Insufficient Optimization of Variadic Functions for Some Targets

[jonathan-gruber-jg]

Clang insufficiently optimizes variadic functions for some target architectures. I only tested the target architectures aarch64 and riscv64, so I'm unsure about other target architectures.

Below is a minimal test case:

#include <stdarg.h>

void yes_ap(int dummy, ...) {
	va_list ap;

	va_start(ap, dummy);

	va_end(ap);
}

void no_ap(int dummy, ...) {
	/* Nothing. */
}

Host system: Arch Linux, x86_64.

Clang version: official Arch Linux package of clang, version 18.1.8-4.

Command line to reproduce results: clang -c test.c -O<opt-level>.

aarch64 assembly with -O2, -O3, -Os, or -Oz is along the lines of this:

yes_ap:
	sub	sp, sp, #192
	stp	x1, x2, [sp, #136]
	stp	x3, x4, [sp, #152]
	stp	x5, x6, [sp, #168]
	str	x7, [sp, #184]
	stp	q0, q1, [sp]
	stp	q2, q3, [sp, #32]
	stp	q4, q5, [sp, #64]
	stp	q6, q7, [sp, #96]
	add	sp, sp, #192
	ret

no_ap:
	sub	sp, sp, #192
	stp	x1, x2, [sp, #136]
	stp	x3, x4, [sp, #152]
	stp	x5, x6, [sp, #168]
	str	x7, [sp, #184]
	stp	q0, q1, [sp]
	stp	q2, q3, [sp, #32]
	stp	q4, q5, [sp, #64]
	stp	q6, q7, [sp, #96]
	add	sp, sp, #192
	ret

riscv64 assembly with -O2, -O3, -Os, or -Oz is along the lines of this:

yes_ap:
	addi	sp, sp, -64
	sd	a7, 56(sp)
	sd	a6, 48(sp)
	sd	a5, 40(sp)
	sd	a4, 32(sp)
	sd	a3, 24(sp)
	sd	a2, 16(sp)
	sd	a1, 8(sp)
	addi	sp, sp, 64
	ret

no_ap:
	addi	sp, sp, -64
	sd	a7, 56(sp)
	sd	a6, 48(sp)
	sd	a5, 40(sp)
	sd	a4, 32(sp)
	sd	a3, 24(sp)
	sd	a2, 16(sp)
	sd	a1, 8(sp)
	addi	sp, sp, 64
	ret