---------------------------------

OPTIMIZE.TXT

---------------------------------

Some useful informations about how to optimize code on a 386/486/Pentium

This initially was a posting on the Internet made by

        Michael Kunstelj.
        Contact at :  virteam@smoke.canberra.edu.au
                        or
                      mick@cairo.anu.edu.au

Initially it included 486 and Pentium optimizations with some "holes"
(like the profiling instructions), i filled the Pentium "holes"
then refined some exaplanations (uh! maybe i made 'em worse! :} ).
Then added some not-so-obviuos cache optimization techniques
and other infos about specific optimizations you can make on 386.

        Lorenzo Micheletto
N.B.
        I added some generic examples about how a cpu works
        BUT they are just examples, intel cpus work in slightly different ways.

N.B./2
        Always check in the books for things you are not sure
        there may be typos here.

----------------------------------
Brief intro about how a CPU works
----------------------------------

Now a brief intro on the innards Intel CPUs.

Most processors these days have within in them a system termed a "pipeline".
The 486 / 586 are certainly within this category.  However, most people 
aren't quite sure what exactly a pipeline is, here follows a complete
explanation:

The execution of a machine code instruction can be roughtly split
in five stages:   [n.b. this is a generic pipeline]
        1) FETCH instruction opcode and immediate data
        2) DECODE opcode and align immediata data into temporary registers
        3) CALCULATE ADDRESS OF OPERANDS and load 'em into memory.
           (calculate the address of memory locations to access
            and select the register operands to use)
           (sometimes called DECODE 2)
        4) EXECUTE instruction (loading memory operands if needed)
           & write results to INTERNAL registers
        5) WRITEBACK memory-operand results to memory

Every stage takes ONE OR MORE cpu cycles to be completed, but usually
a modern cpu needs just one cpu cycle for every execution stage
(excluding stages 1 and 5 that have to deal with memory/cache access
 and stage 4 when "complex" microcoded instructions are executed).

A cpu-core ( the "real" cpu, not cache support nor the other things
that are usually integrated into a modern cpu) has five "specialized"
units, one for every stage of instruction execution, plus a "register file"
(and ultra-fast on-chip memory where internal register values
 are stored) and a CONTROL UNIT.

The control unit "coordinates" the action of all the other units
so they can work together.

Here follows an extended ascii drawing
of ONE on the many ways these units can be connected
(use a ms-dos text editor to see this, or convert it to the
 Windows line-drawing character set if you are using a Windows editor):

              MEMORY INPUT ("memory load" unit)
                        ³                                ÚÄÄÄÄÄÄÄÄÄÄÄÄ¿
                        ÀÄÄÄÄÄÂÄ>ÚÄÄÄÄÄÄÄÄÄÄÄÄ¿<ÄÄÄÄÄÄÄÄ>³            ³
                             ÚÅ¿ ³ FETCH      ³          ³            ³
       ÚÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÙ³À>ÀÄÄÄÄÄÄÄÄÄÄÄÄÁÄÄ¿       ³            ³
       ³ (instruction pointer)³                  ³       ³            ³
       ³                      ³  ÚÄÄÄÄÄÄÄÄÄÄÄÄ¿<ÄÙ       ³            ³
       ³                      ³  ³ DECODE     ³<ÄÄÄÄÄÄÄÄ>³            ³
       ³                      ³  ÀÄÄÄÄÄÄÄÄÄÄÄÄÁÄÄ¿       ³            ³
       ³                      ³                  ³       ³            ³
       ³                      ³  ÚÄÄÄÄÄÄÄÄÄÄÄÄ¿<ÄÙ       ³ CONTROL    ³
       ³                     ÚÙ  ³ ADDRESS C. ³<ÄÄÄÄÄÄÄÄ>³            ³
       ³                ÚÄÄÄÄÅÄÄ>ÀÄÄÄÄÄÄÄÄÄÄÄÄÁÄÄ¿       ³ UNIT       ³
       ³                ³    À¿                  ³       ³            ³
 ÚÄÄÄÄÄÁÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÁÄ¿   ÀÄ>ÚÄÄÄÄÄÄÄÄÄÄÄÄ¿<ÄÙ       ³            ³
 ³ REGISTER FILE          ÃÄÄÄÄÄ>³ EXECUTE    ³<ÄÄÄÄÄÄÄÄ>³            ³
 ÀÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÙ<ÄÄÄÄÄÁÄÄÄÄÄÄÄÄÄÄÄÄÁÄÄ¿       ³            ³
                                                 ³       ³            ³
                                 ÚÄÄÄÄÄÄÄÄÄÄÄÄ¿<ÄÙ       ³            ³
                                 ³ WRITEBACK  ³<ÄÄÄÄÄÄÄÄ>³            ³
                                 ÀÄÄÄÄÄÄÄÄÄÄÄÄÁÄÄ¿       ³            ³
                                                 ³       ÀÄÄÄÄÄÄÄÄÄÄÄÄÙ
                               MEMORY OUTPUT <ÄÄÄÙ
                              ("memory store" unit)

If you look the drawing, the control unit needs to communicate
with all the other units to make 'em work together.
Usually the "control unit" is scattered in little units that coordinates
intruction and data flow between adiacent units, but you can think
at it as a single indipendent unit.

The five execution units are what gives the "raw" cpu power
but the control unit is what lets you fully utilize 'em.

