1 * use a pool of MBState structures to speedup monoburg instead of using a
3 * the decode tables in the burg-generated could use short instead of int
4 (this should save about 1 KB)
5 * track the use of ESP, so that we can avoid the x86_lea in the epilog
10 * the ORP people avoids optimizations inside catch handlers - just to save
11 memory (for example allocation of strings - instead they allocate strings when
12 the code is executed (like the --shared option)). But there are only a few
13 functions using catch handlers, so I consider this a minor issue.
15 * some performance critical functions should be inlined. These include:
16 - mono_mempool_alloc and mono_mempool_alloc0
17 - EnterCriticalSection and LeaveCriticalSection
19 - mono_metadata_row_col
20 - mono_g_hash_table_lookup
22 * the managed/unmanaged boundary is quite slow:
23 - it calls mono_get_lmf_addr, which calls TlsGetValue, which calls
24 pthread_getspecific (). This means that 3 function calls are needed for
25 each native function call.
27 * if a function which involves locking is called from another function which
28 acquires the same lock, it might be useful to create a separate _inner
29 version of the function which does not re-acquire the lock. This is a perf
30 win only if the function is called a lot of times, like mono_get_method.
32 * the frame_state_for function in glibc 2.3.2 can't correctly decipher the
33 unwind tables generated by gcc 3.3. It allways tells the runtime that not all
34 callee saved registers are saved, even when the icall is marked with
35 MONO_ARCH_SAVE_REGS. This forces the runtime to generate wrapper functions
36 for all icalls, slowing things down greatly.
38 * we can avoid calls to class init trampolines if the are multiple calls to the
39 same trampoline in the same basic block. See:
41 http://bugzilla.ximian.com/show_bug.cgi?id=51096
46 * Remove the various optimization list of flags description, have an
47 extra --help-optimizations flag.
49 * Remove the various graph options, have a separate --help-graph for