[DebugInfo] Adding DIExpression into SIL

[mshockwave]

This PR adds DIExpression infrastructure into SIL.

Previously, we can not use debug_value to mark the updated (SIL) SSA of a sub-field inside an aggregate-type source variable. For example:

struct MyStruct {
    var x : Int
}
%2 = alloc_stack $MyStruct, name "my_struct", loc "file.swift":3:4, scope 3
...
%5 = struct_element_addr %2 : $*MyStruct, #MyStruct.x, loc ...

In this snippet we are not able to generate a correct LLVM debug info for %5 using debug_value or debug_value_addr. Because the current SIL debug infrastructure lacks the equivalent feature of LLVM's !DIExpression. More specifically, the DW_OP_fragment operator.

This is not a problem in -Onone, but an outstanding one in optimized build where optimizations -- SROA for example -- might break an aggregate-type object apart or even eliminate the original aggregate type. In which case the debug info quality for variable values will degrade.

This PR (when it's finished) essentially adds functionalities similar to !DIExpression at the SIL level. Take the code above as the example, we're now able to generate the following SIL:

struct MyStruct {
    var x : Int
}
%2 = alloc_stack $MyStruct, name "my_struct", loc "file.swift":3:4, scope 3
...
%5 = struct_element_addr %2 : $*MyStruct, #MyStruct.x, loc ...
debug_value_addr %5 : $*Int, var, (name "my_struct", loc "file.swift":3:4, scope 3), type $MyStruct, expr op_fragment:#MyStruct.x, loc "file.swift":10:11, scope 3

Here are the new parameters / features added to debug_value / debug_value_addr:

• The debug variable can have (optional) SILLocation and/or SILDebugScope attached (the loc and scope directives inside the parens enclosing name), which represents the source location where it was originally declared. If this extra info is absent, the instruction's source location will be used.
• The type directive can be used to indicate the type where the debug variable was originally declared. If this info is absent, the type of the SSA value (i.e. %5 in this case) will be used.
• Finally, the expr directive is used to carry the DIExpression that basically has a one-to-one mapping to LLVM's !DIExpression. Currently we only have op_fragment and op_deref operator, corresponding to DW_OP_LLVM_fragment and DW_OP_deref, respectively.

The part that is slightly controversial is probably how we carry these extra information on in-memory representation for the affected SIL instructions. This change will make SILDebugVariable, the class that represents debug variable and used by all debug-info-carrying instructions (including the alloc_* family), more expensive. Luckily most of the time SILDebugVariable is generated on-demand (it is actually not stored inside aforementioned SIL instructions) so hopefully it won't affect conventional SIL instruction transformations (like copying or moving SIL instructions).

I'm splitting this PR into two commits: The first commit add DIExpression support into SIL instructions classes (e.g. DebugValueInst) and changes regarding SILParser / SILPrinter; The second one add IRGen support.

[mshockwave]

I'll add test cases shortly

[mshockwave]

I'll squash all the commits into one right before it's ready to merge (or I can do it right now if you prefer)

[mshockwave]

@swift-ci test

[mshockwave]

@swift-ci please test

[swift-ci]

Build failed
Swift Test Linux Platform
Git Sha - 045a0ae249842a6c3313de7cb6bd08b983879498

[swift-ci]

Build failed
Swift Test OS X Platform
Git Sha - 045a0ae249842a6c3313de7cb6bd08b983879498

[mshockwave]

@swift-ci please test

[mshockwave]

Add documentations
@swift-ci please test

[adrian-prantl]

Thanks, this is looking great! It would be good to also have test cases for the Verifier assertions that are added by this patch.

[adrian-prantl]

@jckarter Do you have any opinion on the proposed SIL syntax extension?

[mshockwave]

Thanks, this is looking great! It would be good to also have test cases for the Verifier assertions that are added by this patch.

Sorry I almost missed this comment.
And hopefully the windows build bot is back to work 😛

@swift-ci please test

[adrian-prantl]

LGTM.