Let's suppose every unit performs one operation in one step.

With a "cheap" control unit, to execute a complete machine language
instruction you first enable  FETCH, then you enable DECODE
and so on until WRITEBACK is completed and the control unit
enables FETCH again to execute the next instruction.

This is what happens into an "absolutely not pipelined" cpu
with "one cycle" stages, the fastest instructions takes 5 cycles
but the control unit is very simple
(just a circular shift register that enables one unit at a time).

Of course this is the worst case, nobody is so jerk to build
a cpu like that.

Every cpu stage-execution unit is a stand alone thing, it has its hardware
and its "temporary registers" so it is capable to operate "alone".

So, IF THE CONTROL UNIT can SINCHRONIZE the operations
of all the stage-execution units, it can make 'em WORK IN PIPELINE
(like in a factory, where every worker/robot in a production line
 a) receives a partially refined product from the previous worker in line
 b) performs some simple operations to refine the product a little more
 c) pass the result to the next worker in line
).
A cpu with such a control unit is called a pipelined CPU.

If you have to execute in sequence instructions A,B,C,D,E
on a pipelined processor ....

While the WRITEBACK unit is performing stage 5 of A
the EXECUTION         unit can perform stage 4 of B
the ADDRESS CALCULATE unit can perform stage 3 of C
the DECODE            unit can perform stage 2 of D
and the FETCH         unit can perform stage 1 of E

So if you start the cpu at instruction A
it looks like that instruction A takes 5 cycles
while instructions B,C,D,E (immediately one stage behind) looks like they take
JUST ONE CYCLE!!!

So if you execute a SEQUENCE of 8 instructions
on a "not pipelined" processor with a five stage "pipe"
it takes 40 steps to execute the sequence
while on a pipelined processor this takes 5+7=12  steps!!!
( 5 steps for the first instruction to get thru the pipeline
 then 7 steps for the other instructions immediately "one step" behind)
And the more instructions are executed in sequence, the more it looks like
every instruction takes "just" one cycle to execute.


This is of course the optimal situation, when the processor pipeline
is filled and WHEN EVERY INSTRUCTION DOES NOT USE MEMORY/REGISTER OPERANDS
"still under processing" IN SUCCESSIVE EXECUTION-STAGES!!!!!!

-------------------------------------------------------------------------------
THINGS A CPU MUST CHECK TO MAKE A PIPELINE WORK CORRECTLY

A pipelined processor control-unit must "check" :
 A) at stage 1 (FETCH)
    IF in stage 2..4 there are JUMP instructions
             [ they change the program counter
               (the same register used by the fetch unit)
             ]
    THEN stage 1 must "wait" until the jump is completed
         and then "restarts", loading  the memory location pointed by the
         new program counter value.

    Because the jump opcode must pass thru the decode stage ...
    ... if the DECODE unit "detects" a jump opcode, it makes the FETCH unit
    "stops" AT LEAST for two cycles until the jump is performed.
    This of course means the processor pipeline "wastes" some cpu cycles.

    A 486 handles jumps in a smarter way, when the FETCH unit loads
    a jump instruction, it goes on assuming the jump WON'T BE TAKEN
    (it read the instructions following the jump).
    If the jump is taken, the values in the DECODE 1 and DECODE 2 units
    are "discarded" and the FETCH unit starts loading from the new
    program counter value.
    This explains why on 486 relative jumps (the fastest jumps) takes
    3 cycles if taken    (two cycles "lost")
    1 cycle if not taken (no cycles lost)

    A "smarter" control unit can recognize in advance unconditional jumps
    and start immediately loading opcodes from the new program counter
    location
    AND/OR try to "predict" the jump destination of conditional jumps
    (like the Pentium does).


 B) at step 3  (ADDRESS CALCULATION)
    IF in stage 3 a register needed for address calculation is under processing
       in stage 4 (EXECUTE)
    THEN the stage 3 unit generates a "do nothing" control sequence
         for one cpu cycle.

 C) memory access conflicts
    These can be caused by lots of different things, they are usually
    resolved by the memory load/store units.

Told in a short way, in a fully pipelined processor
when the pipeline is "working on a sequence"
the first instruction in the sequence takes the full execution time
while the other instructions takes a shorter time (usually one cycle)
because they are immediately behind, this means they "look faster"
(and in fact they are, because the pipeline fully utilizes all the functional
units it has).

If some instructions "modify" values needed in the previous cpu stages
the control unit stops the unit needing those values (and the others behind it)
and inserts "do nothing" codes into the successive units in the processor pipe
to make things work as expected (this is called a PIPELINE STALL
or a PIPELINE FLUSH)

This explains why usually a taken jump takes more time than one that's not taken
(when a jump is performed you must flush the pipeline).

The newer processors includes JUMP PREDICTION UNITS that handles
jump faster, but the less you jump, the better.

------------------------------------------------------------------------------
PIPELINING IN INTEL PROCESSORS:

The 386 is a "partially pipelined" processor with some optimizations.
It checks for jumps ( A check) in a quite slow way
and most of times can execute every stage in one cycle.
It cannot perform the address generation check (B check)
so it must execute stage 3 and 4 in "not pipelined mode".
This explains why THE FASTEST 386 INSTRUCTIONS TAKE TWO CYCLES
( stage 3 unit must wait stage 4 to complete before managing
  the next instructions).

