digitalmars.D.learn - FP magic in std.math.pow
- Seb (54/54) Jul 31 2016 Consider this short program:
- Stefan Koch (5/12) Jul 31 2016 It's an anoying feature.
- Seb (4/19) Aug 01 2016 It's about 1000 instructions with std.math.pow.
Consider this short program:
void main()
{
alias S = float;
S s1 = 0x1.24c92ep+5;
S s2 = -0x1.1c71c8p+0;
import std.math : std_pow = pow;
import core.stdc.stdio : printf;
import core.stdc.math: powf;
printf("std: %a\n", std_pow(s1, s2));
printf("pow: %a\n", s1 ^^ s2);
printf("pow: %a\n", powf(s1, s2));
version(LDC)
{
import ldc.intrinsics : llvm_pow;
printf("ldc: %a\n", llvm_pow(s1, s2));
}
}
std: 0x1.2c155ap-6
pow: 0x1.2c155ap-6
powf: 0x1.2c1558p-6
As you can see below the C powf compiles to the assembly powf and
LDC compiles to powf too. The output of std.math.pow is rather
large, hence not listed.
1) Is this a bug in Phobos or just a very annoying "feature"?
2) I thought that DMD was decoupled from Phobos? So I was very
astonished to see that it's not (see [1])
[1]
https://github.com/dlang/dmd/blob/master/src/expression.d#L14781
```
.text._Dmain segment
assume CS:.text._Dmain
_Dmain:
push RBP
mov RBP,RSP
sub RSP,010h
movss XMM0,FLAT:.rodata[00h][RIP]
movss -8[RBP],XMM0
movss XMM1,FLAT:.rodata[00h][RIP]
movss -4[RBP],XMM1
movss XMM1,-4[RBP]
movss XMM0,-8[RBP]
call powf PC32
cvtss2sd XMM0,XMM0
mov EDI,offset FLAT:.rodata 32
mov AL,1
call printf PC32
xor EAX,EAX
leave
ret
0f1f
add 0[RCX],AL
.text._Dmain ends
```
Jul 31 2016
On Sunday, 31 July 2016 at 22:38:59 UTC, Seb wrote:
Consider this short program:
void main()
{
alias S = float;
S s1 = 0x1.24c92ep+5;
S s2 = -0x1.1c71c8p+0;
[...]
It's an anoying feature.
The reason this is not implemented in dmd is that pow does not
map to a simple cpu instruction on x86.
I will have another shot at fixing this once the CTFE stuff is in.
Jul 31 2016
Seb <seb wilzba.ch> On Sunday, 31 July 2016 at 22:45:16 UTC, Stefan Koch wrote:On Sunday, 31 July 2016 at 22:38:59 UTC, Seb wrote:It's about 1000 instructions with std.math.pow. Yeah any improvement of the FP magic in DMD would be highly appreciated ;-)Consider this short program: void main() { alias S = float; S s1 = 0x1.24c92ep+5; S s2 = -0x1.1c71c8p+0; [...]It's an anoying feature. The reason this is not implemented in dmd is that pow does not map to a simple cpu instruction on x86. I will have another shot at fixing this once the CTFE stuff is in.
Aug 01 2016