| commit | 93ac0eaba030a055ceccade34e9c2ed6cd4ba7ae | |
| author | Douglas Katzman <dougk@google.com> | |
| Fri, 22 Jan 2016 02:05:31 +0000 (21 21:05 -0500) | ||
| committer | Douglas Katzman <dougk@google.com> | |
| Fri, 22 Jan 2016 02:15:10 +0000 (21 21:15 -0500) | ||
| tree | 30c54fa7f202d2e1f4aeba5f5fa6bc7173bcd619 | treesnapshot (tar.gz zip) |
| parent | 0f750f2697a4185b6bba65a085d9c8b5cc21a9d7 | commitdiff |
sb-disassem: Better call convention for prefilters.
The following changes are limited to x86[-64] based on limited
ability to test other backends (it's architecture-independent,
but the prefilter functions have to be updated).
* Don't construct and compile a new lambda expression that calls
each prefilter for the current instruction - instead just loop
over operands based on the instruction's metadata.
* Call each prefilter with the right number of arguments: 0, 1, >1,
depending on how many byte specifiers correspond to the operand.
This makes the ModR/M prefilter receive 2 separate bytes instead
of a list, and eliminates most of the (IGNORE) declarations.
Despite the unoptimized calls of LDB in CALL-PREFILTERS, this is a
win, because bit extraction does not cons, nor do we cons arglists.
The following changes are limited to x86[-64] based on limited
ability to test other backends (it's architecture-independent,
but the prefilter functions have to be updated).
* Don't construct and compile a new lambda expression that calls
each prefilter for the current instruction - instead just loop
over operands based on the instruction's metadata.
* Call each prefilter with the right number of arguments: 0, 1, >1,
depending on how many byte specifiers correspond to the operand.
This makes the ModR/M prefilter receive 2 separate bytes instead
of a list, and eliminates most of the (IGNORE) declarations.
Despite the unoptimized calls of LDB in CALL-PREFILTERS, this is a
win, because bit extraction does not cons, nor do we cons arglists.