implement ctfe yl2x and yl2xp1

[IgorStepanov]

This PR intriduces builtin compile-time yl2x and yl2xp1 for some platforms.
Supported platforms: DMC, VS (x86, x86_64), MinGW (x86 and x86_64), GCC (x86, x86_64), SunC(x86, x86_64).
If platform doesn't support this builtin, "cannot evaluate unimplemented builtin..." error will be raised as it was early.

[IgorStepanov]

@ibuclaw Please, check my *nix inline assebler code. Is it ok?

[rainers]

VS (x86_64) doesn't support inline assembler, so I can't implement yl2x for it.

You can implement x64 asm for VC in vcbuild/ldfpu.asm.

[IgorStepanov]

Ok

[WalterBright]

Why float64 instead of float80?

[IgorStepanov]

I don't know. I did it similar examples like eval_sqrt

Expression *eval_sqrt(Loc loc, FuncDeclaration *fd, Expressions *arguments)
{
    Expression *arg0 = (*arguments)[0];
    assert(arg0->op == TOKfloat64);
    return new RealExp(loc, Port::sqrt(arg0->toReal()), arg0->type);
}

Maybe CTFE stores all floating numbers with op == TOKfloat64

[WalterBright]

Hmm, you're right, the others do it too.

[ibuclaw]

GDC can't use this, but this is not touching any shared code anyway.

[ibuclaw]

Actually, rndtol may be CTFE-able in GDC - I'll have to check this in detail.

[ibuclaw]

> cat test.d 
import std.math : rndtol;
static x = rndtol(1.2);
pragma(msg, rndtol(1.2));

> gdc test.d -S
1L

[IgorStepanov]

Maybe I should remove this test?

[IgorStepanov]

I've removed it.

[IgorStepanov]

@ibuclaw Please comment the next code:

void Port::yl2x_impl(longdouble* x, longdouble* y, longdouble* res)
{
    __asm__ ("fyl2x": "=t" (*res): "u" (*y), "0" (*x) : "st(1)" );
}

Is it ok? It works, but I'm scared by unused assembler syntax. Do I need to add additional instructions like fwait or finit?

[IgorStepanov]

Added solaris support: tested with Sun C 5.12 Linux_i386 in x86 and x86_64 mode.

[WalterBright]

@IgorStepanov you don't need fwait with modern CPUs, and finit is only needed at program startup.

[WalterBright]

Auto-merge toggled on

[IgorStepanov]

Thanks!

@IgorStepanov you don't need fwait with modern CPUs

It was needed only when x87 was implemented as separate chip?

and finit is only needed at program startup.

I thinked, finit should be called after using xmm indtructions.
Is it wrong?

[CyberShadow]

This broke building DMD x64 with MSVC 2010.

root\port.c(296): error C2065: 'long_double' : undeclared identifier [C:\Projects\Extern\D\dmd\src\dmd_msc.vcxproj]

[CyberShadow]

If I change long_double to longdouble I get a linking error.

Supported platforms: DMC, VS (x86, x86_64),

Does it work for you?

[CyberShadow]

Fixed it, the functions were not declared as extern "C".

[CyberShadow]

#4041

[MartinNowak]

This shouldn't be here.

[IgorStepanov]

Should I create PR, which fix it? I'll do it after 3-5 hours.

[MartinNowak]

FR https://issues.dlang.org/show_bug.cgi?id=3749

[IgorStepanov]

Created #4052 Is it ok?