Bug 73 – Functions used to initialize variables are not inlined.

/* Compile with -O -inline -release This program compares performance of 2 loops, one initializing a scope local variable w/ a function and the other not. The function used as an initializer is not inlined. There is about a 11x difference on my system. */ import std.stdio, std.date, std.math; // local copy of std.math.abs() // (imported functions are not inlined - see DMD bug #67) int abs(int x) { return x >= 0 ? x : -x; } void main() { int sum = 0; d_time s = getUTCtime(); for(int i = 0; i < 100_000_000; i++) { int val = abs(i); sum += val - val; } d_time e = getUTCtime(); writefln("std.math.abs(): sum= ",sum,", secs = ",(e - s)/cast(double)TicksPerSecond); d_time tmp = e - s; sum = 0; s = getUTCtime(); for(int i = 0; i < 100_000_000; i++) { int val; val = abs(i); sum += val - val; } e = getUTCtime(); writefln("local abs(): sum= ",sum,", secs = ",(e - s)/cast(double)TicksPerSecond); writefln("ratio = ", tmp / cast(double)(e - s)); }

There's a comment in inline.c for DeclarationExp's saying: // Should scan variable initializers But it isn't happening right now. Adding a basic scan of the ExpInitializer (if present) the same way DeclarationExp::inlineCost works seems to resolve this, but considering how trivial that is... I can't help but expect that there's more to it than just that. -[Unknown]

(In reply to comment #1) > There's a comment in inline.c for DeclarationExp's saying: > > // Should scan variable initializers > > But it isn't happening right now. Adding a basic scan of the ExpInitializer > (if present) the same way DeclarationExp::inlineCost works seems to resolve > this, but considering how trivial that is... I can't help but expect that > there's more to it than just that. > I've noticed that too and agree there must be more to it for some cases. But couldn't that just be re-written by the compiler as two expressions and then inlined? int i = foo(); => int i; i = foo(); => doInline(); Are there cases where that wouldn't be semantically identical? Thanks, - Dave > -[Unknown]

Created attachment 8 Add a basic scan to DeclarationExp. AFAICT, that's exactly what adding a scan in there does. I'm just compiling DMD without backend hooks, but I can see with my changes that it now wants to inline things (by turning on CANINLINE_LOG.) It's just calling the inline stuff on ExpInitializer's exp, which is probably an AssignExp or something, which is a BinExp, which then has the CallExp checked, which then finally gets inlined. I'm just worried it wasn't added because it might cause fallout. I don't really know the code base, so I'm not privy to what fallout might be caused. Anyway, here's a patch that shows what I changed to enable inlining in these cases. If I had the time on my hands, I might try running DStress with this patch on gdc, to see if it causes any regressions. -[Unknown]

Fixed in DMD 0.151.