🚀 go-pugleaf

In <1992Dec24.070621.2821@sol.ctr.columbia.edu> penev@venezia (Penio Penev) writes:

| Does any one know a reason, why a program, runnng on R3000 would not
| run on R4000?
| 
| I'm using only ALU, memory and branch instructions. No coprocessor 
| (0 or 1), not mul/dev.
| 
| The program runs on R3000 (IRIX 4.0.1), but fails on R4000 (4.0.4).
| The error is reproducible, occures allways in the same region of the
| program, and behaves identically on successive runs. It seems to me
| like a jump to a bad address, resulting from a bad value in the $ra
| register.The problem is that whenever I put brakepoints "close enough"
| and do noting, but cont in them, it disappers.
|
| I have two explanations: 1) I do not fill the load delay slot
| properly, which results in jump to garbage. 2) I have made a mistake,
| but such a one, that the R3000 hazard (un)handling corrects it.

You may be running afoul of a bug in the rev 2.2 part of the r4k, if your
program is either some form of interpreter, or is otherwise executing
out of the data segment.  To work around that bug, branch and jump
instructions must be double word aligned (low 3 bits of address 0).

Your last paragraph above makes it sound like you are generating
instruction streams without the compiler.

The compilers normally work around this, and the kernel has a workaround
for it for pages in the instruction segment (of user programs).

This is fixed in rev 3 of the chip (which we aren't shipping yet, and
may not ship, depending on availability of the r4400, and cost of
qualifying a new chip rev).  It is also fixed in the r4400.

Programs like interpreters have fallen afoul of this problem.
It is documented in the IRIX 4.0.3 and later release notes.
--
Let no one tell me that silence gives consent,  |   Dave Olson
because whoever is silent dissents.             |   Silicon Graphics, Inc.
    Maria Isabel Barreno                        |   olson@sgi.com