optimization-tricks.txt

A collection of cute optimization tricks, mostly things I haven't had time to follow up on...

Loads from constant globals with patterns in data
(Note: LLVM currently handles some cases w/compares, but that's all I can find)
a = constant global [1,3,5,7,9]
a[i] -> -1 + i * 2
(and other assorted idioms)

Use masked loads w/pass through to eliminate selects of addresses
a = select c, a1, a2
v = load a
->
v1 = masked.load(c, a1, undef)
v = masked.load(!c, a2, v1)
(can do better if either are safe to speculate)

trivial specialization (no cost dispatch)
foo(..., bool b) { if (b) C; else D; }
->
foo(...) { if(b) tail C(); else tail D(); }
(potentially very interesting w/early return idioms)


Tricks for encoding long jumps with fewer bytes using implicit fault handlers
LongJCC: # 6 bytes
  jle baz

ShortJCC: # 2 bytes
  jle bar
  
forward_jump_int3: # 3 bytes
  jle continue
  int3 # bad to have in instruction stream?
continue: 

forward_jump_ud2: # 4 bytes
  jle continue2
  ud2 # bad to have in instruction stream?
continue2: 

jcc_payload:  # 4 bytes
  test %rax, %rcx
  .byte 116 #0x74, je
fault:
  .byte 204 #0xcc, int3 as payload of je
  jle fault #backbrach into previous jcc

and_payload: % 4 bytes
  test %rax, %rcx
  .byte 4 #0x04, add AL, 204
fault2:
  .byte 204 #0xcc, int3 as payload of add
  jle fault2 #backbrach into previous add
  
bar:
  retq
.section ".other"
baz:
  retq