Re: [PATCH v7] RISC-V: Support CALL for RVV floating-point dynamic rounding

[Jeff Law <jeffreyalaw@gmail.com>]

On 7/26/23 07:08, Robin Dapp via Gcc-patches wrote:
> So after thinking about it again - I'm still not really sure
> I like treating every function as essentially an fesetround.
> There is a reason why fesetround is special.  Does LLVM behave
> the same way?
> 
> But supposing we really, really want it and assuming there's consensus:
> 
> +  start_sequence ();
> +  emit_insn (gen_frrmsi (DYNAMIC_FRM_RTL (cfun)));
> +  rtx_insn *backup_insn = get_insns ();
> +  end_sequence ();
> 
> A comment here would be nice why we need a sequence for a single
> instruction.  I'm not fully aware what insert_insn_end_basic_block
> does but won't a
> 
>    rtx_insn *last = BB_END (bb);
>    emit_insn_before_noloc (gen_frrmsi (DYNAMIC_FRM_RTL (cfun)), last, bb);
> 
> suffice?  One way or another need these kinds of non-local
> constructs here don't seem entirely rock solid.
Typically an LCM algorithm needs to insert on edges rather than at the 
end of blocks -- this is particularly important to preserve its property 
that on no path through the CFG can we have more evaluations of the 
expression after PRE/LCM than before PRE/LCM.

The other thing edge insertions do is simplify the abnormal critical 
edge problems.  I'd have to dig into the precise details, but in the 
generic PRE/LCM code we clobber the available expressions on critical 
edges so that we don't try to hold a value live across that edge.

Thus the insertion point will tend to be the normal edge of a block that 
ends with a call to a potentially throwing function.

Inserting on the edge also significantly simplifies handling of 
conditional branches ;-)

Jeff