bearophile wrote:
Walter Bright:
It turns out that dmd's runtime library function had a crummy
implementation of long division in it.<
In that case I have spent few hours narrowing the problem to a very small
benchmark. Then I have left to you to spot and fix the precise cause of the
low performance. And today I know x86 assembly a bit better :-)
Please forgive me for ragging on about this, but you had originally concluded
that it was the code generator's fault. My point is that drawing such a
conclusion would lead one to spend many hours fruitlessly looking in the wrong
place for a solution, and that looking at the assembler would quickly point one
in the right direction to getting it fixed.
I wrote about this issue because I see it all the time, and by experienced
programmers.
I know a person who has done it (looked at the JIT output) with Java using
a debugger,
I have done this often, many of my bug reports for LLVM compare the asm
produced by LLVM with the asm produced by the Sun JavaVM:
http://weblogs.java.net/blog/2008/03/30/deep-dive-assembly-code-java See for
example: http://llvm.org/bugs/show_bug.cgi?id=5501
Very nice, you should post that on reddit!
