🚀 go-pugleaf

In article <1992Nov18.191730.1044@meiko.com> richard@meiko.com (Richard Cownie) writes:
>1. Conditional move instruction is certainly useful.  I was involved
>   in writing a compiler for the Acorn RISC Machine (ARM), which has
>   the unusual feature that all instructions are conditionally
>   executed, with a 4-bit field to specify the condition (the most
>   common condition being "always").  Any conditional branch over
>   a short sequence of instructions was optimised into a short
>   sequence of conditionally-executed instructions, which reduced
>   the code size and saved flushing the pipeline (there was no smart
>   branch prediction or delayed branches).

The last sentence is key here. If branches can be predicted with decent
accuracy then it is hard to obtain benefits from conditional moves
over using conditional branches. In cases where branches can be predicted,
a good processor front-end can provide the correct instructions
"most of the time" and basically eliminate branches ahead of time.
If the branch is correctly predicted than the full scalarity of
the processor is available for computation. With conditional moves,
the general case is that half of the functional units are available
since you have to duplicate each instruction (half for the setup
and half for the conditional execution).
For hard to predict branches (50% ~ 60%), with a small number of
instructions to guard, then conditional moves may be useful only if the
pipeline is designed to handle the forwarding of the data efficiently
(not free, requires junk in the pipeline).

- Marc Tremblay.
Sun Microsystems.