218,10 @@ SYM(syscall_lazy_fp_switch):
.Lfp_save_done:
/* Restore the floating point context if necessary */
+ st %g0, [%l4 +
%lo(SPARC_THREAD_CONTROL_REGISTERS_FP_CONTEXT_OFFSET)]
cmp %l6, 0
be .Lfp_restore_done
- st %g0, [%l4 +
%lo(SPARC_THREAD_CONTROL_REGISTERS_FP_CONTEXT_OFFSET)]
+ nop
ldd [%l6 + SPARC_FP_CONTEXT_OFFSET_F0_F1], %f0
ldd [%l6 + SPARC_FP_CONTEXT_OFFSET_F2_F3], %f2
ldd [%l6 + SPARC_FP_CONTEXT_OFFSET_F4_F5], %f4
Some areas use the __FIX_LEON3FT_B2BST define to enable this
workaround. I am not sure if using the delay slots gives a measurable
performance boost. It would be good to mention the errata in the
documentation and give some advice how the workarounds should be
applied to RTEMS.
https://docs.rtems.org/branches/master/cpu-supplement/sparc.html
Indeed it is a bit messy, and we don't what to have comments all over
the SPARC code about 0009 either. At Gaisler we scan the images at every
release (and master from time to time) to discover inconsistencies and
then we address them if new instances are found. It makes sense to add a
LEON specific errata section. That will be good since it also document
what errata is worked around by RTEMS and toolchain. I will at least
document:
* listing all errata
* linking all the documents, for reference.
* describe how they are activated them in the build,
* what BSP and multilib is affected
I have updated the three patches related to TN-0009, now referencing
#4154. However I will consider #4154 closed once I submit proper
documentation in the sparc cpu supplement as propose. UI hope this is
acceptable.
Thanks,
Daniel
_______________________________________________
devel mailing list
devel@rtems.org
http://lists.rtems.org/mailman/listinfo/devel