The 486 has an improved EXECUTION unit (stage 4), improved memory access
AND its control unit can perform the address generation check
so that it can pipe stage 3 and 4 if there are no
register/memory conflicts with two successive instructions.
This explains why lots of 486 instructions take just one cycle
(if the pipeline is filled and no pipeline stalls are produced).
What's more, pipeline reloading is improved, so even jumps have less
impact on execution time.
It also has a 4Kbyte internal cache that boosts memory access
and allows big gains from strip-mining optimizations.
Another nice thing is the BURST READ access that accelerates memory reads
when it needs to update its internal cache.

PIPELINED,SUPERSCALAR & BRANCH PREDITING CPU: THE PENTIUM

Well, what's beyound a fully pipelined processor like the 486 ?
There is a 586/Pentium processor, that's SUPER SCALAR (more than one pipeline)
and BRANCH PREDICTING (it has a specialized unit that annotates the
position and the result of some og the most recent  jumps, and so
it can try to "guess" from where the next instruction after the jump execution
will be loaded, if it guess correcly a pipeline stall is avoided, else
it stalls)(this helps speeding up the execution of nested loops).

The 586 can execute UP TO TWO integer operations in a single step
(because it has TWO pipelines, but with some limitations)
it can BURST READ/WRITE and has TWO indipendent 8k caches
(Harvard style CPU to cache architecture) this means you can
strip-mine to the max. if strips fit into the 8Kbyte data cache.

Well, now you know how they work, let's see how to make 'em work to the max.

------------------------------------------------------------------------------
SPEEDING UP 486/586 CODE.
------------------------------------------------------------------------------

-------------------------------
ADDRESS GENERATION STALLS (AGI)
-------------------------------
An Address Generation Interlock occurs when a register which is
currently being used as the base or an index was the destination component 
of a previous instruction (the stage 3 to 4 pipeline stall i described above)

For example,:

        add edx, 4
        // AGI, ONE CLOCK STALL
        mov esi, [edx]

For the 486, AGI's can only occur on adjacent instructions.

On the 586 or Pentium instructions up to 3 locations away can cause an AGI.
( i.e. an instruction waiting to execute step 4 on pipeline 1
       may need to wait the result produced by an instruction
       still executing on pipeline 2)

                pipeline step              pipe1  pipe2
        addres_calculate/fetch_operand       C      D    |
        execute                              B      A    |
                                                         V

      If C needs the results of A, it must stall pipeline 1 for one cycle
      to wait for A to "exit from" pipeline 2.
      If C needs the results of D, it must stall pipeline 1 for two cycles
      to wait for D to "exit from" pipeline 2.

An example of 3 instruction away AGI:

        add esi, 4
        pop ebx
        dec ebx
        mov edx, [esi]

Takes 4 clocks on a 486.  On a 586, the move command must wait for the 
add command, thus AGI stalling for one clock.  The code above then 
would run in three clock cycles on a 586.

Remember that even instructions that read or write implicitly to registers
cause AGI stalls:

    mov esp,ebp
    pop ebp.

    is executed as...

    mov esp,ebp
    [agi]          ; pop uses the esp as a pointer
    pop ebp

----------------------------
INSTRUCTION DECODING STALLS:
----------------------------

When decoding an instruction with an IMMEDIATE VALUE
AND
either an  INDEX OFFSET or an IMMEDIATE DISPLACEMENT
486's have a one clock penalty. (586's do not)

        Example:

                mov  spam, 248
                mov dword ptr [esp+4], 1

This happens because the decode stage of the pipeline can read and decode
ONE "immediate" value at a time
while an immediate value and an immediate offset are TWO immediate values.
The 586 instead has not such problems (when decoding) (execution is different)

---------------------------------
REGISTER ACCESS CPU STALLS:
---------------------------------

486's have a 1 clock penalty when modifying the lower word of a DWORD register
that's used immediately after it, 586's do not.
        Example:
                mov al,0
                mov [ebp], eax
                (a one clock penalty on 486, no penalty on a 586)


-------------------------------
CODE ALIGNMENT
-------------------------------

To speed up the instructions, alignment of code should be on the
beginning of cache boundaries.
(32 bytes on 586, 16 bytes on 486)

Thus, start your routine on the start of the cache page.
This way, when someone calls the routine, the processor
will just have to load a cache line to get the first sequence to feed
into the pipeline, and while they are executing it will have
enough time to load the other cache lines needed.

This will speed up the processing  of all the instructions within that page.
The effect of this speed up on the 586 is not a dramatic as is it is on the 486
because the 586 has bigger caches and faster access to memory
so a cache line load/store has less impact on cpu.

------------------------------
DATA ALIGNMENT
------------------------------

Misaligned access in the data cache causes an extra 3 cycles on both the 
486 and 586.

Ways to speed up data:
For DWORD data, alignment should be on a 4-byte boundary.
For WORD data, alignment should be on a 2 byte boundary for the 486, 
and simply within the 4-byte page for the 586.
For 8 byte data (64 bits), data should be aligned on a 8-byte boundary
so it is possible to use BURST READS (and burst writes too, on a Pentium).

And yes, on many applications with tight inner loops, these things do 
accumulate to make a noticeable speed-up.

