Add LDC-specific traits for CTFE information about the target machine.

[JohanEngelen]

Edit: updated to current state of the PR. (removed __traits(targetFeatures))

This adds the following LDC-specific traits:

__traits(targetCPU) == "broadwell"
__traits(targetHasFeature, "sse2") == bool

Example usage:

import std.stdio;
void main() {
  writeln("CPU = ", __traits(targetCPU));
  writeln("Has 'sse3' = ", __traits(targetHasFeature, "sse3"));
  writeln("Has 'sse4' = ", __traits(targetHasFeature, "sse4"));
  writeln("Has 'sse4.1' = ", __traits(targetHasFeature, "sse4.1"));
}

Outputs on my machine:

❯ ../bin/ldc2 -wi -mcpu=native -run ../../tests/codegen/target_traits.d
CPU = haswell
Has 'sse3' = true
Has 'sse4' = false
Has 'sse4.1' = true

[JohanEngelen]

@9il what do you think?

[9il]

Awesome! Thank you a lot!

[dnadlinger]

Ping @WalterBright @ibuclaw – what do you think about the API?

[WalterBright]

__traits(targetFeatures) == Tuple("sse", "sse2", ...)
__traits(targetHasFeature, "sse2") == bool

The latter is redundant. We should try to avoid redundant core language features.

[UplinkCoder]

so is hasMember

[JohanEngelen]

The latter is redundant. We should try to avoid redundant core language features.

I left targetHasFeature for convenience.
any!q{a = "sse4.1"}([__traits(targetFeatures)]) is quite ugly I think for functionality that is probably most used out of the three traits.

[ibuclaw]

I can't say that gcc readily provides this information, but it seems reasonable.

[WalterBright]

so is hasMember

True, but we learn from our mistakes rather than repeat them.

is quite ugly

__traits is always ugly. It was intended to be buried in library code. Things that are easily done in the library should not be in the core language.

[JohanEngelen]

I can't say that gcc readily provides this information

Neither does LLVM :(
For -mcpu=native the information is easy to obtain, for -mcpu=haswell it is not...

[ibuclaw]

Neither does LLVM :(

Well at least you have getTargetFeatureString and getTargetCPU exposed to the frontend. ;-)

[JohanEngelen]

OK, I've been able to squeeze the feature info out of LLVM with a small LLVM patch :)
I'll submit that patch to LLVM, and see how that goes over. Then for LLVM<3.9, __traits(targetHasFeature, ...) will only work when either explicitly mentioning the feature on the cmdline, or with cpu=native.

[ibuclaw]

I guess you could cache this information in Target also. That may be the best way for me to do it without adding a new gcc target "hook".

[JohanEngelen]

http://reviews.llvm.org/D19189

[mleise]

Nice! I hope this makes it into all compilers. There is always the option to define a safe fall-back like "i386" for dmd on x86. targetHasFeature would be the most frequently used trait in the set. The list version is really only good for "compiled with " output ... unless someone writes a nice wrapper.

@ibuclaw Maybe checking the implementation of this helps:

gcc -march=native -dM -E - < /dev/null | egrep "MMX|SSE|AVX" | sort
#define __AVX__ 1
#define __AVX2__ 1
#define __MMX__ 1
#define __SSE__ 1
…

[ibuclaw]

@mleise - I'm aware of how gcc generates macros, gdc uses the same mechanism for generating pre-defined version conditions. ;-)

All I said was that there is no "ready to use" way of getting the information.

[JohanEngelen]

I removed __traits(targetFeatures) as this list is not obtainable without hacking LLVM. __traits(targetHasFeature) is more useful and this can be obtained from LLVM with some effort.
Updated the PR.

[JohanEngelen]

For LLVM version < 3.7, hasTargetFeature will always return the safe false default. I don't know how to obtain the necessary information with LLVM < 3.7.

[dnadlinger]

We might as well just raise the minimum requirement to 3.7 at some point…

[JohanEngelen]

Ping! :)

[9il]

LGTM

This would be very helpful for Mir's BLAS Micro Kernel parameter estimation. LDC generates (see gist) efficient micro kernel code, and #1472 will allow to optimize it even more :-)

[JohanEngelen]

@redstar Can I merge this? The "hook" is the same as #5771 so it is easily modified once that is merged.

[JohanEngelen]

Will merge when green.

[JohanEngelen]

@9il Can you update us on how you are using this in your code? :)

[9il]

Will do with BLAS release :)