[PATCH] D26769: [IR] Remove the DIExpression field from DIGlobalVariable.

Fri Nov 18 10:23:25 PST 2016

> On Nov 18, 2016, at 10:10 AM, David Blaikie <dblaikie at gmail.com> wrote:
> 
> 
> 
> On Fri, Nov 18, 2016 at 10:00 AM Adrian Prantl <aprantl at apple.com <mailto:aprantl at apple.com>> wrote:
>> On Nov 18, 2016, at 9:51 AM, David Blaikie <dblaikie at gmail.com <mailto:dblaikie at gmail.com>> wrote:
>> 
>> Would it be reasonable/useful to add another level of indirection while we're here - that would describe which part of the variable is being described by this location:
>> 
>> DIGlobalVariableExpr(value: DW_OP_const 42) -> DIGlobalVariableFragment(offset: 32 bits) -> DIGlobalVariable(name: foo)
>> 
>> (just guessing/pretending there's a DW_OP_const for constant values - I forget how it all works, but purely for demonstration purposes)
>> 
>> That way code manipulating DIGlobalVariableExprs wouldn't need to have explicit knowledge of which part of the variable is being described here. (I suppose that could end up with us creating a bottom-up expression tree in metadata, which probably isn't ideal (too heavyweight))
> 
> The nice thing about the Dwarf expressions being a stack-based post-fix-notated programming language is that it you can compose functions by simply concatenating them. The typical kind of manipulations that we do (for example, adding an "add offset", "dereference" operation) can done by simply putting the new expression in front of the old expression. I don't think it can get much simpler than that :-)
> 
> Wouldn't the piece/bit_piece need to go on top level of the expression tree? I don't think you can simply concatenate things on top of it... 

The expression tree of a stack-based postfix language degenerates to a list. The works in the language can consume a different number of stack arguments, but under the assumption that one well-formed DIExpression takes and produces exactly one value (which is true for all non-constant expressions we support), it is safe to concatenate the expressions.

> 
> Oh, also, DW_OP_bit_piece is ordered, judging by the examples in the spec:
> 
> DW_OP_reg3 DW_OP_piece 4 DW_OP_reg10 DW_OP_piece 2
>  A variable whose first four bytes reside in register 3 and whose next two bytes reside in register 10. 
> 
> So how do we know which DIGlobalVariableExpr goes first? Or if there are holes between exprs?
> 
> (am I making sense? not sure - I can try to explain differently, or perhaps just not understanding)

Please let me know if I'm missing the point. I think there may be some confusion about how multi-piece values are represented in LLVM Metadata.

One DIExpression only ever represents one DW_OP_piece (with the DW_OP_piece always being the very last element of the expression). To represent a variable with multiple pieces, we use two DIExpression. For example (extracted from the split-global.ll testcase):

; struct { int x,y; } Point;
@point.x = global i32 1, align 4, !dbg !12
@point.y = global i32 2, align 4, !dbg !13

!0 = distinct !DIGlobalVariable(name: "point", scope: !1, file: !2, line: 1, type: !5, isLocal: false, isDefinition: true)
!12 = !DIGlobalVariableExpression(var: !0,  expr: !DIExpression(DW_OP_bit_piece,  0, 32))
!13 = !DIGlobalVariableExpression(var: !0,  expr: !DIExpression(DW_OP_bit_piece, 32, 32))

The DWARF backend collects all the DIGlobalVariableExpressions and their respective GlobalVariables and translates this into a single location:

DW_AT_location ([0x0000000000000000], piece 0x00000004, [0x0000000000000004], bit-piece 32 32)
                            @point.x, !12, @point.y, !13.

does that make sense?

-- adrian

>  

> 
> --- adrian
> 
>> 
>> Perhaps keeping the bit_piece parts out of the DWARF expression bytes and in a separate field of the DIGlobalVariableExpr?
>> 
>> DIGlobalVariableExpr(offset: 32, value: DW_OP_const 42)
>> 
>> Then at least when a variable moves around (does that happen? maybe? - I suppose something like asan, smooshing everything into a single alloca would move something around without splitting it up) the expression can be updated relatively simply without unpacking through a DW_OP_bit_piece, etc?
>> 
>> Maybe this doesn't matter - Ih aven't looked at much of the expression/location handling code.
>> 
>> On Fri, Nov 18, 2016 at 9:32 AM Adrian Prantl <aprantl at apple.com <mailto:aprantl at apple.com>> wrote:
>> aprantl removed rL LLVM as the repository for this revision.
>> aprantl updated this revision to Diff 78544.
>> aprantl added a comment.
>> 
>> Improved type-safety by adding a DIGlobalVarExpr pointer union.
>> 
>> 
>> https://reviews.llvm.org/D26769 <https://reviews.llvm.org/D26769>
>> 
>> Files:
>>   include/llvm/Bitcode/LLVMBitCodes.h
>>   include/llvm/IR/DIBuilder.h
>>   include/llvm/IR/DebugInfo.h
>>   include/llvm/IR/DebugInfoMetadata.h
>>   include/llvm/IR/GlobalVariable.h
>>   include/llvm/IR/Metadata.def
>>   lib/Analysis/ModuleDebugInfoPrinter.cpp
>>   lib/AsmParser/LLParser.cpp
>>   lib/Bitcode/Reader/BitcodeReader.cpp
>>   lib/Bitcode/Writer/BitcodeWriter.cpp
>>   lib/CodeGen/AsmPrinter/CodeViewDebug.cpp
>>   lib/CodeGen/AsmPrinter/DwarfCompileUnit.cpp
>>   lib/CodeGen/AsmPrinter/DwarfCompileUnit.h
>>   lib/CodeGen/AsmPrinter/DwarfDebug.cpp
>>   lib/IR/AsmWriter.cpp
>>   lib/IR/DIBuilder.cpp
>>   lib/IR/DebugInfo.cpp
>>   lib/IR/DebugInfoMetadata.cpp
>>   lib/IR/LLVMContextImpl.h
>>   lib/IR/Metadata.cpp
>>   lib/IR/Verifier.cpp
>>   lib/Transforms/IPO/StripSymbols.cpp
>>   lib/Transforms/Instrumentation/AddressSanitizer.cpp
>>   test/Assembler/diglobalvariable.ll
>>   test/Assembler/diglobalvariableexpression.ll
>>   test/Bitcode/DIGlobalVariableExpr.ll
>>   test/Bitcode/DIGlobalVariableExpr.ll.bc
>>   test/Bitcode/diglobalvariable-3.8.ll
>>   test/Bitcode/diglobalvariable-3.8.ll.bc
>>   test/DebugInfo/X86/multiple-at-const-val.ll
>>   test/DebugInfo/X86/pr12831.ll
>>   test/DebugInfo/X86/split-global.ll
>>   test/DebugInfo/X86/stack-value-dwarf4.ll
>>   test/DebugInfo/X86/unattached-global.ll
>>   test/Transforms/GlobalMerge/debug-info.ll
>>   unittests/IR/MetadataTest.cpp
>> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20161118/1ffd22fa/attachment.html>