(address . guix-patches@gnu.org)(name . Simon South)(address . simon@simonsouth.net)
20200607151709.21821-1-simon@simonsouth.net
This patch allows JamVM 2.0.0 to be built without using outdated versions of
gcc and glibc, by disabling the branch-patching optimization JamVM normally
applies to code it has inlined.
I've successfully built IcedTea 1.13.13 on x86_64 with this patch applied,
without encountering any "Invalid instruction" errors. It also seems to work
fine on aarch64.
Lengthy explanation, for anyone interested in the details:
JamVM's branch-patching optimization tries to improve the performance of
Java's branching opcodes by replacing the load-operand-and-jump portion of
their implementation with a single jump instruction directly to the handler
for the opcode to which the operand points. However, the current
implementation
- Doesn't attempt much of an optimization, since it only replaces the final
jump instruction with another (the operand is still loaded in either case).
- Doesn't have any actual effect since the instruction it replaces is a
duplicate synthesized by gcc that is never executed anyway.
- Is prone to breakage, since it doesn't ensure the new jump instruction is
the same length as the original. (This is the specific reason for the
failure on x86_64: The replacement jump instruction is longer and clobbers
part of the following instruction, making it invalid.)
This patch simply disables the optimization within JamVM, leaving the other
optimizations intact. (gcc is smart enough to no longer generate duplicate
jump instructions in this case so it reduces the code size slightly, too.)
Alternate solutions I considered and rejected:
- Fixing the implementation. JamVM uses a label to mark within each opcode's
handler where the jump instruction should be placed, and moving this label
to the start of the load-and-jump sequence rather than the end appears (from
stepping through the code with gdb) to fix all the issues above. However,
JamVM then fails at startup, reporting an unhandle-able exception during
initialization. So presumably some other part of the code relies on the
current, broken implementation of this feature, and while it's probably
possible to fix _that_ as well I doubt it would be worth the effort.
- Disabling the optimization at runtime. Invoking JamVM with "-Xnopatching"
also solves the problem, and we could just update the bootstrap procedure to
do this. However JamVM 1.5.1 doesn't recognize this option and fails if it's
specified, so we'd need to change the ecj-javac-wrapper package to handle
the two interpreters differently or to accept an argument specifying the
flags to use, and in my opinion this would be a step backwards from what
exists currently.
Also, since JamVM would only ever work reliably when this option is
specified, it's not really much of an "option" and making its effect
permanent in the code seems like a more sensible approach.
--
Simon South
simon@simonsouth.net
Simon South (1):
gnu: jamvm: Fix to work with current gcc and glibc.
gnu/packages/java.scm | 7 ++---
.../jamvm-2.0.0-disable-branch-patching.patch | 27 +++++++++++++++++++
2 files changed, 29 insertions(+), 5 deletions(-)
create mode 100644 gnu/packages/patches/jamvm-2.0.0-disable-branch-patching.patch
--
2.26.2