Turbo 2.3.0 incompatible with Tideways #15407

Activity

1. DuckThom commented on Oct 13, 2026

   

   DuckThom

   on Oct 13, 2026

   Author

   More actions

   I have done some more digging by adding debug code to <turbo-ext>/src/TypeCombinator.cpp as that is where the segmentation faults are originating from.

   The segfault is actually triggered by this line in Composer: https://github.com/composer/composer/blob/main/src/Composer/PHPStan/ConfigReturnTypeExtension.php#L185

   Adding a var_dump right before it reveals the following value for $types:

   array(2) {
     [0]=>
     object(PHPStan\Type\StringType)#1465 (0) {
     }
     [1]=>
     object(PHPStan\Type\Constant\ConstantBooleanType)#1464 (1) {
       ["value":"PHPStan\Type\Constant\ConstantBooleanType":private]=>
       bool(false)
     }
   }
   

   However, inside the turbo extension, with the following debug lines added to turbo-ext/src/TypeCombinator.cpp:4353:

           cls.method(sigs::union_, [](INTERNAL_FUNCTION_PARAMETERS) {
                   zval *types;
                   uint32_t count;
   
                   ZEND_PARSE_PARAMETERS_START(0, -1)
                           Z_PARAM_VARIADIC('*', types, count)
                   ZEND_PARSE_PARAMETERS_END();
   
                   fprintf(stderr, "\n--- [STACK INSPECTION] ---\n");
                   fprintf(stderr, "execute_data:      %p\n", (void*)execute_data);
                   fprintf(stderr, "prev_execute_data: %p\n", (void*)execute_data->prev_execute_data);
                   fprintf(stderr, "ZEND_CALL_ARG 1:   %p\n", (void*)ZEND_CALL_ARG(execute_data, 1));
                   fprintf(stderr, "ZEND_CALL_ARG 2:   %p\n", (void*)ZEND_CALL_ARG(execute_data, 2));
                   fprintf(stderr, "types pointer:     %p (count=%u)\n", (void*)types, count);
   
                   // Dump 6 zval slots around types to locate #1464
                   for (int k = -1; k <= (int)count + 1; k++) {
                           zval *zv = &types[k];
                           fprintf(stderr, "  slot[%2d] at %p: type=%u", k, (void*)zv, Z_TYPE_P(zv));
                           if (Z_TYPE_P(zv) == IS_OBJECT) {
                                   zend_object *obj = Z_OBJ_P(zv);
                                   fprintf(stderr, " (OBJECT obj=%p, handle=%u)", (void*)obj, obj ? obj->handle : 0);
                           }
                           fprintf(stderr, "\n");
                   }
                   fprintf(stderr, "--------------------------\n\n");
   
                   if (UNEXPECTED(!requireTypeArguments(count, types))) RETURN_THROWS();
                   PT_RETURN_VAL(TypeCombinator::union_(count, types));
           });

   The following is echoed:

   --- [STACK INSPECTION] ---
   execute_data:      0x7fa4a7c16410
   prev_execute_data: 0x7fa4a7c162b0
   ZEND_CALL_ARG 1:   0x7fa4a7c16460
   ZEND_CALL_ARG 2:   0x7fa4a7c16470
   types pointer:     0x7fa4a7c16460 (count=2)
     slot[-1] at 0x7fa4a7c16450: type=0
     slot[ 0] at 0x7fa4a7c16460: type=8 (OBJECT obj=0x7fa4a373c200, handle=1465)
     slot[ 1] at 0x7fa4a7c16470: type=8 (OBJECT obj=0x7fa4a7c162b0, handle=0)
     slot[ 2] at 0x7fa4a7c16480: type=4
     slot[ 3] at 0x7fa4a7c16490: type=4
   --------------------------
   

   Here, the object pointer for object(PHPStan\Type\Constant\ConstantBooleanType)#1464 is suddenly replaced with the prev_execute_data pointer. Presumably because Tideways is also wrapping internal function calls, messing with the stack.

   While I am not sure if this conflict is easily fixable, I do think the turbo extension might need an extra guard against this so it doesn't completely crash the php process if the object name cannot be resolved.
2. staabm commented on Oct 13, 2026

   

   staabm

   on Oct 13, 2026

   Contributor

   More actions

   phpstan is already doing a auto-restart when xdebug is present. maybe we should just restart for tideways extension as well
3. ondrejmirtes commented on Oct 13, 2026

   

   ondrejmirtes

   on Oct 13, 2026

   Member

   More actions

   Yes, and pcov 😊
4. staabm

   added

   bug

   on Oct 14, 2026
5. staabm

   closed this as completedin phpstan-src#6757on Oct 17, 2026
6. staabm commented on Oct 22, 2026

   

   staabm

   on Oct 22, 2026

   Contributor

   More actions

   //cc @TimWolla just fyi, in case there is room for improvement on the tideways end