issue-#209 Compiler trap in Real64 op ADR(x)

[luowy]

a compiler bug, adimetrius have made a fix.
https://community.blackboxframework.org ... f=48&t=258

[Robert]

For the issue see : https://redmine.blackboxframework.org/issues/209 .

[luowy]

update another choice https://community.blackboxframework.org ... 1665#p1665

[Josef Templ]

for a bug fix as proposed by adimetrius see https://redmine.blackboxframework.org/p ... 9e59a688be

- Josef

[Robert]

luowy - Why have you voted against this - Is there something wrong with the solution?

[luowy]

well, there are two points;
1, DevCPL486.CheckMemForm used by Float load or store, the operands it accepted is only integer and real form;

Code: Select all

FILD – Load Integer
	16-bit memory     	1101 1111 : mod 000 r/m	
	    FILD m16int	DF         /0  	Push m16int onto the FPU register stack.
	32-bit memory    	1101 1111 : mod 101 r/m	
	    FILD m32int 	DB        /0 	Push m32int onto the FPU register stack.
	64-bit memory      	1101 1011 : mod 000 r/m	
	    FILD m64int 	DF        /5 	Push m64int onto the FPU register stack.

FLD – Load Real
	32-bit memory 	1101 1001 : mod 000 r/m
	    FLD m32fp 	D9       /0	Push m32fp onto the FPU register stack.
	64-bit memory 	1101 1101 : mod 000 r/m
	    FLD m64fp 	DD       /0	Push m64fp onto the FPU register stack.
	80-bit memory 	1101 1011 : mod 101 r/m
	    FLD m80fp 	DB       /5	Push m80fp onto the FPU register stack.
	ST(i) 	1101 1001 : 11 000 ST(i)
	    FLD ST(i)	D9      C0+i	Push ST(i) onto the FPU register stack.

obviously, the bug comes from the front step: the Pointer form should not come to here;


2, Pointer is a unsigned integer actually, e.g.
r := r0 + S.ADR(x); Post should : ASSERT(r >= r0)
this fix output such code:

Code: Select all

8D 45 F8 	LEA	   EAX, [EBP-8]	
50 	       PUSH	EAX	
DB 04 24 	FILD	  [ESP]	
83 C4 04 	ADD	   ESP, 4

FILD tread the operand as a signed integer,
if ADR(X) >2G, the resoult will be r< r0;
P.S. because cp language has no unsigned integer, this requirement may be inappropriate,It will break consistency:
e.g. y:=ADR(x); r:=r0+y; VS r:=r0 + ADR(x)

[Josef Templ]

This is a very good argument, luowy. We will have to look into this issue again.
In the next week I am on vacation and will probably not be able to work on this.

But the voting should be suspended, I think.

- Josef

[Josef Templ]

I have looked at this issue again.
My conclusion so far is that the current solution is not as bad as I thought.

@luowy
1. the alternative proposals for the fix are also in the backend, as far as I understand them. So this argument does not really apply.

2. The "signed"-argument does also not really apply either because it is the very nature of CP that it does not have unsigned numeric types.
Introducing the new concept of an "unsigned" address conversion for this very exotic case does not feel right to me.

@all
For checking the abstract syntax tree constructed by the frontend and passed to the backend it would be
good to have a tool that dumps the tree to a window in readable textual form (with enum constants expanded into names).
Otherwise it is very tedious to debug such issues.
Is there any such tool around?

- Josef

[luowy]

Is there any such tool around?

I have no such tool, I used only three papers, copied all enum names with a pencil, and the builtin post debugger(trap window), HALT(100) at the position which I want to inspect!
sometimes spend only a minute if the bug near the trap, sometimes will need several hours to find the solution if the bug far away from it...

[Josef Templ]

I would like to come back to the other questions resp. comments

@luowy
1. the alternative proposals for the fix are also in the backend, as far as I understand them. So this argument does not really apply.

2. The "signed"-argument does also not really apply either because it is the very nature of CP that it does not have unsigned numeric types.
Introducing the new concept of an "unsigned" address conversion for this very exotic case does not feel right to me.

- Josef