PHP's two-pass compiler
On my way to Istanbul I was looking at Xdebug bug #422 . For some reason Xdebug was crashing while doing code-coverage analysis, in the part that analyses which code was dead (ie. opcodes that could never be reached). The crash occurred with a JMPZ (jump-if-zero) instruction, that suddenly saw a jump-to position of 572222864. That position resembles more a jump-address.
Xdebug uses the same branch analysis implementation as VLD so I used the latter tool to find out why it would crash. Unfortunately, it was working just all nice and fine with VLD. After digging around some more, I saw from the back trace that the crash in Xdebug only occurred when a user-defined error-handler was called while parsing a file. The latter gave me the insight of looking at which phase the compiler was in. I remembered that PHP has a two phase compiler. The first pass is quick and dirty, and only records the opcode line number to jump to. Xdebug however was expecting an memory address as jump target. Because a memory address is a much larger number than an opcode number—the latter usually not being much higher than a thousand—Xdebug was setting the "visited" flag in a part of memory that wasn't allocated. And writing to unallocated memory makes a process die with a segmentation fault.
The compiler in PHP is two-pass. During the first pass, it will find out to which opcode it needs to jump in the jump instructions. However, the PHP engine (and Xdebug) expects a memory address to jump to while executing your script. In the second pass, the compiler will then go over the generated opcodes and calculate the memory address to jump to from the jumps to opcode numbers. It will also do a few other things, such as collapsing sequential EXT_STMT opcodes, calling Zend extension's functions to finalize the opcode arrays—Xdebug uses this for caching whether an opcode array has been scanned already—and re-allocating the opcode array itself to save space.
Now, the thing is, that usually VLD and Xdebug kick in after the whole opcode array has been created, which includes running the second pass of the compiler. However, Xdebug also tries to analyze opcode arrays while executing them. In the case of a user defined error handler, that happens before the second pass has been run. Preventing the crash was therefore as easy as making sure that the compiler's second pass had been run while scanning the opcode arrays for executable code.
Life Line
I know my French is pretty terrible, but I'm sure I'm closer to the correct answer than what's shown here...
Merge branch 'v2022'
Merge pull request #169 from psumbera/solaris-2
I walked 7.0km in 1h6m48s
Fixed some ffing sidewalks again.
I walked 10.5km in 1h40m26s
Updated a pet_grooming shop
I walked 8.6km in 2h12m58s
I walked 8.7km in 1h24m16s
Updated a restaurant
I walked 2.4km in 24m20s
I walked 6.6km in 1h4m32s
I walked 0.6km in 4m38s
I walked 8.5km in 1h22m35s
Merged pull request #1029
Reflow some comments
Add comments, add end of file newlines, fix php 8.5 compilation
Benchmark Xdebug performance
Merged pull request #1051
PHP 8.6: printf() is now optimised out if it only uses %s and %d (and…
PHP 8.6: The object IDs of objects changed in tests
PHP 8.6: ZSTR_INIT_LITERAL() no longer takes non-literals, so use zen…
PHP 8.6: WRONG_PARAM_COUNT has been removed in favour of zend_wrong_p…
PHP 8.6: zval_dtor() has been deprecated in favour of zval_ptr_dtor_n…
Update test for version constraints, as well as the error messages


Shortlink
This article has a short URL available: https://drck.me/ptc-6za