Creator of LLVM, Clang, and Swift Says To Not Write Security Critical Code In C/C++
via Digitalmars-d
digitalmars-d at puremagic.com
Wed Jul 15 02:36:00 PDT 2015
On Wednesday, 15 July 2015 at 08:32:19 UTC, Kagamin wrote:
> On Wednesday, 15 July 2015 at 00:28:47 UTC, Ola Fosheim Grøstad
> wrote:
>> I am happy if I see that I have consistent 20% headroom, and I
>> am not going to be happier by making my program faster...
>
> But then why optimizations would matter? If the program is
> fast, you won't improve it by improving performance by 2 times:
> it will remain fast, and if it's slow, it's probably an
> algorithmic complexity.
Ability to optimize later matters because I need to keep the
deadline, or else the real time thread will be killed by the OS.
If I miss the deadline occasionally, maybe I only need 10%
improvement.
I have many options. I can reduce fidelity in an audio
application. I can put more work into integrating two loops into
one loop and keep values in SIMD registers based on the number of
SIMD register the particular CPU supports etc...
What I don't want to do is restructure the entire dataset, so I
put more work into memory layout than the initial loop. If the
loop completes in time I'm good, if not, I put more work into it
(or reluctantly reduce fidelity).
> You said it yourself that to get performance from C you need
> extensions, it's not provided by C semantics itself.
No? I said I write "cpu independent simd" in my core performance
oriented loop as a starting point. Whether I need to do that is
debatable... but it ensures that data structures are designed to
be simd friendly from the start.
More information about the Digitalmars-d
mailing list