RE: [boost] Re: [Review] Named parameters library
From: Doug Gregor [mailto:dgregor@cs.indiana.edu]
Since I've been researching in this area for several years...
Dave is correct: no compiler out there is going to optimize this, because the side effects are killer. It is possible to analyze that behavior and perform the optimization Thorsten is describing, but
(1) There are better ways than looking at it from the memory subsystem/pointer magic angle (2) This is very much an open research area; don't expect this from a research compiler for at least 2 years. (3) This is very much an open research area; don't expect this from a commercial compiler for at least 10 years.
Copy elision and the RVO are all we have. We can't rely on anything more than that.
out of curiosity: what about the LD_PRELOAD trick that's used on ELF systems to replace symbols at startup ? It's notably used by some tools to swap in special versions of operator new/delete to analyze memory management without the need to recompile anything. Is there any way for a compiler to know that some calls can be fully evaluated at compile-time ? Thanks, Stefan
On Nov 18, 2004, at 10:08 AM, Stefan Seefeld wrote:
From: Doug Gregor [mailto:dgregor@cs.indiana.edu]
Since I've been researching in this area for several years...
Dave is correct: no compiler out there is going to optimize this, because the side effects are killer. It is possible to analyze that behavior and perform the optimization Thorsten is describing, but
(1) There are better ways than looking at it from the memory subsystem/pointer magic angle (2) This is very much an open research area; don't expect this from a research compiler for at least 2 years. (3) This is very much an open research area; don't expect this from a commercial compiler for at least 10 years.
Copy elision and the RVO are all we have. We can't rely on anything more than that.
out of curiosity: what about the LD_PRELOAD trick that's used on ELF systems to replace symbols at startup ? It's notably used by some tools to swap in special versions of operator new/delete to analyze memory management without the need to recompile anything.
The problem goes well beyond replacing symbols... we'd need to so some nontrivial analysis to determine that the operation had no side effects, ignore side effects in the memory subsystem, and then rewrite substantial pieces of code.
Is there any way for a compiler to know that some calls can be fully evaluated at compile-time ?
Sure. It's generally called "partial evaluation", where part of the program is evaluated at compile time, and that may include function calls. Doug
participants (2)
-
Doug Gregor -
Stefan Seefeld