Created attachment 2126[details]
Fix uninitialized sections of VM instruction stream
I know I noted before (possibly informally in the mailing list) that in
vm_interpreted.c:VM_PrepareInterpreter(), the use of a single 'pc' variable
while expanding opcodes from a byte -> word (mainly for RISC platforms where
unaligned access causes a crash) leaves 3 uninitialized instructions in the
instruction stream.
The code uses the same PC for both the int array and the byte array when
loading 32-bit values. Since the PC is updated via += 4, this causes the
integer array to have 3 uninitialized values in them. It just so happens that
the interpreter loop also does a += 4 (i.e. skipping 4 integer values), so this
bug doesn't manifest itself as a crash. The solution is almost trivial -- use a
byte PC and an integer PC and update the interpreter to not skip 4 words on
certain opcode -- which is much more consistent.
Tested on IRIX 6.5.28 on Octane2 (interpreted QVM, no *.so files)
Created attachment 2126 [details] Fix uninitialized sections of VM instruction stream I know I noted before (possibly informally in the mailing list) that in vm_interpreted.c:VM_PrepareInterpreter(), the use of a single 'pc' variable while expanding opcodes from a byte -> word (mainly for RISC platforms where unaligned access causes a crash) leaves 3 uninitialized instructions in the instruction stream. The code uses the same PC for both the int array and the byte array when loading 32-bit values. Since the PC is updated via += 4, this causes the integer array to have 3 uninitialized values in them. It just so happens that the interpreter loop also does a += 4 (i.e. skipping 4 integer values), so this bug doesn't manifest itself as a crash. The solution is almost trivial -- use a byte PC and an integer PC and update the interpreter to not skip 4 words on certain opcode -- which is much more consistent. Tested on IRIX 6.5.28 on Octane2 (interpreted QVM, no *.so files)Comment on attachment 2126 [details] Fix uninitialized sections of VM instruction stream Changing mimetype of attachment.