I have retargeted LCC for my own custom designed CPU. LCC actually works quite well and I have the basis of an OS running; all compiling correctly.

Today I compiled the Dhrystone benchmark code and suddenly LCC does something very strange: it is screwing up its register allocation; overwriting values by using the same register where it shouldn't. Look at the example below:

ldw r4,-2(bp)
ldi r3, -84
add r3, r3, bp
stw r0(r3),r4

ldw r3,-4(bp)
ldi r3, -82
add r3, r3, bp
stw r0(r3),r3

MULT16(r1,r3,r4)

The code is setting up 2 parameters for a multiplication. In the first block it is loading a local var into R4 and then storing it somewhere using R3 (i clobber all registers when i call a pseudo op like MULT).

However, in the second block it is loading a local var in R3, then immediately overwriting that register for the address calculation.

The relevant code generation rules are:

reg: ADDRLP2 "\tld16\t%c, %a\n\tadd\t%c, %c, bp\n" 2
stmt: ASGNU2(reg,reg) "\tstw\tr0(%0),%1\n" 2
reg: MULU2(reg,reg) "MULT16(%c,%0,%1)\n" 1

the clobber function has this:

case MUL+U:
spill( (1<<R2) | (1<<R3) | (1<<R4), IREG, p);
break;

I don't really know where to look for a problem. The compiler has been behaving splendidly and I never had trouble with register allocations.

Any pointers or tips greatly appreciated!
Erwin

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

Can you send me the C source that produces this bug?

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)