[llvm-dev] extending liveness of 'this' pointer via FAKE_USE opcode

Mon Sep 21 13:41:22 PDT 2015

From: David Blaikie [mailto:dblaikie at gmail.com]
Sent: Monday, September 21, 2015 11:23 AM
To: llvm-dev; Pieb, Wolfgang; Adrian Prantl; Frédéric Riss
Cc: Eric Christopher
Subject: Fwd: [llvm-dev] extending liveness of 'this' pointer via FAKE_USE opcode

Oops, forgot to reply-all (& adding a few other people to the 'to' line)
---------- Forwarded message ----------
From: David Blaikie <dblaikie at gmail.com<mailto:dblaikie at gmail.com>>
Date: Mon, Sep 21, 2015 at 11:22 AM
Subject: Re: [llvm-dev] extending liveness of 'this' pointer via FAKE_USE opcode
To: "Pieb, Wolfgang" <Wolfgang_Pieb at playstation.sony.com<mailto:Wolfgang_Pieb at playstation.sony.com>>

On Mon, Sep 21, 2015 at 11:16 AM, Pieb, Wolfgang via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:
Hello!

At Sony we've seen some serious customer interest in having the 'this' pointer visible throughout an entire function during debugging. However, optimizations may eliminate it after its last use, so we've been looking for a way to artificially extend its liverange to the end of the function.

How much have you observed the actual (usually a register, yes?) register being reused, versus the debug info just not successfully describing its location? LLVM's optimized debug info still does a pretty bad job tracking optimized values (especially anything that gets promoted out of an alloca into a register)

With respect to the ‘this’ pointer, (or parameters in general) I have observed register reuse after the last actual use of ‘this’. I haven’t seen any instance of the register still holding the value while the debug info is not describing it, but the sample of code I’ve looked at is not very large.

I’m still investigating locals in general.

 So far, the most compelling way we can think of, and one we have used successfully in the past in at least one other compiler,
is to create a 'fake use' of the 'this' pointer at the end of the function, compelling the rest of the compiler to not optimize it away.

Is there some more general heuristic we should consider rather than special casing the 'this' pointer? (something that might help other cases of parameters being optimized away entirely that would still be valuable to the user? Should we try to preserve all parameter values? All variables?)

Preserving all variables is something we have done in a different compiler with relatively minor performance degradation, and we would like to do this in LLVM as well. The ‘this’ pointer and parameters in general are easier to implement because their scope is easy to determine (and there’s been some definite customer interest in the ‘this’ pointer).

 At the moment there doesn't seem to be a good way to create such a fake use in LLVM (please enlighten us if you know of one), so we are
proposing to introduce a new intrinsic (e.g. llvm.fake_use), which would take a single value argument, representing a use of that value.
The intrinsic would be lowered to a new invariant TargetOpcode (e.g. FAKE_USE), which serves the same purpose at the MI level.
Code emission would simply ignore the new opcode.

Frontends could use the intrinsic to extend liveranges of variables as desired. As a first use case, clang would accept a new option
(e.g. -fkeep-this-ptr) which would cause a fake use of 'this' to be inserted at the end of a function, making it available for inspection
throughout the entire function body.

Do you have much in the way of performance numbers to get a sense of the impact of this feature? (& possibly the impact with the sliding scale of "this pointer or all parameters or all variables" sort of thing)

In a different (proprietary) compiler we experimented with making all local variables available throughout their respective lexical scope. The body of evidence is not particularly large, but the performance degradation was in the single digits. There is no data yet for LLVM/clang, but we are planning to collect some.

 One important note is that since such an option would affect code generation, it cannot be automatically enabled by -g. However, should there be
eventually support for a -Og mode (optimize for debugging), that mode could enable it.

Is it worth considering adding -Og now/for this feature?

Yes, we believe this would be a good idea.

Any comments or alternative ideas are appreciated.

_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150921/77106592/attachment.html>