spirv-opt: Fix OpCompositeExtract relaxation with struct operands

[bejado]

When running the ConvertToHalf pass, an OpCompositeExtract instruction with a OpTypeStruct operand should not be demoted to half. This was (partially) fixed in #5201. However, that PR did not address a couple of scenarios:

1. The OpCompositeExtract instruction is explicitly decorated with RelaxedPrecision.

               OpDecorate %9 RelaxedPrecision
    %v4float = OpTypeVector %float 4
  %_struct_6 = OpTypeStruct %v4float
          %7 = OpUndef %_struct_6
               ...
          %9 = OpCompositeExtract %float %7 0 3
               OpReturn

2. The OpCompositeExtract instruction's usages are also relaxed (or it has no uses), so it is automatically decorated with RelaxedPrecision. See this block in convert_to_half_pass.cpp.

    %v4float = OpTypeVector %float 4
  %_struct_6 = OpTypeStruct %v4float
          %7 = OpUndef %_struct_6
               ...
          %9 = OpCompositeExtract %float %7 0 3
; %9 is not used, or its uses are also relaxed
               OpReturn

Otherwise, spirv-opt ends up generating SPIR-V which does not pass validation:

error: line 15: Result type (OpTypeFloat) does not match the type that results from indexing into the composite (OpTypeFloat).
  %9 = OpCompositeExtract %half %7 0 3

This PR is an attempt to address these issues by more strictly disallowing an OpCompositeExtract instruction to be demoted to half precision if it has a OpTypeStruct operand.

[CLAassistant]

All committers have signed the CLA.

[greg-lunarg]

Please change reviewer to @jeremy-lunarg

[jeremy-lunarg]

LGTM