Codegen: every boxed assignment open-codes a 15-block type switch with duplicated error blocks — 111k-line IR module and 51 KB helper functions for a 4-line script (lib/JIT/JitValueBox.php)

[PurHur]

Category

Foundation: · generated-code size / compile time · child of #36188

Problem

JitValueBox::copyBetweenPointers() (lib/JIT/JitValueBox.php:623-800) appends value_copy_{string,hashtable,object,long,double,bool,null,done} plus seven after_* blocks — 15 basic blocks, tagged by a static counter so nothing is shared — at every boxed assignment site (176 copy* call sites in lib/). The same pattern repeats for error paths: each op gets its own __compiler_jit_raise_type_error + abort block.

Measured on the committed IR dump build/micro/m_loop.ll (4-line typed loop, 111,127 IR lines, 485 defined functions):

function	IR lines	of which
PHPCompiler_ext_standard_IniJitHelper__inicfgget	16,508	1,176 __value__valueDelref, 759 __compiler_jit_raise_type_error + abort pairs, 167 __string__separate
…IniJitHelper__iniget	12,187	(51 KB of .text in the linked binary)
…PendingHeadersJitHelper__statuscodefromhttpversionline	—	51 KB of .text for a status-line parser
user function loop	40

The user program is 0.04 % of the module. Everything else is helper corpus re-lowered per unit, and its size is set by this open-coding. It costs LLVM verify/codegen time on every build, unit.o size (see the helper-unit duplication issue), and makes any optimisation pass 25x more expensive than the program warrants.

php-src reference

• Zend/zend_variables.h — ZVAL_COPY / ZVAL_COPY_VALUE are one inline function each; the type switch lives in zval_add_ref, not at every call site.

PHP implementation target

• Emit the body of copyBetweenPointers once per module as __value__copy(%__value__* dst, %__value__* src) (and __value__assign = delref old + copy) with alwaysinline so that the minimal pass pipeline (sibling issue) can still inline where profitable; call sites become one call.
• Outline the cold error paths: one __jit_type_error_abort(msg, code) noreturn function per module; call sites br to a single shared cold block per function.
• Same treatment for JitValueNumeric::emitBoxedNumericResult (lib/JIT/JitValueNumeric.php:373-448): load each operand's tag once; valueIsDouble/valueIsString/valueBoxToDouble re-load it 4x today.
• Measure with PHP_COMPILER_DUMP_IR=1 (writes /tmp/phpc-last.ll): IR lines, define count, and .text of the hello-world binary before/after.

Repro

./script/docker-exec.sh -- bash -lc 'source script/php-env.sh && printf "<?php echo 1;\n" > build/one.php \
  && PHP_COMPILER_DUMP_IR=1 php bin/compile.php -o build/one build/one.php && wc -l /tmp/phpc-last.ll \
  && grep -c "value_copy_string_" /tmp/phpc-last.ll && size build/one'

Done when

• IR line count of the echo 1 module drops ≥ 40 %; value_copy_string_ labels ≤ 1 per module
• Hello-world .text (size) drops accordingly; committed helper-unit cache refreshed once (make helper-runtime-prelink-refresh) and its total size recorded in the PR
• script/aot-smoke.sh 8/8; script/differential-sweep.sh --aot --repeat 3 failing-name set unchanged; PHP_COMPILER_OPT_LEVEL=2 build of Ack(3,9) still 4093

[PurHur]

claim: implementer-worker-lane-b — starting this run (lane B tie-break #36193 % 3 == 1; lane overflow from closed #36143–#36148)