`asm volatile ("":::"memory")`

Found compiling busybox 1.37.0, with [[bug-c-logical-not-is-not-folded-in-a-constant-expression]] and [[bug-c-a-ternary-cannot-be-the-callee-of-a-call]].

What was there: nothing

written	before
`asm("")`	`error: call to undeclared function: asm`
`__asm__ volatile ("")`	`error: call to undeclared function: __asm__`
`asm volatile ("":::"memory")`	`Expected: ), but got: (Kind: 79)` — the first `:`
`asm volatile ("mov %0,%0" : "+r"(x))`	same

asm was an ordinary identifier, so it parsed as a call.

busybox reaches it from include/libbb.h:

#define barrier()             asm volatile ("":::"memory")
#define ASSIGN_CONST_PTR(pptr, v) do { *(void**)(pptr) = (void*)(v); barrier(); } while (0)
#define SET_PTR_TO_GLOBALS(x) ASSIGN_CONST_PTR(&ptr_to_globals, x)

INIT_G() / INIT_S() expand that, and applets call it as their function's first statement — so coreutils/test.c, editors/ed.c and util-linux/acpid.c each died immediately, at line 926, 1006 and 266.

What landed, and the line it draws

An empty template with no output operands is a compiler barrier: it orders the compiler, not the machine. pxx does not reorder across a statement boundary, so compiling it to nothing is not an approximation — it is what the construct means. That form is accepted, in every spelling:

asm(""); asm volatile (""); __asm__ volatile (""); __asm ("");
asm volatile ("":::"memory");     asm volatile ("" ::: "memory");
asm volatile ("" : : : "memory"); asm volatile ("" "" ::: "memory");

Everything else is refused, by name. Parsing real instructions and dropping them would silently miscompile precisely the code that cares most about what the machine does — the incomplete step reporting in the vocabulary of a complete one class this project treats as its most expensive:

error: C: inline asm with a non-empty template is not supported
       — the instructions would be silently dropped
error: C: inline asm with an output operand list is not supported
       — nothing would ever be written to it

The second exists because an empty template with an OUTPUT section still claims to produce a value, which nothing wrote. ":::" (clobbers only) carries no such claim, which is why the section COUNT distinguishes them rather than the template alone.

This is deliberately not an implementation of extended asm. Operand constraints, %0 substitution, register allocation and clobber honouring are a real feature; this is the honest boundary in front of it, and the diagnostic names which part is missing so a reader is not left inferring it from a parse error.

Measured


`coreutils/test.c`	refused → compiles
`editors/ed.c`	refused → compiles
`util-linux/acpid.c`	refused → compiles
`run_c_conformance.sh`	220 pass / 0 fail — baseline
`run_pascal_conformance.sh`	346 pass / 0 fail — baseline
self-host	converged after 1 round
`gate.sh quick`	GREEN

busybox, cumulative for the three C fixes today

Restricted to the 286 files gcc actually compiles in libbb + coreutils + editors + util-linux — an earlier count of 307 included include-fragments like pw_encrypt_des.c that the real build never compiles standalone, so it inflated the failure list:

	files compiling clean
session start	174
after the `!` / negative-bound fix	176
after the ternary callee	176 (+ `stat.c`, − none)
after inline asm	178

and the residue is now sharply separated:

remaining failure	count
missing crtl function — a LIBRARY gap	100
real compiler gaps	8

The eight, which are the whole remaining compiler work list:

file
`libbb/copy_file.c:351`	`expected C expression`
`editors/vi.c:791`	`expected C expression`
`util-linux/fdformat.c:82`	`expected C expression`
`util-linux/rtcwake.c:102`	`expected C expression`
`libbb/unicode.c:124`	`IR_UNSUPPORTED: AST node (kind 1)`
`util-linux/rev.c:51`	`IR_UNSUPPORTED: AST node (kind 1)`
`libbb/lineedit.c:367`	`stray token at top level: 'wchar_t'`
`editors/awk.c`	`#error … RAND_MAX` — a header value, not the compiler

Four share expected C expression and two share IR_UNSUPPORTED, so this is plausibly three or four distinct causes rather than eight. Do not assume it: test.c/ed.c/acpid.c shared a symptom string with copy_file.c today and turned out to be a different bug, which is why they were separated before being fixed.

Gate

test/casm_barrier.c (gcc -O0 oracle: every accepted spelling, plus a barrier between two increments in a function and at top level) and test/casm_nonempty_template_fails.c (must NOT compile), both wired into test-core.