[patch] Fix PR middle-end/78468

[Eric Botcazou]

> I'd say let's not have a middle ground - this stuff is sufficiently
> brain-twisting that I'd rather go back to a known working state. If
> there was an error in the previous patch, let's roll it back until we
> fully understand the situation.

In fact it's not exactly a middle ground; it's already almost a roll back 
since only PowerPC, PA-RISC and s390 define STACK_DYNAMIC_OFFSET.

PowerPC and s390 are OK according to their maintainers, so we're left with 
only PA-RISC.  IIUC the question boils down to whether STACK_DYNAMIC_OFFSET 
garantees that (stack pointer + STACK_DYNAMIC_OFFSET) is always aligned to 
STACK_BOUNDARY, which is 2 * BITS_PER_WORD on PA-RISC.

Dave, can you give the answer here?

-- 
Eric Botcazou