incorrect "optimization"

Index > Compiler Internals > incorrect "optimization"

Author

Thread

Japheth

Joined: 26 Oct 2004
Posts: 151

Japheth 18 Oct 2012, 04:59

There was a brief "discussion" in masm32 forum about a probably incorrect optimization in fasm. The testcase is:

Code:


    format mz

    entry _text:start
    stack 400h

segment _text

start:
    mov ax,_data
    mov ds,ax
    mov ebp,1
    mov dx,[table+ebp*2]
    mov ah,9
    int 21h
    mov ah,4ch
    int 21h

segment _data

string0 db "string0",13,10,'$'
string1 db "string1",13,10,'$'
string2 db "string2",13,10,'$'

table dw string0, string1, string2

expected result is "string1" to be displayed, but this isn't the case with recent fasm.

18 Oct 2012, 04:59

JohnFound

Joined: 16 Jun 2003
Posts: 3499
Location: Bulgaria

JohnFound 18 Oct 2012, 05:53

ebp works by default on ss, not on ds. This is not because of FASM optimizations.

18 Oct 2012, 05:53

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20309
Location: In your JS exploiting you and your system

revolution 18 Oct 2012, 07:00

fasm does not track the contents of the segment registers. fasm does not know that the SS register is not initialised to point to the data segment. Use this:

Code:

mov dx,[ds:table+ebp*2]

18 Oct 2012, 07:00

sinsi

Joined: 10 Aug 2007
Posts: 789
Location: Adelaide

sinsi 18 Oct 2012, 07:43

The original "problem" was that FASM will take LEA EAX,[6+EBX*2] and change it to the shorter version of LEA EAX,[6+EBX+EBX].
Fine by me, since it works out the same. Then there was a gotcha about using EBP.

MASM will encode LEA EAX,[6+EBP+EBP] with a DS override, FASM doesn't.
I've looked briefly through the intel docs but can't find anything. Anyone?

18 Oct 2012, 07:43

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20309
Location: In your JS exploiting you and your system

revolution 18 Oct 2012, 07:55

sinsi wrote:

The original "problem" was that FASM will take LEA EAX,[6+EBX*2] and change it to the shorter version of LEA EAX,[6+EBX+EBX].
Fine by me, since it works out the same. Then there was a gotcha about using EBP.

MASM will encode LEA EAX,[6+EBP+EBP] with a DS override, FASM doesn't.
I've looked briefly through the intel docs but can't find anything. Anyone?

LEA does not use any segment register, it doesn't access memory at all.

18 Oct 2012, 07:55

JohnFound

Joined: 16 Jun 2003
Posts: 3499
Location: Bulgaria

JohnFound 18 Oct 2012, 08:00

AFAIK, EBP based addressing always works on SS by default. That is because EBP is designed to be used for stack access.

So, [2*ebp+6] or [ebp+ebp+6], both should work on SS.

P.S. I mean in memory accessing instructions. LEA of course will not use any segments at all.

18 Oct 2012, 08:00

Japheth

Joined: 26 Oct 2004
Posts: 151

Japheth 18 Oct 2012, 08:12

I quite intentionally didn't provide a link to the masm32 forum thread because it's mainly about LEA and LEA is ABSOLUTELY irrelevant for this issue.

Masm and assumed behavior of Masm is also irrelevant. This problem is not because Fasm doesn't ASSUME segment register contents.

I agree that adding a DS segment prefix makes the test case "work" - but this is not the point. The point is that the test case is to work WITHOUT any segment prefix.

18 Oct 2012, 08:12

JohnFound

Joined: 16 Jun 2003
Posts: 3499
Location: Bulgaria

JohnFound 18 Oct 2012, 08:21

Japheth wrote:

The point is that the test case is to work WITHOUT any segment prefix.

Explain this more detailed. Why it should work? IMHO, in both cases [ebp+ebp+const] or [ebp*2+const] the CPU will try to access SS.

_________________
Tox ID: 48C0321ADDB2FE5F644BB5E3D58B0D58C35E5BCBC81D7CD333633FEDF1047914A534256478D9

18 Oct 2012, 08:21

sinsi

Joined: 10 Aug 2007
Posts: 789
Location: Adelaide

sinsi 18 Oct 2012, 08:24

>MASM will encode LEA EAX,[6+EBP+EBP] with a DS override, FASM doesn't.
That should be MOV, not LEA. Sorry.

>Masm and assumed behavior of Masm is also irrelevant.
Relevant, because FASM does differently.

Japheth, do you know if it says anything in the intel docs?
Also, what does jwasm do?

18 Oct 2012, 08:24

Japheth

Joined: 26 Oct 2004
Posts: 151

Japheth 18 Oct 2012, 08:33

JohnFound wrote:

IMHO, in both cases [ebp+ebp+const] or [ebp*2+const] the CPU will try to access SS.

That's the point. SS is only the default segment register if the base register is EBP (or ESP or BP). The index register doesn't matter.

18 Oct 2012, 08:33

JohnFound

Joined: 16 Jun 2003
Posts: 3499
Location: Bulgaria

JohnFound 18 Oct 2012, 08:49

Ah, I got it now, after some tests ( Smile

)
It is a bug and has to be fixed. Tomasz probably already thinks about it.

18 Oct 2012, 08:49

Tomasz Grysztar

Joined: 16 Jun 2003
Posts: 8351
Location: Kraków, Poland

Tomasz Grysztar 18 Oct 2012, 09:20

No, it is not a bug. It was already discussed on this board, see my stance here: http://board.flatassembler.net/topic.php?p=116138#116138 (starting with "As for the EBP*2 problem" words).
In short: by default fasm optimizes as for flat mode, where segments do not matter (and so addresses containing EBP can be most agressively optimized).
When working in segmented environment and unsure about the base segment that will get used, just use segment prefix in the address expression and fasm will happily optimize instruction that does exactly what it is told to.

You can also find some related release notes about rare optimization cases here: http://board.flatassembler.net/topic.php?p=118691#118691

18 Oct 2012, 09:20

JohnFound

Joined: 16 Jun 2003
Posts: 3499
Location: Bulgaria

JohnFound 18 Oct 2012, 09:31

I missed this discussion. Embarassed

But your arguments looks reasonable.

18 Oct 2012, 09:31

Japheth

Joined: 26 Oct 2004
Posts: 151

Japheth 18 Oct 2012, 12:52

JohnFound wrote:

I missed this discussion. But your arguments looks reasonable.

I also missed this. Personally, I find the argumentation a bit "borderline" ( somewhat similiar to the SSSO "arguments" ), but never mind!

18 Oct 2012, 12:52

shutdownall

Joined: 02 Apr 2010
Posts: 517
Location: Munich

shutdownall 18 Oct 2012, 19:46

Tomasz Grysztar wrote:

When working in segmented environment and unsure about the base segment that will get used, just use segment prefix in the address expression and fasm will happily optimize instruction that does exactly what it is told to.

Or simply use EDI instead of EBP because EDI has DS as default segment.

18 Oct 2012, 19:46

< Last Thread | Next Thread >

Forum Rules:

You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum
You cannot attach files in this forum
You can download files in this forum