[ir] SNodeOpStmt use GlobalPtrStmt for all SNodeOpTypes

[xumingkuan]

Related issue = #742

This PR removes the indices field from SNodeOpStmt.

I have a question in lower_access: what if a global pointer is lowered with activate == false by a SNodeOpStmt or a GlobalLoadStmt, but later we need to activate that pointer?

[Click here for the format server]

---

[yuanming-hu]

Awesome! LGTM, and welcome back!

I have a question in lower_access: what if a global pointer is lowered with activate == false by a SNodeOpStmt or a GlobalLoadStmt, but later we need to activate that pointer?

I'm not exactly sure what "later we need to activate that pointer" refers to, but here's what I remember about the activate flag:

• When created, each GlobalPtrStmt has activate=true
• In flag_access (before lower_access), for read-only pointers we mark activate=false

(There doesn't seem to be a place where we mark a pointer with activate=false as activate=true? :-) Correct me if I'm wrong.)

[xumingkuan]

Awesome! LGTM, and welcome back!

I have a question in lower_access: what if a global pointer is lowered with activate == false by a SNodeOpStmt or a GlobalLoadStmt, but later we need to activate that pointer?

I'm not exactly sure what "later we need to activate that pointer" refers to, but here's what I remember about the activate flag:

• When created, each GlobalPtrStmt has activate=true
• In flag_access (before lower_access), for read-only pointers we mark activate=false

(There doesn't seem to be a place where we mark a pointer with activate=false as activate=true? :-) Correct me if I'm wrong.)

IIUC, flag_access in fact does the following:

• Mark each GlobalPtrStmt as activate=false
• For global stores and atomic ops, mark the pointers as activate=true
• For global pointers with activate=true inside struct fors, mark them as activate=false if they share sparsity with the loop indices

[yuanming-hu]

Awesome! LGTM, and welcome back!

I have a question in lower_access: what if a global pointer is lowered with activate == false by a SNodeOpStmt or a GlobalLoadStmt, but later we need to activate that pointer?

I'm not exactly sure what "later we need to activate that pointer" refers to, but here's what I remember about the activate flag:

• When created, each GlobalPtrStmt has activate=true
• In flag_access (before lower_access), for read-only pointers we mark activate=false

(There doesn't seem to be a place where we mark a pointer with activate=false as activate=true? :-) Correct me if I'm wrong.)

IIUC, flag_access in fact does the following:

• Mark each GlobalPtrStmt as activate=false
• For global stores and atomic ops, mark the pointers as activate=true
• For global pointers with activate=true inside struct fors, mark them as activate=false if they share sparsity with the loop indices

This sounds reasonable to me - maybe you can show a specific case where "activate that pointer" becomes problematic?

[xumingkuan]

IIUC, flag_access in fact does the following:

• Mark each GlobalPtrStmt as activate=false
• For global stores and atomic ops, mark the pointers as activate=true
• For global pointers with activate=true inside struct fors, mark them as activate=false if they share sparsity with the loop indices

This sounds reasonable to me - maybe you can show a specific case where "activate that pointer" becomes problematic?

I see... I thought each pointer should be lowered only once, but I re-read lower_access and found that it is the operation (global load/store, atomic op, snode op) rather than the pointer that gets replaced (with lowered statements). There should be no problem here.