[llvm-dev] Unify debug and optimized variable locations with llvm.dbg.addr [was: DW_OP_LLVM_memory]

Robinson, Paul via llvm-dev llvm-dev at lists.llvm.org
Mon Sep 11 13:49:04 PDT 2017

Burying the value/address distinction in an expression opcode makes it pretty inconvenient for code that wants the distinction to make a difference, I would think.  I don't know how often that actually matters though.

From: David Blaikie [mailto:dblaikie at gmail.com]
Sent: Monday, September 11, 2017 12:58 PM
To: Reid Kleckner; Adrian Prantl
Cc: Robinson, Paul; llvm-dev; Alex Bradbury; Chandler Carruth
Subject: Re: Unify debug and optimized variable locations with llvm.dbg.addr [was: DW_OP_LLVM_memory]

On Mon, Sep 11, 2017 at 12:35 PM Reid Kleckner <rnk at google.com<mailto:rnk at google.com>> wrote:
On Fri, Sep 8, 2017 at 10:32 AM, Adrian Prantl <aprantl at apple.com<mailto:aprantl at apple.com>> wrote:
> On Sep 7, 2017, at 2:18 PM, Reid Kleckner <rnk at google.com<mailto:rnk at google.com>> wrote:
> On Thu, Sep 7, 2017 at 11:11 AM, Robinson, Paul <paul.robinson at sony.com<mailto:paul.robinson at sony.com>> wrote:
>> Different intrinsics sounds like a good solution to me. J
>> So what happens with the case where a variable is registerized but later we decide to spill it?  Presumably we'd have a dbg.addr to point to the spill slot.  In past compilers I've used, spill slots were treated analogous to register allocation, i.e. some effort was made to minimize the number of spill slots and a variable might be spilled to different slots at different points.  If LLVM does that, then dbg.addr will have to be allowed to associated different addresses with the variable.  On the other hand, if LLVM allocates a unique memory "home" for each spilled variable, then dbg.addr can retain the property you suggest, that the address expression is always the same.
> dbg.addr is really IR only. Machine DBG_VALUE instructions can already represent addresses or values depending on their second argument. At this point, I don't see any reason to change that.
If we can write a verifier to check the validity of a dbg.addr's address, why do we need the separate intrinsic? I guess the answer is that while every address must be a pointer value, not every pointer value is an address. Is this correct?

Mainly just for readability. We're encoding one bit of information: is the result of DWARF expression on the LLVM value argument the variable's address or value? The DW_OP_LLVM_memory proposal encodes that bit as a special opcode in the expression. The dbg.addr proposal makes it more first class: it's part of the IR, the intrinsic, not some possibly semantically unimportant metadata. People seem to prefer that.

People do? I'd actually lean the other way myself (closer to DWARF, fewer first class constructs in the IR/more orthogonality, etc) - FWIW, or in case my other rambling emails were confusing. (well, maybe the intermediate state/mismatch between LLVM IR and DWARF is enough that it can't be close enough to DWARF to be sensible/tidy, so taking it further away as you're suggesting here might be better).
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170911/4570f5bc/attachment.html>

More information about the llvm-dev mailing list