Phobos range test

This simple program to test the zero cost range using the Phobos v2.

The sample source can be found in the range_test.d file in this repository.

The test makes a benchmark between a for loop and range using Phobos.

The CUT (Code Under Test) is a simple pointer list with an inner Range that implements a InputRange.

The List is implemented as a class. The CUT can be compiled as an FINAL version or virtual version.

    version(FINAL) {
        final Range opSlice() pure nothrow @nogc {
            return Range(this);
        }
    }
    else {
        Range opSlice() pure nothrow @nogc {
            return Range(this);
        }
    }

The CUT can be run with the make tag run, run-o, run-final, run-final-o

The sub-tag final means vesion=FINAL and the sub-tag o means that the CUT is compiled with optimizing.

First, the tests are performed with dmd.

1. CUT without optimizing and virtual opSlice function.

make clean run DMD=1

DMD64 D Compiler v2.109.1
range sum = 49995000
for sum   = 49995000
benchmark range test = 42 secs, 528 ms, 196 μs, and 9 hnsecs
benchmark for test   = 28 secs, 202 ms, 327 μs, and 9 hnsecs
Reation 150.797%

2. CUT with optimizing and virtual opSlice function.

make clean run-o DMD=1

DMD64 D Compiler v2.109.1
range sum = 49995000
for sum   = 49995000
benchmark range test = 40 secs, 916 ms, 941 μs, and 3 hnsecs
benchmark for test   = 9 secs, 569 ms, 864 μs, and 8 hnsecs
Reation 427.560%

3. CUT without optimizing and final opSlice function.

make clean run-final DMD=1

DMD64 D Compiler v2.109.1
range sum = 49995000
for sum   = 49995000
benchmark range test = 38 secs, 193 ms, 321 μs, and 2 hnsecs
benchmark for test   = 25 secs, 960 ms, 902 μs, and 1 hnsec
Reation 147.119%

4. CUT wih optimizing and final opSlice function.

make clean run-final-o DMD=1

DMD64 D Compiler v2.109.1
range sum = 49995000
for sum   = 49995000
benchmark range test = 51 secs, 2 ms, and 840 μs
benchmark for test   = 10 secs, 572 ms, 25 μs, and 6 hnsecs
Reation 482.432%

The same test using ldc2.

5. CUT without optimizing and virtual opSlice function.

make clean run

LDC - the LLVM D compiler (1.39.0):
range sum = 49995000
for sum   = 49995000
benchmark range test = 1 minute, 180 ms, 540 μs, and 9 hnsecs
benchmark for test   = 23 secs, 800 ms, 303 μs, and 1 hnsec
Reation 252.856%

6. CUT with optimizing and virtual opSlice function.

make clean run-o

LDC - the LLVM D compiler (1.39.0):
range sum = 49995000
for sum   = 49995000
benchmark range test = 11 secs, 66 ms, 150 μs, and 3 hnsecs
benchmark for test   = 10 μs and 5 hnsecs
Reation 105391907.619%

7. CUT without optimizing and final opSilce function.

make clean run-final

LDC - the LLVM D compiler (1.39.0):
range sum = 49995000
for sum   = 49995000
benchmark range test = 1 minute, 1 sec, 380 ms, 484 μs, and 2 hnsecs
benchmark for test   = 29 secs, 598 ms, 563 μs, and 1 hnsec
Reation 207.377%

8. CUT with optimizing and final opSlice function.

make clean run-final-o

LDC - the LLVM D compiler (1.39.0):
range sum = 49995000
for sum   = 49995000
benchmark range test = 20 μs
benchmark for test   = 19 μs and 3 hnsecs
Reation 101.027%

Conclusion

The benchmark results are listed below.

X	virtual	virtual-opt	final	final-opt
dmd	159.8%	427.6%	147.1%	482.4%
ldc2	252.9%	105391907.6%	207.4%	101.0%

It's clear to see that the virtual function opSlice in what takes up the time.

For the test bench, it can also be concluded if a final opSlice is used and instead and the code is compiled with ldc2 optimized with -O3 the range in this case has zero cost.

This is the code produced by ldc2

This code is the range_sum function with the virtual opSlice.

nothrow @nogc @safe int example.range_sum(example.List!(int).List):
        push    rax
        mov     rax, qword ptr [rdi]
        call    qword ptr [rax + 48] /// virtual function call
        xor     ecx, ecx
        test    rax, rax
        je      .LBB0_3
.LBB0_1:
        add     ecx, dword ptr [rax + 8]
        mov     rax, qword ptr [rax]
        test    rax, rax
        jne     .LBB0_1
.LBB0_3:
        mov     eax, ecx
        pop     rcx
        ret

This is the for loop.

nothrow @nogc @safe int example.for_sum(example.List!(int).List):
        mov     rcx, qword ptr [rdi + 16]
        xor     eax, eax
        test    rcx, rcx
        je      .LBB3_3
.LBB3_1:
        add     eax, dword ptr [rcx + 8]
        mov     rcx, qword ptr [rcx]
        test    rcx, rcx
        jne     .LBB3_1
.LBB3_3:
        ret

So the only difference is the indirect function call and the stack cleanup.

This code is produced with -O3 and the final opSlice function.

nothrow @nogc @safe int example.range_sum(example.List!(int).List):
        mov     rcx, qword ptr [rdi + 16]
        xor     eax, eax
        test    rcx, rcx
        je      .LBB0_3
.LBB0_1:
        add     eax, dword ptr [rcx + 8]
        mov     rcx, qword ptr [rcx]
        test    rcx, rcx
        jne     .LBB0_1
.LBB0_3:
        ret

This is the for loop

nothrow @nogc @safe int example.for_sum(example.List!(int).List):
        mov     rcx, qword ptr [rdi + 16]
        xor     eax, eax
        test    rcx, rcx
        je      .LBB3_3
.LBB3_1:
        add     eax, dword ptr [rcx + 8]
        mov     rcx, qword ptr [rcx]
        test    rcx, rcx
        jne     .LBB3_1
.LBB3_3:
        ret

The code produced by ldc2 is in this case the same.

This means we have zero cost in this case.