[llvm-commits] [llvm] r83207 - /llvm/trunk/lib/CodeGen/AsmPrinter/DwarfDebug.cpp
clattner at apple.com
Sun Jan 3 21:22:07 PST 2010
On Oct 1, 2009, at 2:35 PM, Evan Cheng wrote:
> Can't we compute these on demand so codegen passes don't have to change these?
Devang, I never saw an answer to this. I have serious concerns about DbgScope and the presence of two MachineInstr*'s that can dangle seems very dubious.
> On Oct 1, 2009, at 2:25 PM, Devang Patel wrote:
>> On Oct 1, 2009, at 1:38 PM, Dan Gohman wrote:
>>> On Oct 1, 2009, at 11:25 AM, Devang Patel wrote:
>>>> Author: dpatel
>>>> Date: Thu Oct 1 13:25:23 2009
>>>> New Revision: 83207
>>>> URL: http://llvm.org/viewvc/llvm-project?rev=83207&view=rev
>>>> Record first and last instruction of a scope in DbgScope.
>>> How does this interact with Post-RA scheduling, MachineLICM, and MachineHoist?
>> It depends on when we run this DwarfDebug pass. Today, at iSel time we put the stake in the ground and emit label node (or instruction) to mark scope boundaries. In future, this will be divided into three steps
>> 1 - Just before AsmPrinter, the DwarfDebug will note down scope boundaries in DbgScope based on info attached with an machine instruction.
>> 2 - @processDebugLoc() in AsmPrinter the labels will be emitted to mark scope boundaries and DW will be updated according.
>> 3 - At the end, the scope DIEs will be created based on DbgScope entries.
>> So MachineLICM will have as much freedom as standard LICM.
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-commits