Universal global regs 3

[iluuu1994]

No description provided.

[dstogov]

I didn't completely understand what problem do you like to solve.

In case you like to eliminate VM code that moves CPU register used for opline to/from EX(opline), it's better to move some functions into zend_execute.c and make them responsible to that movements (e.g. zend_throw_...()) .

Completely disabling some CPU register for the whole binary looks too aggressive and useless.

[iluuu1994]

@dstogov Sorry, I should have included some more context. To clarify a bit, the idea comes from this comment: php#12758 (comment)

There is a known issue where EX(opline) may be outdated on OOM errors where the current opline->lineno is emitted. Worse, if the first opline is the one that OOMs, then EX(opline) may be completely uninitialized and cause a segfault. There were some recent attempts to fix some or all of them.

• fix segfault in ZEND_BIND_STATIC php/php-src#12758
• fix segfault in ZEND_FUNC_GET_ARGS php/php-src#12768
• SAVE_OPLINE before allocating memory php/php-src#12648
• [wip] Add missing SAVE_OPLINE() php/php-src#10004

They involve adding SAVE_OPLINE calls to basically all handlers. With most handlers doing SAVE_OPLINE, global registers become less useful. I understand they still avoid passing opline to non-inlined helper functions. The motivation of this PR was to dodge SAVE_OPLINE by making the IP register reserved universally. Also note that GCC does use the fixed register as long as it can deduce that it is not accessed or written to during the execution of the compiled function. This does seem profitable according to this PoC, just not by as much as I had hoped.

That said, I do think the gain is too small to fragment extensions in another way (like nts/zts).