Rework the intrinsic dependencies for SVE instructions

[yuxiliu-arm]

Supersedes #963.

• Add iico_ctrl to AnyActiveElement checks (for scatter load and gather store)
• Add iico_ctrl to ActivePredicateElement checks
• Add iico_order to SVE store¹
• Fix the semantics for the SVE NEG instruction

Example graph change for herd/tests/instructions/AArch64.sve/Y01.litmus:

LD1D {Z0.D},P0/Z,[X0] (load, predicated) before:

after:

ST1D {Z0.D},P0,[X0] (store, predicated) before:

after:

For herd/tests/instructions/AArch64.sve/V16.litmus's LD1W {Z2.S},P0/Z,[X1,Z1.S,UXTW #2] (gather load), before:

after:

For herd/tests/instructions/AArch64.sve/V17.litmus's ST1W {Z1.S},P0,[X1,Z1.S,UXTW #2] (scatter store), before:

after:

Footnotes

1. Arm ARM B2.5 SVE memory ordering relaxations, R_CJHWV: "When a single SVE vector store instruction generates multiple writes to the same location, the instruction ensures that these writes appear in the coherence order for that location, in order of increasing vector element number. No other ordering restrictions apply to memory effects generated by the same SVE store instruction." ↩

[yuxiliu-arm]

Actually, the load semantics is still a bit off. There should be iico-data from the predicate reading event to the branching event. Let me fix it.

[yuxiliu-arm]

The new commit fixes the predicate register read iico-data issue and also adjust the base register read and the data register read. Now for herd/tests/instructions/AArch64.sve/Y01.litmus, the "after" figure for continuous load is as follows:

and continuous store:

Notice that in the continuous case, the base register is read no matter the result of AnyActive branching (e.g. in the continuous store figure above, y is not iico-ctrl after ev26). This is due to the way its ASL spec is written. This is different from the gather load and scatter store case, where the base register is only read if there are active elements. Whether this is intentional remains to be seen, but we should be able to construct a test to illustrate the differences.

[relokin]

Thank you so much for taking this over @yuxiliu-arm!

This looks very good! I've added a couple of comments just on the documentation/comments of the code.

[relokin]

The code looks good to me. As it stands we have two options:

• You rework the commits to clean up the history, or
• We squash all commits.

Otherwise this looks ready and I am happy to merge by the end of the week unless there are any objections.

[relokin]

Thanks @yuxiliu-arm, I will merge this on Friday unless there any objections.

[relokin]

Merged into master. Thanks @yuxiliu-arm !