Through profiling I have found a performance problem in my code. I have later reduced the case and created this demo code: import std.typecons: tuple; import std.c.stdio: printf; auto foo() { printf("foo\n"); return tuple(1, 2); } void main() { foreach (x; foo().tupleof) printf("%d\n", x); } Its output, DMD 2.052: foo 1 foo 2 In my code foo() was an expensive computation. In my opinion in this situation foo() needs to be computed only once (or I'd like some other solution to avoid this invisible multiple calls, like some kind of error, etc).
Kai Meyer has shown this related program: import std.typecons: tuple; import std.c.stdio: printf; auto foo() { printf("foo\n"); return tuple(1, 2); } void main() { auto f2 = foo().tupleof; } Its output: foo foo So maybe the problem seems isn't caused by the static foreach.
The problem is in .tupleof, where (expr).tupleof is rewritten to TypeTuple!((expr).field0, (expr).field1, (expr).field2, ...) even if expr have side effect, e.g. ----------- import std.c.stdio : printf; struct S { int x; int y; } void main() { cast(void) (printf("Hi\n"), S(2,3)).tupleof; } // print "Hi" twice. ----------- This should be rewritten as (tmp=expr, TypeTuple!(tmp.field0, ...)). (The ',' here is a comma expression.)
With 2.059 Win32 First example gives: PS E:\DigitalMars\dmd2\samples> rdmd bug foo 1 2 Second example: PS E:\DigitalMars\dmd2\samples> rdmd bug foo Third example: PS E:\DigitalMars\dmd2\samples> rdmd bug Hi
This issue has been already fixed as part of bug 4940.