[Python-Dev] Reordering opcodes (PEP 203 Augmented Assignment)

[Guido van Rossum]

> Separating out LOAD_FAST from the switch shows a nice effect.
> SET_LINENO is removed by -OO anyway, so there's really no
> use in optimizing this one.

I tried this and found about three percent speed increase on pystone,
for what that's worth.  This is with python -OO on Linux x86.  Note
that removing the (now redundant) case from the switch seemed to make
a small difference too.

Alas, I have no time to play with optimizing the main loop in a more
rigorous way... :-(

Here's the patch I came up with:

Index: ceval.c
===================================================================
RCS file: /cvsroot/python/python/dist/src/Python/ceval.c,v
retrieving revision 2.187
diff -c -r2.187 ceval.c
*** ceval.c	2000/07/25 12:56:38	2.187
--- ceval.c	2000/07/30 16:13:23
***************
*** 608,616 ****
  		f->f_lasti = INSTR_OFFSET();
  #endif
  		
  		opcode = NEXTOP();
! 		if (HAS_ARG(opcode))
  			oparg = NEXTARG();
  #ifdef DYNAMIC_EXECUTION_PROFILE
  #ifdef DXPAIRS
  		dxpairs[lastopcode][opcode]++;
--- 608,631 ----
  		f->f_lasti = INSTR_OFFSET();
  #endif
  		
+ 	  get_opcode:
  		opcode = NEXTOP();
! 		if (HAS_ARG(opcode)) {
  			oparg = NEXTARG();
+ 			if (opcode == LOAD_FAST) {
+ 				x = GETLOCAL(oparg);
+ 				if (x != NULL) {
+ 					Py_INCREF(x);
+ 					PUSH(x);
+ 					goto get_opcode;
+ 				}
+ 				PyErr_SetObject(PyExc_UnboundLocalError,
+ 					   PyTuple_GetItem(co->co_varnames,
+ 							oparg));
+ 				goto on_error;
+ 			}
+ 		}
+ 
  #ifdef DYNAMIC_EXECUTION_PROFILE
  #ifdef DXPAIRS
  		dxpairs[lastopcode][opcode]++;
***************
*** 1282,1300 ****
  			}
  			Py_INCREF(x);
  			PUSH(x);
- 			break;
- 
- 		case LOAD_FAST:
- 			x = GETLOCAL(oparg);
- 			if (x == NULL) {
- 				PyErr_SetObject(PyExc_UnboundLocalError,
- 					   PyTuple_GetItem(co->co_varnames,
- 							oparg));
- 				break;
- 			}
- 			Py_INCREF(x);
- 			PUSH(x);
- 			if (x != NULL) continue;
  			break;
  
  		case STORE_FAST:
--- 1297,1302 ----