-------------------------------------------------------------------------------
SPEEDING UP REGISTER AND OPERAND USAGE
-------------------------------------------------------------------------------

Use the EAX as much as possible.
Many instructions are 1 byte shorter when using EAX.
That's less code you have to move around and slightly less internal processing.
Use the DS register as much as possible for roughly the same reason,
In the case of several references being made to a variable addressed with a 
displacement, load the displacement into a register.

Try to use the ESP register to reference the stack in lower levels of your 
subroutines (faster than push/pop things and leaves EBP free for other
uses)


-------------------------------------------------------------------------------
HANDY INFO ON SPEEDING UP INTEGER INSTRUCTIONS
-------------------------------------------------------------------------------

1. Avoid using complex instructions like LEAVE, ENTER, LOOP, string 
   instructions etc  Simpler instructions will get the job done faster.
   Simpler instructions have been getting faster with every new processor
   that has come along.

2. With 8-bit operands, do your best to use the byte opcodes, rather than 
   using the 32 bit instructions  with sign and zero extended modes.
   Internal sign extension is expensive, and speed increases can be found by
   simplifying the process as much as possible.

3.  LEA's generally increase the chance of AGI's.  However, LEA's can be
    advantageous because:
    *  In many cases an LEA instruction may be used to replace constant
       multiply instructions. (a sequence of LEA, add and shift for example)
    *  LEA may be used as a three/four operand addition instruction.
       LEA ECX, [EAX+EBX*4+ARRAY_NAME]
    *  Can be advantageous to avoid copying a register when both operands to
       an ADD are being used after the ADD as LEA need not overwrite its
       operands.

    The general rule is that the "generic"

    LEA A,[B+C*INDEX+DISPLACEMENT]

        where A can be a register or a memory location and B,C are registers
        and INDEX=1,2,4,8
        and DISPLACEMENT = 0 ... 4*1024*1024*1024
                           or (if performing signed int operations)
                           -2*1024*1024*1024 ... + (2*1024*1024*1024 -1 )

    replaces the "generic" worst-case sequence

    MOV X,C    ; X is a "dummy" register
    MOV A,B
    MUL X,INDEX    ;actually  SHL X, (log2(INDEX))
    ADD A,DISPLACEMENT
    ADD A,X

    So using LEA you can actually "pack"  up to FIVE instructions into one
    Even counting a "worst case" of TWO OR THREE AGIs caused by the LEA
    this is very fast compared to "normal" code.
    What's more, cpu registers are precious, and using LEA
    you don't need a dummy "X" register to preserve the value of B and C.

4. Zero-Extension with short ints terror tales

  The MOVZX  takes four cycles to execute due to due zero-extension
  wobblies. A better way to load a byte into a register is by:
     xor eax,eax
     mov al,memory
 
  As the xor just clears the top parts of EAX, the xor may be placed on the
  OUTSIDE of a loop that uses just byte values.
  The 586 shows greater response to such actions.

  It is recommended that 16 bit data be accessed with the MOVSX and
  MOVZX if you cannot place the XOR on the outside of the loop.

  N.B. Do the "replacement" only for movsx/zx inside loops.

5. When comparing a value in a register with 0, use the TEST command.
  
  TEST operands by ANDing the operands together without spending any
  internal time worrying about a destination register.
  Use test when comparing the result of a boolean AND command with an
  immediate constant for equality or inequality if the register is EAX.
  You can also use it for zero testing.
  (i.e. test ebx,ebx  sets the zero flag if ebx is zero)

6. Address Calculations

  Pull address calculations into load and store instructions.  Memory
  reference instructions have 4 operands: a relocatable time segment base, a
  base register, a scaled index register and a displacement.  Often several
  integer instructions can be eliminated by fully using the operands of
  memory addresses.  (more on this later)

7.  INTEGER DIVIDE
  In most cases, an Integer Divide is preceded by a CDQ instruction.
  This is as  divide instructions use EDX:EAX as  the dividend and CDQ sets
  up EDX.
  It is better to copy EAX into EDX, then arithmetic-right-shift EDX 31 places
  to sign extend.
  The copy/shift instructions take the same number of clocks as CDQ,
  however, on 586's allows two other instructions to execute at the same
  time.  If you know the value is a positive, use XOR EDX,EDX.

8. INTEGER MULTIPLY
  The integer multiply by an immediate can usually be replaced with a faster
  and simpler series of shifts, subs, adds and lea's.
  As a rule of thumb when 6 or fewer bits are set in the binary representation
  of the constant, it is better to look at other ways of multiplying and not use
  INTEGER MULTIPLY. (the thumb value is 8 on a 586)
  A simple way to do it is to shift and add for each bit set, or use LEA.

  Here the LEA instruction comes in as major cpu booster, for example:
      LEA ECX,[EDX*2]       ; multiply EDX by 2 and store result into ECX
      LEA ECX,[EDX+EDX*2]   ; multiply EDX by 3 and store result into ECX
      LEA ECX,[EDX*4]       ; multiply EDX by 4 and store result into ECX
      LEA ECX,[EDX+EDX*4]   ; multiply EDX by 5 and store result into ECX
      LEA ECX,[EDX*8]       ; multiply EDX by 8 and store result into ECX
      LEA ECX,[EDX+EDX*9]   ; multiply EDX by 9 and store result into ECX

  And you can combine leas too!!!!
      lea ecx,[edx+edx*2]   ;
      lea ecx,[ecx+ecx*8]   ;  ecx <--  edx*27
  (of course, if you can, put three instructions between the two LEA
   so even on Pentiums, no AGIs will be produced).

