Re: [alpha] cleanup opDEC workaround

[Richard Henderson <rth@redhat.com>]

On Tue, Nov 20, 2001 at 08:51:05PM +0300, Ivan Kokshaysky wrote:
> Ok, but with opDEC_fix = 8 we actually skip to addt/stt, so that asm
> probably should be rearranged to ...

No, read the comment and think about it some more.

 		"fmov $f31, $f0\n\t"
 		"cvttq/svm $f31, $f31\n\t"
 		"cmpteq $f31, $f31, $f0\n\t"
 		"addt $f31, $f31, $f31\n\t"
 		"stt $f0, %0"

In the broken case, we'll enter entIF with pc==cvttq.  Add 8 gives
addt, subtract 4 gives cmpteq, which gives $f0 = 2.0.

In the working case, we'll enter entIF with pc==cmpteq.  Add 8 gives
stt, subtract 4 gives addt, which is engineered to be a noop.

> Further, if fp emulation isn't compiled in, we'll have kernel mode
> instruction fault. A quick fix appears to be

Hmm.  If fp emulation isn't compiled in, we shouldn't bother
testing this, I think.  Means you can't debug fp emulation via
modules on Multia, but I'm pretty sure I don't care.

I suppose the other alternative to get the testing code out of
the normal entIF is to create a custom entIF that is installed
only during opDEC testing.  Seems too much work...


r~