Bug 859 – float vector codegen after inlining very different from manual inlined code

Compiler inlining of functions gives much worse performance than manually inlined functions (at least in some cases). In the attached example, the performance is 6 times slower. C:\>dmd -O -inline -release -g testinline.d C:\>testinline.exe compiler inlined time: 374058 manually inlined time: 61362 C:\>obj2asm testinline.obj -ctestinline.asm See line 486 for the compiler inlined code See line 544 for the manually inlined code The compiler inlined code extra instructions like the following: lea ESI,-080h[EBP] lea EDI,-048h[EBP] movsd movsd movsd lea ESI,-074h[EBP] lea EDI,-03Ch[EBP] movsd movsd movsd These instructions are absent in the manually inlined code, and may be the cause of the poor performance.

Created attachment 92 Example to test inlining of a simple function

Created attachment 93 Assembly code from the example

To avoid opening a new bug, I'll reuse this ancient bug report, since the summary is pretty much the same I'll write for this. I'm having some performance problems moving some stuff from a lower-level C-style to a higher-lever D-style. Here is an example: --- int find_if(bool delegate(ref int) predicate) { for (int i = 0; i < 100; i++) if (predicate(i)) return i; return -1; } int main() { // for (int i = 0; i < 100; i++) // if (i == 99) // return i; // return -1; return find_if((ref int i) { return i == 99; }); } --- The program produced by this source executes 4 times more instructions than the more direct (lower-level) version commented out. I would expect DMD to inline all functions/delegates and produce the same asm for both, but that's not the case. This is a reduced test-case, but I'm working on improving the GC and I'm really hitting this problem. If I use this higher-level style in the GC, a Dil run for generating the Tango docs is 3.33 times slower than the C-ish style used by the current GC. So I think this is a real problem for D, it's really important to be able to encourage people to use the higher-level D constructs.

@Leandro Lucarella: ldc seems to inline the predicate just fine, although the generated code is still slightly different.

(In reply to comment #4) > @Leandro Lucarella: ldc seems to inline the predicate just fine, although the > generated code is still slightly different. Yes, LDC is better at inlining because it doesn't use the front-end inlining code, it let the LLVM optimizer do the job instead (I think they inhibited the DMDFE inliner precisely because of this issues). This bug report is about the DMD implementation.

An improved version of the test program, that allows to compare dmd and ldc on this inlining problem: version (Tango) { import tango.stdc.stdio: printf; import tango.stdc.stdlib: atof; } else { import std.c.stdio: printf; import std.c.stdlib: atof; } struct Vec3 { float x, y, z; } float dot(Vec3 A, Vec3 B) { return A.x * B.x + A.y * B.y + A.z * B.z; } struct Timer { long starttime; static long getTime() { asm { naked; rdtsc; ret; } } void start() { starttime = getTime(); } void stop() { long endTime = getTime(); printf("time: %lld\n", endTime - starttime); } } void main() { int n = 30_000; Vec3 a = Vec3(atof("1.0"), atof("2.0"), atof("3.0")); Vec3 b = Vec3(atof("4.0"), atof("5.0"), atof("6.0")); Timer t; float sum; printf(" Auto inlined "); sum = 0.0; t.start(); for (int i; i < n; i++) { a.x++; a.y++; a.z++; sum += dot(a, b); } t.stop(); printf("sum: %f\n", sum); printf("Manually inlined "); sum = 0.0; t.start();; for (int i; i < n; i++) { a.x++; a.y++; a.z++; sum += a.x * b.x + a.y * b.y + a.z * b.z; } t.stop(); printf("sum: %f\n", sum); }

Guys, piling more stuff into a bug report isn't a good idea. In fact, I need to re-classify this bug since its not a problem with inlining at all. The call to DOT in the original code _is_ being inlined. The resulting code is different than the manually inlined version, but the code IS inlined. While they might be the same, they're different enough right now to call them different bugs. I just split the new report into bug 4440

This was fixed by the changes that fixed bug 2008. This report passes static arrays as a parameter which was one of the things that caused the inliner to reject a function. I'm going to close this bug. I've opened bug 4447 to track a remaining issue regarding oddities involving the first function taking significantly longer to execute, regardless of which it is.