9. Clearing Registers
  Using   XOR reg,reg is fast but sets up conditions codes.
  A slightly slower way to do it of course is to mov reg,0
  which preserves condition codes.

10. Avoid ENTER, instead, try something like:
  PUSH EBP
  mov ebp, esp
  sub esp, BYTE_COUNT

11. JUMP INSTRUCTIONS
  Jump instruction come in two forms, one real near that jumps between -127
  and 128 of the current location, and a 32 bit version that does a full jump.
  The short form is quite a bit faster, however unfortunately many compilers
  put long jumps where short jumps would suffice.
  To ensure that short jumps can be used (when you know it is possible),
  explicitly specify the destination as being byte length
  (i.e use jmp short instead of plain jmp, if you can)

12. Task Switching
  Task Switching is evil.  It is slow,  real slow.  Avoid task switching too
  often, as more and more of your time will be spent in processing the task
  switch.
  For faster task switching, perform you task switching in software.  This
  allows a smaller processor state to be saved and restored.  Anything that
  shaves time off  75+ clock cycles isn't all that bad in my book.

13.  Minimize segment register loads and the use of far pointers as dearly
  much as you can.  If you are doing a lot of processing on a small piece of
  data far away, it might be faster just to copy the data to nearby and work
  on it there (by the way, this is a good reason to use flat protected mode).

...and....

14. THINK ABOUT WHAT YOU WANT TO DO
  All the other optimisations of Unrolling Loops, moving the invariant data
  etc still apply.  That, however, is more an algorithmic problem for you, but
  still keep it firmly in mind.

_______________________________________________________________________________
586 Specific Optimizations
-------------------------------------------------------------------------------

The 586Pent has a five stage pipeline structure.
However, the pipeline is split into two different pipelines, known 
as Pipeline U and Pipeline V.   This allows two instructions to be
executed in parallel and thus presents the opportunity of 
executing two/instuctions per clock.
The U pipeline is very much like the 486 pipeline, in that it can handle
the entire set of instructions.  Pipeline V on the other hand can
handle only "simple" instructions.
The fast parts of the U and V pipelines are possible as much of the 
functions are hardwired not microcoded.

Anyway, I've blurted on about that for a bit, but what you want to know
is "How to I get two instructions running in one clock/cycle?"

Lament no further, here are the criteria:

  1. Both instructions must be simple (see below)
  2. There must not be a read-after-write or write-after-read
      register/flag dependencies between them.
  3. Neither can have a displacement and an immediate
  4. Instructions with prefixes can only occurr in the U-pipe
     (except for branches on condition jz,jc, etc.etc.)

The following are considered "simple" instructions,
the ALU means any ALU command (ADD etc):

 1. Mov reg, reg/mem/immed
 2. mov mem, reg/imm
 3. ALU reg, reg/mem/immed
 4. ALU mem, reg/immed
 5. inc reg/mem
 6. dec reg/mem
 7. push reg/mem
 8. pop reg
 9. nop
 10. lea reg/mem
 11. Jmp/ Call / Jcc near

N.B Remember only the U pipe can perform SHIFTS/ROTATIONS
    so "couple" every shift/rol with a simple instruction
    before and after to be sure every pipeline is filled.

Note, however, than while both pipelines can perform ALU
instructions concurrently, this is not the case when
one ALU instruction requires the output of the
other pipeline ALU instruction.
Example:  ALU instruction in pipe U and
          performing ADC in pipe V.

There are two exceptions to this rule:
1) In the case of a compare/branch combination.
2) With push/pop combinations (because of optimized stack access)

--------------------------
Branch Prediction
-------------------------

Another nice optimisation with the 586 hardware is that whenever there is 
a possibility of a jump, for example, Jg, the 586 can automatically
start "reading ahead" code from the jump location just in case the the
Jump is accepted.  Remember the CODE alignment thing above?  Well,
it couldn't hurt your grandmother to align the labels on the code pages.

------------------------------------------
Amazing new 586 Optimizations and speedups
------------------------------------------

The 586 has a lot of really nice optimizations and kick-ass
features in addition to what I've mentioned above, unfortunately,
this information is proprietary and will only be given
to people who sign non-disclosure agreements and shell out a
 $$ or two...  Bummer...
(Meethinks, 586 clone-makers won't be too happy about that... :-) )
Compiler/OS makers will probably be the purchasers.
Quite a few of these optimisations won't work on anything less than
a 586, so don't sweat too much about it.  Hey, just write
for 386/486/586 and you'll be ok.

Hovewer, i found a nice article on July 1994 Byte about some
"secret weapons of Pentium coders"....

============================================================================
PENTIUM'S PROFILING REGISTERS:

Pentiums have a set of 64bit "machine specific registers" (MSR)
that are accessed by way of the RDMSR (read MSR) and WRMSR (write MSR)
instructions.

THESE ARE PROTECTED MODE, RING 0 INSTRUCTIONS (CPU LEVEL 0)
( can work in a 386Powered programs only if executing under VCPI
  XMS or "no V86 manager"
  or if you find a way to "toggle" DPMI from ring 3 to ring 0)
(I think they can work even if you boot ms-dos in real mode
 and use the proper instruction prefixes needed to execute
 32bit instructions in real mode).

