'Missing ]' error with LDRH/STRH + shifted offset

In the ARM BBC BASIC assembler, the following instructions both result in a ‘Missing ]’ error:

LDRH R5, [R2, R3, LSL #1]
STRH R5, [R1, R3, LSL #1]

Omitting the shifts leads to successful assembly, as does omitting the ‘H’ suffix.

I thought those would be valid instructions? Might this be down to a bug in BASIC’s assembler?

I realise that I can shift R3 with a preceding MOV R3,R3,LSL #1, but it seems odd that the assembler doesn’t like shifted offset.

Using BBC BASIC V 1.59 (24 Sep 2014) on a Raspberry Pi 2, RISC OS 5.21 (17-Feb-15).

I’m writing graphics routines which handle 16bpp (RGB565) bitmap images, hence the need for the halfword loads & stores.

David.
—

It’s not a bug – LDRH/STRH don’t support shifted register offsets.

(OK, the Thumb versions of the instructions do support small shift values, but BASIC assembler and RISC OS as a whole don’t really support Thumb so that doesn’t help you much)

Thank you, Jeffrey.

It’s not a bug but it is a rather unhelpful error message. Could it be changed to something more clear, say “Illegal instruction”?

…if you are doing image manipulation or similar I can encourage you to use NEON code. There are also tons of load/store variations like VLD/VLDR/VLDMIA…and when I read RGB565 this comes to my mind (in french…) => Link

It’s not a bug but it is a rather unhelpful error message. Could it be changed to something more clear, say “Illegal instruction”?

“Illegal instruction” is even more vague.

If it was a standalone assembler then you could probably go for something like this, similar to the way modern C compilers annotate their error messages:

']' expected
LDRH R5, [R2, R3, LSL #1]
                ^

Usually that would be enough for the programmer to be able to work out what the problem is, and it doesn’t require the author of the compiler/assembler to write hundreds of different context-specific error messages.

However BASIC’s error handling means that annotated error messages will be difficult (essentially all errors need to be a RISC OS error block, and also most programs start with ON ERROR since the default error handler doesn’t list the line number). So I think the only way of improving the message(s) would be the hard way – add extra code to each place an error is generated to try and infer what the programmer intended, and then produce a more specific message like “LDRH doesn’t support shifted register offsets”.

…if you are doing image manipulation or similar I can encourage you to use NEON code. There are also tons of load/store variations like VLD/VLDR/VLDMIA…and when I read RGB565 this comes to my mind (in french…) => Link

Here’s the ARM blog post it links to (the original URL is dead and it doesn’t look like ARM set up a redirect. Grr!)

Kuemmel wrote:

…if you are doing image manipulation or similar I can encourage you to use NEON code. There are also tons of load/store variations like VLD/VLDR/VLDMIA…and when I read RGB565 this comes to my mind (in french…) => Link

Thanks for the link. Yes, I’ve seen some of your NEON-powered graphics code in action! Very instructive, and I’ll probably be looking at it again over the next few weeks or months.

David.
—

Just don’t expect to get astronomical performance improvements from using NEON on the Pi 2 – the Cortex-A7 NEON unit is designed for low power/die space, not speed. For integer performance it’s probably OK, but for float apparently it’s not much faster than VFP.

https://community.arm.com/processors/f/discussions/5289/neon-cortex-a7-is-4-times-slower-than-cortex-a8