[herd] Add LDAP/STLP instructions

[RaghavRoy145]

No description provided.

[relokin]

Thanks for this! I've got two comments.

[relokin]

@HadrienRenaud can I please ask your opinion on the changes to this file?

[HadrienRenaud]

I am happy with the file as it is now.

[relokin]

I am not sure we will retain backwards compatibility, but I think diyone7 will need to support the new instructions. This means we need the annotations PaL -> STLP and PaA -> LDAP. Now I think we can either keep interpreting PaI as either LDIAPP or STILP or we can split that to PaIQ and PaIL. Not that right now I can't generate a test with STILP. Neither this works:

$> diyone7 -metadata false -arch AArch64 PodWW PaIL Rfe PodRR Fre

nor this:

$> diyone7 -metadata false -arch AArch64 PodWW PaI Rfe PodRR Fre

and indeed neither annotation is valid

$> diyone7 -arch AArch64 -show annotations
 P X XL XA XAL L Q A PaL PaA PaIQ PaN Pa

[relokin]

I don't think we should keep this type

[RaghavRoy145]

• I've kept pair_opt because having separate Ld_Pair and St_Pair caused assertion failures in both add_lxm_atom and add_lxm_edge as the pretty-printers for shared variants like Pa and PaN produce identical strings, but compare_atomo sees them as different constructors.
• Using prefixes (LdPa/StPa) breaks backward compatibility with diyone7 and user-facing tests.
• The unified Pair at the was the only approach that preserved backward compatibility while avoiding these issues, and this needs the pair_opt type. The split into ld_pair_opt/st_pair_opt is preserved at the instruction level (I_LDP/I_STP) where direction is known, with pair_opt_to_ld and pair_opt_to_st conversion functions used in emit_access where the direction is relevant again.

[relokin]

why are we turning this into a Polymorphic type?

[RaghavRoy145]

I used polymorphic variants so that ld_pair_opt and st_pair_opt could share common tags (Pa, PaN) while having separate type constraints on I_LDP vs I_STP.

With regular variants, the shared constructors (Pa, PaN) would need to be duplicated across both types (e.g. LdPa/StPa, LdPaN/StPaN), and every conversion and pattern match becomes more verbose. If regular variants are preferred, it's a straightforward change. Happy to switch if there's a preference.

[relokin]

shouldn't we also split this to a ld and st variant?

[RaghavRoy145]

Currently atom_acc uses the Pair of pair_opt * pair_idx.

At the instruction level, the pretty-printing of I_LDP/I_STP doesn't go through pp_pair_opt and it uses the instruction printer directly. So there's no need for split pp_ld_pair_opt/pp_st_pair_opt functions.

If we were to remove the unified pair_opt and go back to Ld_Pair/St_Pair in atom_acc, we cpould split printers, but then we'd hit the add_lxm_atom and add_lxm_edge assertion failures described above since both would produce the same strings for shared variants like Pa and PaN.

[relokin]

@ShaleXIONG can you please review and help with the changes to gen/?

[RaghavRoy145]

I am not sure we will retain backwards compatibility, but I think diyone7 will need to support the new instructions. This means we need the annotations PaL -> STLP and PaA -> LDAP. Now I think we can either keep interpreting PaI as either LDIAPP or STILP or we can split that to PaIQ and PaIL. Not that right now I can't generate a test with STILP. Neither this works:

$> diyone7 -metadata false -arch AArch64 PodWW PaIL Rfe PodRR Fre

nor this:

$> diyone7 -metadata false -arch AArch64 PodWW PaI Rfe PodRR Fre

and indeed neither annotation is valid

$> diyone7 -arch AArch64 -show annotations
 P X XL XA XAL L Q A PaL PaA PaIQ PaN Pa

This I think is because of a bug in my PR, I failed to include "PaIL" in fold_pair (I overwrote it in the second commit)

Here is a working example after I fixed it locally:

herdtools7 % _build/install/default/bin/diyone7 -metadata false -arch AArch64 PodWW PaIL Rfe PodRR Fre
AArch64 MP+poppail+po
{
 int y[2]={0,0};
 0:X1=x; 0:X2=y;
 1:X1=x; 1:X2=y;
}
 P0               | P1          ;
 MOV W0,#1        | LDR W0,[X2] ;
 STR W0,[X1]      | LDR W3,[X1] ;
 MOV W3,#2        |             ;
 SUB W4,W3,#1     |             ;
 STILP W4,W3,[X2] |             ;

exists (1:X0=1 /\ 1:X3=0)

And here is one with STLP:

herdtools7 % _build/install/default/bin/diyone7 -metadata false -arch AArch64 PodWW PaL Rfe PodRR Fre 
AArch64 MP+poppal+po
{
 int y[2]={0,0};
 0:X1=x; 0:X2=y;
 1:X1=x; 1:X2=y;
}
 P0              | P1          ;
 MOV W0,#1       | LDR W0,[X2] ;
 STR W0,[X1]     | LDR W3,[X1] ;
 MOV W3,#2       |             ;
 SUB W4,W3,#1    |             ;
 STLP W4,W3,[X2] |             ;

exists (1:X0=1 /\ 1:X3=0)