-----------------------------------------------------------
RDMSR   Read MSR
        Copies the MSR register indexed by ECX
        into the 64bit pair  EDX:EAX

        RDMSR macro
                db 0Fh,032h
        endm
-----------------------------------------------------------
WRMSR   Write MSR
        Copies into the MSR register indexed by ECX
        the value contained into the 64bit pair  EDX:EAX

        Macro to insert it into code if your assembler
        does not support Pentiums:

        WRMSR macro
                db 0Fh,030h
        endm
-----------------------------------------------------------


Intel Pentium user manuals, documents only MSR 0, 1 and 0Eh
and states that MSR 3, 0Fh and above 13h are reserved and illegal.

So, what's left? And what's useful to you ?

MSR 10h is the counter of CPU CYCLES since last power-up.
That value can also be read with the RDTSC (Read time stamp counter)
instruction)
RDTSC is 0Fh,031h in bytes and must be executed while in ring 0 too.

MSR 10h is the most precise timer available on the Pentium
because it "ticks" at the CPU frequency.

Then comes MSR 11h,12h and 13h there you will find the hot stuff

The lower 32bits of MSR 11h are actually two INDEXES
INTO THE CPU PROFILING REGISTERS!!!!

The profiling registers are CPU EVENT TIMERS AND COUNTERS
they can keep track of what's going on inside and outside
your super-duper processor, they can detect nearly everything the CPU does.

The first 16bits of MSR11h selects
the profiling register accessible thru MSR 12h.
The second 16bits of MSR11h selects
the profiling register accessible thru MSR 13h.

Here comes the format of the "profiling register indexes":
bit 0..5 : type of the profiling register to access
bit    6 : Set if you want to monitor the events in cpu ring 0,1,2 (system)
bit    7 : Set if you want to monitor the events in cpu ring 3 (user level)
bit    8 : 0 = access count-of-hardware-events
           1 = access count-of-total-cpu-cycles used to process the cumulated
               events.
           (i'm not sure of this, maybe 0 means count time and 1 count events)
bit 9..15: UNKNOWN, DO NOT MODIFY



Profiling registers types:
INDEX  NAME
0       data write
1       data read
2       data TLB miss (translation look-aside buffer)
3       data read miss
4       data write miss
5       write (hit) to M or E state lines
6       data cache lines written back
7       data cache snoops
8       data cache snoop hits
9       memory accesses in both pipes
0Ah     bank conflict
0Bh     misaligned data memory reference
0Ch     code read
0Dh     code TLB miss
0Eh     code cache miss
0Fh     segment load
10h     ????
11h     ????
12h     branches
13h     BTB hits  (Branch Target Buffer)
14h     taken branch OR BTB hit
15h     pipeline flushes
16h     instructions executed
17h     instructions executed in V-pipe
18h     bus utilization (clocks)
19h     pipeline stalled by write backup
1Ah     pipeline stalled by data memory write
1Bh     pipeline stalled by write to E or M line
1Ch     locked bus cycles
1Dh     i/o read or write cycles
1Eh     non cacheable memory references
1Fh     AGI (Address Generation Interlock)
20h     ????
21h     ????
22h     FPU operations
23h     breakpoint 0 match
24h     breakpoint 1 match
25h     breakpoint 2 match
26h     breakpoint 3 match
27h     hardware interrupts
28h     data read or data write
29h     data read miss or data write miss

So if you want to profile things:
0) Set properly the environment of the test
   (cache empty, cache filled with all the routine code, etc. etc.)
1) Set MSR 11h to the two counters you want to read
2) Read MSR 12h,13h to get the initial values,
3) Execute the code you want to profile
4) Read the final values of MSR 12h,13h
   (without setting MSR 11h again this time).
5) Calculate the difference between the initial and final values.

This way if you are not sure how to optimize things
you can actually test how they work and what's going on.

USING THE PROFILING REGISTER YOU CAN "TEST ON THE ROAD"
YOUR CODE DOWN TO THE LAST CPU CYCLE!!!!!

-------------------------------------------------------------------------------
386 Optimization
-------------------------------------------------------------------------------

Well, nobody buys a 386 now, right?
But still lots of people has one....
So if you wanna be 386 friendly remember .....


--------------------------
DECODE-AFTER-JUMP OVERHEAD
--------------------------

When a jump is performed, a 386 spends some extra time decoding
the next instruction to execute and flushing the pipeline.

THE FIRST INSTRUCTION requires exactly ONE EXTRA CPU CYCLE
for every COMPONENT ( prefixes, opcode, mod/rm , sib ,offset, immediate value )
of the instruction.  Notice i said COMPONENT!!!
An 8bit displacement or a 32bit one, count always as an 1 extra cycle.

So, in 32bit mode (where no prefixes are needed for 32bit instructions):

        loopme: inc esi              ; opcode
                mov [edi+1234h],eax  ; opcode, mod/rm , immediate offset
                dec ecx              ; opcode
                jne short loopme     ; opcode short_displacement

IS FASTER THAN:

        loopme: mov [edi+1234h],eax
                inc esi
                dec ecx
                jne short loopme

Because placing first the three component mov instruction
adds 2 cycles to the instruction execution time, weird, uh?
But if you remember this thing, you can boost 386 code a lot.

