[ixp1200] Problem in microcode optimizer

[Heaton, Eric D]

All,

This assembler bug is documented in the uca_rel_notes.txt file and in the
Release Notes (found in the IXA SDK installation), as well as one
additional:

o) The optimizer has the following defects:

   1. The optimizer will illegally place a P1 stage ctx_arb immediately
after a
      P3 stage branch.

   2. The optimizer will incorrectly move +ifsign and +carry ALU operations,
      changing the instructions from which they get the condition codes.

One more <undocumented> optimizer bug concerns using it with the find_bset
instruction and indirect shifting:

	alu[--, shift, b, tmpa]
	find_bset[search_reg, >>indirect]	

That is, the optimizer will move the alu statement used as the indirect
shift value incorrectly somewhere below the find_bset instruction.

Also note that the optimizer does not support is not currently supported for
multiple segments (i.e. across the 2k control store), but will be in a
future release.

Be sure to keep these in mind when using the optimizer in the assembler!

     Eric Heaton
     Technical Marketing Engineer
     Intel Corporation

-----Original Message-----
From: Alok Kumar [mailto:alok at cs.utexas.edu]
Sent: Monday, April 30, 2001 6:28 PM
To: ixp1200 at CS.Princeton.EDU
Subject: [ixp1200] Problem in microcode optimizer


Hello Everyone,

I found a problem in microcode optimizer in the Developer Workbench, so I
thought to send it through in the mailing list so that you can turn off
the code optimizer is your optomized code has similar problem.

The problem is when we use "alu" instruction with operation  "+ifsign".
Probably, optimizer does not account for the fact the resulf of alu with
+ifsign depends on the condition code. For example the following code is
optimized incorrectly

.local num den res diff
alu[num, num, -, den]						; instr 1
alu[--,--, B, num]						; instr 2
alu[res, 1, +IFSIGN, res, <<1]					; instr 3
alu[num, den, +IFSIGN, num]					; instr 4

alu[den, --, B, den, >>1]					; instr 5
alu[diff, diff, -, 1]						; instr 6
br<0[header_stored#]						; instr 7
.endlocal

The optimizer expands it as:

.local num den res diff
alu[l000!num, l000!num, -, l000!den]				; instr 1
alu[--,--, b, l000!num]						; instr 2
alu_shf[l000!res, 1, +ifsign, l000!res, <<1]			; instr 3
alu_shf[l000!diff, l000!diff, -, 1, 0]				; instr 6
br<0[header_stored#],						; instr 7
	defer[2]
	; BRANCH LATENCY FILL OPTIMIZATION:  the uword below was "pushed"
down 2 positions
alu[l000!num, l000!den, +ifsign, l000!num]			; instr 4
	; BRANCH LATENCY FILL OPTIMIZATION:  the uword below was "pushed"
down 2 positions
alu_shf[l000!den, --, b, l000!den, >>1]				; instr 5
.endlocal

Here semantics of instr 4 changes in optimized code and it works on
condition code set by instr 6 instead of condition code set by instr 2 as
it should. Using +carry in alu instruction also has similar problem. I
think optimizer assumes that alu instruction does not depend on any
condition
codes. Please, turn off optimizer if you are using +ifsing or +carry.

Thanks
Alok


____________________________________________________________________________
___

Alok Kumar

Residence:
3401 Red River St #201
Austin, TX 78705

Office:
Dept of Computer Science
ACES Building
University of Texas
Austin, TX 78712

Ph:  Home   - (512) 472-6443
     Office - (512) 232-7883
Fax: (401) 679 8171

Homepage: http://www.cs.utexas.edu/~alok
____________________________________________________________________________
___