Have you? I used to write device drivers mostly on ARM architectures and I have newer seen an instruction which would let me precisely control CPU instruction ordering or what it does with cache that I couldn't do in C.

That is why stuff like V-Tune exist with micro-ops profiler and pipeline debugging, maybe ARM should provide better tooling.

This tool helps as much with C (or even higher level languages) as it does with Assembly. It doesn't change the fact both languages work on the same abstract machine model.

What is C's abstract model for NEON?

A simple c library:

https://github.com/projectNe10/Ne10

You have not yet shown an argument for why c is not a portable assembly.

Access to CPU-specific flags like the carry flag. When doing a multiple-precision addition, if the CPU carry bit is inaccessible, then it must instead be compared to the result to see if it overflowed, which takes 3-5 more instructions per limb or worse on modern superscalar processors that are serial in terms of data accesses.

Portable by definition means it cannot capture every specific detail of every architecture out there.

The carry flag is not a specific architectural detail; it is part of the output from the prime example of a basic electronic circuit that is the digital adder.

It is an architectural detail. MIPS for example doesn't have a carry flag. You have to do the exact same procedure in assembly on MIPS you have described in your previous comment when doing multiple word addition.

MIPS along with DEC alpha do not have adder circuits as a design goal was not to have flags. This along with the scenario of multi-precision arithmetic resulted in the implementation of multi-precision arithmetic on MIPS and DEC alpha using a form of branch like instruction with an ALU stage instead of the unavailable instructions dependent on the alternative adder circuits. The choice of having or not having a carry flag alone is not an architectural detail as the choices of [micro]architecture involve how operations are implemented in circuits/hardware. The carry flag specifically has to be considered in the context of its [the adder] circuit when making architecture decisions. The lack of control over flags leading to the inability to compile the same C program to the same assembly operations with same behavior on different hardware demonstrates that a strictly C program is unable to express a consistent portable assembly program across all hardware.

You are basically arguing CF == !CF. On MIPS you literally have to compare your result with an operand in order to establish if there was an overflow. On Arm you just look at the status register.

When I was doing embedded development, I used C as portable assembly in practice and I know I wasn't the only one. Can you do everything an architecture enables you in a portable way? No, of course not.

By definition there isn't such thing as portable Assembly.

Not by definition, but by practice as generating assembly is usually done to optimize performance. Languages restricted to generate a common feature set between isomorphic ISAs and bijectively associated assembly languages are at a performance disadvantage compared to the full ISA specific assembly.