By the way, remember that "pipeline overhead" is not so obvious to calculate.
        Look at this:
        add eax,ebx   ; opcode, mod/rm
        add eax,1234h ; opcode, immediate_value
        stosd         ; opcode    <--- these "slow" instruction pipes-in faster
        pop eax       ; opcode    <--- so if you can, put 'em first

------------------
SHORT INSTRUCTIONS
------------------

Short instructions are loaded and decoded faster than longer ones
and since 386 has no internal cache and less memory access bandwidth
than a plain 486, this helps a lot.

Well that's all for a 386.

-------------------------------------------------------------------------------
CACHE OPTIMIZATION TECHNIQUES  (THEY CAN WORK ON ANY CPU!!!!!!!!!!!!)
-------------------------------------------------------------------------------

Usually "code optimization" is cpu-based, but there more things
up in the sky and down on earth than just a cpu... the cache for example!

Well, the difference between a cache hit and a cache miss
means lots of cpu cycles lost, so better hit the cached locations to the max.

386 usually have and external 64k ... 128k cache (mine has none, sigh! :( )
486 have a 4k internal cache and a 64k...512k (usually 256k) second level cache
Pentiums have an Harvard 8k code + 8k data cache
plus a 256k..1Mbyte second level cache.

Use "short" loops so there is more probability that the code to execute
will reside fully into the cache.

Then remember you can count on an external cache of at least 64k
(usually 128k..256k for a 486).

So, if you have to process big arrays or big bitmaps with multiple passages
do not "scan" all the array for every pass, use a "strip by strip"
strategy instead so every "strip" fully fits into the cache
and is ready for the second and third pass without cache misses.

This technique is called STRIP-MINING, you can include into your program
a routine that checks the optimal "strip size" trying a multi-pass test
on 64k,128k,256k,512k strips and "complete array"
and then sets the optimal "strip size" to use
when perfoming "strip-mining" routines.

On a Pentium you can use 8k "strips" so your "strip mined" routines
will work with the full speed of the internal data cache
(remember the internal cache works at full cpu speed while
 the secondary cache may be runnin at half that).

The advantage of strip mining is caused by the fact that
the additional jumping/looping needed to process arrays in "strips"
of adiacent memory locations that can fit together into the cache
is A LOT LESS than the time caused by a single cache miss.

-------------------------------------------------------------------------------
NOT SO OBVIOUS OPTIMIZATIONS
-------------------------------------------------------------------------------

----------------------
COMPLEX INSTRUCTIONS
----------------------
Intel says complex instruction are evil on 486 and up, but there are
exceptions... expecially if want to make things run smooth on 386 too.

STRING instructions are FASTER on 386, expecially if they are the first
in a loop.

If you have to move data in 32bit chunks you'd better use REP MOVSD if you can
because it alone replaces this "simple instructions only"
super-optimized sequence:

        rep_loop:
                mov eax,[esi]
                add esi,4      ; or -4
                mov [edi],eax
                add edi,4      ; or -4
                dec ecx
                jne rep_loop

REP MOVSD takes  2+7*n cycles on a 386 or 486

While the "optimized" sequence uses EAX
and takes [(2+4+2+2+2+2+7)*n - 4 ] = [21*n - 4] cycles on a 386
      and [  (1+1+1+1+1+3)*n - 2 ] = [ 8*n - 2] cycles on a 486

Cache-line aligning the "optimized" code on a Pentium
i think it takes  [ 3*n ]  cycles.
So if 486s are your base system don't throw away REP MOVSD

Also remember that REP MOVSD take 2 bytes instead of 13 bytes
does not need to use EAX (one more register free for other things)
and it does not use the Pentium's BTB (Branch Target Buffer)
so you can be sure it does not miss and the outer loops
with entries in the BTB can be one level deeper.

What's more, the AMD K5 automatically splits a complex instruction
into an optimized sequence of simple instructions
and issues them to fully utilize the four execution units it has.
Guess what it means for poor old movsd :)


Heck! I think those Intel engineers lost a big thing when they
decided to not improve MOVS/LODS/STOS. A "blitter" unit directly
controlled by string instructions (with "fetch ahead","bufferize"
and "auto align" capabilities) could be a great thing for us.

Think about this: a REP MOVSB/W/D "translated" to
a "head" byte move (to align things)
a "central" qword move (burst read/burst writes)
and a "tail" byte move (to align things).
And all this without trashing the processor data cache!!!!
Can you immagine the boost your code may get?
Heck! Sun engineers ADDED "blitter" support in their new 64bit sparc cpus
because they seen their software moving data a lot, why Intel ones did not?

On a Pentium you can get the optimal 2-cycle memory to memory MOV
by way of pipelining, but this cannot take advantage of the fact
that when performing REP MOVS the processor KNOWS AHEAD how many locations
will be moved (read: on a "smarter" cpu this can help optimize to the
max the processor-to-memory bandwidth and avoid caching overhead)
nor it can take advantage of the full power of burst read/writes
neither it can take advantage of the fact that WHILE THE STRING MOVE
IS GOING ON, the processor pipelines can execute OTHER instructions after
the MOVS and "not dealing" with the memory range "under transfert"
or with ECX,ESI,EDI and Direction Flag.
I hope they will take note of this for P7.



