Convention and intent for grammar for aliases. (Was: Possibly wrong `extensions` and `version` for the aliased operands?)

[cmarcelo]

Should GOOGLE aliases be tagged with 1.4? (I'm guessing no) Should the non-GOOGLE variants have empty "extensions" field? (I'm guessing yes)

I'd expect decorate string and hlsl-related annotations to follow the same convention, but they disagree in both questions.

Decorate string (with params omitted):

    {
      "opname" : "OpDecorateString",
      "opcode" : 5632,
      "extensions" : [ "SPV_GOOGLE_decorate_string", "SPV_GOOGLE_hlsl_functionality1" ],
      "version" : "1.4"
    },
    {
      "opname" : "OpDecorateStringGOOGLE",
      "opcode" : 5632,
      "extensions" : [ "SPV_GOOGLE_decorate_string", "SPV_GOOGLE_hlsl_functionality1" ],
      "version" : "1.4"
    },
    {
      "opname" : "OpMemberDecorateString",
      "opcode" : 5633,
      "extensions" : [ "SPV_GOOGLE_decorate_string", "SPV_GOOGLE_hlsl_functionality1" ],
      "version" : "1.4"
    },
    {
      "opname" : "OpMemberDecorateStringGOOGLE",
      "opcode" : 5633,
      "extensions" : [ "SPV_GOOGLE_decorate_string", "SPV_GOOGLE_hlsl_functionality1" ],
      "version" : "1.4"
    },

Hlsl-related annotations (with params omitted):

        {
          "enumerant" : "CounterBuffer",
          "value" : 5634,
          "version" : "1.4"
        },
        {
          "enumerant" : "HlslCounterBufferGOOGLE",
          "value" : 5634,
          "extensions" : [ "SPV_GOOGLE_hlsl_functionality1" ],
          "version" : "None"
        },
        {
          "enumerant" : "UserSemantic",
          "value" : 5635,
          "version" : "1.4"
        },
        {
          "enumerant" : "HlslSemanticGOOGLE",
          "value" : 5635,
          "extensions" : [ "SPV_GOOGLE_hlsl_functionality1" ],
          "version" : "None"
        },

[johnkslang]

We've agreed to evolve this so that:

• the first appearance of a token specifies everything about it
• the spec. itself merges this and all its aliases into a single box/row (as already done for instructions, but not yet done for other classes of enumerants)
• subsequent aliases of the first appearance list only the name and number, neither replicating nor contradicting any of the other fields

There was some tooling that wasn't quite ready for this yet, so the JSON file is not yet sanitized in this way.

[cmarcelo]

So, taking the examples above and modify them to follow the new rules. The decorate string ones would only omit fields in aliases, without changes:

    {
      "opname" : "OpDecorateString",
      "opcode" : 5632,
      "extensions" : [ "SPV_GOOGLE_decorate_string", "SPV_GOOGLE_hlsl_functionality1" ],
      "version" : "1.4"
    },
    {
      "opname" : "OpDecorateStringGOOGLE",
      "opcode" : 5632,
    },
    {
      "opname" : "OpMemberDecorateString",
      "opcode" : 5633,
      "extensions" : [ "SPV_GOOGLE_decorate_string", "SPV_GOOGLE_hlsl_functionality1" ],
      "version" : "1.4"
    },
    {
      "opname" : "OpMemberDecorateStringGOOGLE",
      "opcode" : 5633,
    },

And the HLSL related ones:

        {
          "enumerant" : "CounterBuffer",
          "value" : 5634,
          "extensions" : [ "SPV_GOOGLE_hlsl_functionality1" ],
          "version" : "1.4"
        },
        {
          "enumerant" : "HlslCounterBufferGOOGLE",
          "value" : 5634,
        },
        {
          "enumerant" : "UserSemantic",
          "value" : 5635,
          "extensions" : [ "SPV_GOOGLE_hlsl_functionality1" ],
          "version" : "1.4"
        },
        {
          "enumerant" : "HlslSemanticGOOGLE",
          "value" : 5635,
        },

Is that correct?

If so, why not just add a "aliases" field in the main opcode, and have a single entry per opcode? Room for adding more per-alias data?

[johnkslang]

My understanding is there are still some consumers that want both sets of complete information. I like the idea of adding an aliases field. These questions need to be answered both for instruction enumerants and for non-instruction enumerants, which generally are treated a bit differently.

We need to evolve this while keeping the spec., SPIRV-Tools, and likely other consumers in mind.

(It seems you just cared about the instructions(?), as this style of aliasing has been present for some time in other enumerants; it was not novel, it was following the precedent already set.)

[dneto0]

I agree it's a good direction to ensure consistency for each token. We haven't implemented the update to SPIRV-Tools to understand the new form. I filed an issue (see above) to do that.

[johnkslang]

Also see #108 for any loose ends in documentation.

[spencer-lunarg]

So this "the first item is the truth" isn't followed here

 {
      "opname" : "OpTypeAccelerationStructureNV",
      "class"  : "Reserved",
      "opcode" : 5341,
      "operands" : [
        { "kind" : "IdResult" }
      ],
      "capabilities" : [ "RayTracingNV" , "RayTracingKHR", "RayQueryKHR" ],
      "extensions" : [ "SPV_NV_ray_tracing" , "SPV_KHR_ray_tracing", "SPV_KHR_ray_query" ],
      "version" : "None"
    },
    {
      "opname" : "OpTypeAccelerationStructureKHR",
      "class"  : "Reserved",
      "opcode" : 5341,
      "operands" : [
        { "kind" : "IdResult" }
      ],
      "capabilities" : [ "RayTracingNV" , "RayTracingKHR", "RayQueryKHR" ],
      "extensions" : [ "SPV_NV_ray_tracing" , "SPV_KHR_ray_tracing", "SPV_KHR_ray_query" ],
      "version" : "None"
    },

I guess personally, instead of worrying about ordering, I would also prefer something like a "alias" : "OpTypeAccelerationStructureKHR"

(cc @bashbaug )