-------------------------------------------------------------
THE ADDRESSING MODE ADVANTAGE
-------------------------------------------------------------
Don't be fooled by the current Riscy trend, be cooler than the rest.
Some Intel "complex addressing" modes are really powerful
if you just have enough creativity.
Lets suppose you need to add together data from "four"
streams of data....

A riscy (risc-like) way to do this could be...
                              ; 486 timings when looping
      riscy_way:
               mov eax,[esi]  ; 1
               add esi,4      ; 1
               add eax,[edx]  ; 2
               add edx,4      ; 1
               add eax,[ebx]  ; 2
               add ebx,4      ; 1
               add eax,[ecx]  ; 2
               add ecx,4      ; 1
               mov [edi],eax  ; 1
               add edi,4      ; 1
               dec ebp        ; 1
               jne riscy_way  ; 3
                              ; loop cycles = 17

Now lets see the "intelly" way! :)
Let's suppose the "counter" EBP won't exceed 2^31 -1
we can do the following ...

               ; move pointers ahead ...
               lea esi,[esi+ebp*4]
               lea edx,[edx+ebp*4]
               lea ebx,[ebx+ebp*4]
               lea ecx,[ecx+ebp*4]
               lea edi,[edi+ebp*4]
               neg ebp ; negate increment
               ; then you can fully use the address unit ALU

                                    ; 486 timing when looping
    intelly_way:
               mov eax,[esi+ebp*4]  ; 1
               add eax,[edx+ebp*4]  ; 2
               add eax,[ebx+ebp*4]  ; 2
               add eax,[ecx+ebp*4]  ; 2
               mov [edi+ebp*4],eax  ; 1
               inc ebp              ; 1
               jnz intelly_way      ; 3
                                    ; loop cycles = 12
               

On a Pentium, "riscy" and "intelly" runs at nearly the same speed
BUT ON A 486, the "riscy" way runs 30% slower than "intelly" !!!!

This means that using the "riscy" code
your 486 will look like a slow cow compared to a Pentium
while using the "intelly" code your 486 will look good enough
( not to mention this helps to make the difference between
  "needs a Pentium" and "this flies on 486 too!!").

-------------------------------------------------------------
32bit ACCESS SPEAKS FOR ITSELF, just remember to fully use it
-------------------------------------------------------------

Everywhere you can, use 32bit instructions
and if you can, align data on 32bit.

For example, let's say you have to manipulate lots of strings,
if you align them on dword boundaries (including string lenght)
with null byte paddings, you can boost processing time MORE THAN 8 TIMES!!!!

The same thing applies to manipulating 8bit bitmaps and "software" sound mixing.

Into 386mixer coded a routine that mixed simultaneously 4 sound channels
running only on internal register and mixing up to 4 samples at a time
from the 4 channels (look at the "intelly" example above).

If you need speed, you can even tollerate small calculation errors
or reduced precision
and manipulate successive 8,16bit values in a single 32bit instruction.

Let's say you have and array of unsigned words and you need to multiply them
for a constant, say 45, how can you do that ?
If you know the values won't overflow the 16 bit they use
(if they overflow you will have to choose another method), you can do the
following:
           mov ecx,count
           mov esi,start_address
     handle_two_words:
           mov eax,[esi] ; load two words at a time
           ; an AGI happens, but i can't eliminate it
           lea eax,[eax+eax*4]  ; multiply by 5
           add esi,4  ; increment pointer, avoid 486 AGI
                      ; reduce by 1 the Pentium AGIs
           lea eax,[eax+eax*8]  ; then multiply by 9
           dec ecx  ; decrement counter & keep Pentium at full power
           mov [esi],eax   ;
           jne handle_two_words

And if you have very big arrays, you can even unroll two or three times
the loop to further speed up this code.

------------------
SELF COMPILED CODE
------------------

Sometimes you need to execute lots of times the same "jump filled"
instruction sequence, and you know ahead what jumps will be taken
and what won't.
What's worse if there are lots of conditional jumps (Jcc) "in sequence"
they may be enough to overflow the capability of branch-prediction units!!!

So, what can you do?
My answer to this is a SELF-COMPILER!!!
A small finite state machine that instead of crunching data directly
IT GENERATES SUPER-SPECIALIZED ROUTINES that will crunch the data in the
fastest way the code generator knows.

I've done a thing like that for the _PageFLip1 routine in 386video.asm.
At mode initialization a code generator generates
all the 256 blit-skip sequences the _PageFLip1 routines may need
when copying "modified" pixels on screen.

This way a call is performed only after 32 blitted pixels, instead of jumping
every 2..4 pixels (or every pixels in the worst case situations).

By the way, this is NOT self-modifying code
this is an "integrated code compiler".

------------------------------------------------------------------------------
I hope this information has been helpful for you.
Now make some coffee, brush your teeth and phone up your girlfriend and 
go and see a movie.   This document will be here when you get back, and I 
imagine there is only so much of this you can take in one day.... :-)  :-)  :-)

Live long and code well...

Regards,

   Michael Kunstelj.

Regards from me too,

   Lorenzo Micheletto


Discuss this article in the forums


Date this article was posted to GameDev.net: 10/25/1999
(Note that this date does not necessarily correspond to the date the article was written)

See Also:
x86 Assembly

© 1999-2011 Gamedev.net. All rights reserved. Terms of Use Privacy Policy
Comments? Questions? Feedback? Click here!