[LLVMdev] Maybe OT -- what is this compiler optimization formally known as?

[Hayden Livingston]

This is actually not a LLVM issue, but more an issue of my higher
language,that I'm trying to target LLVM on. I'm seeing that if I do a naive
translation of my language, it generates additional load stores.

For example (in my high level language):

  var1 = expression1;
  var2 = expression2;
  var3 = call(var1, var2);
  var4 = expression3;
  var5 = expression4;
  var6 = call(var4, var5);
  var7 = call(var6);
  var8 = call(var3, var7);

This can be effectively represented in my high level language as

var8 = call(call(expression1, expression2),
call(call(expression3, expression4));

When I think about it naively. I would write code that would start from
var8 and traverse and try to substitute all "vars" with expressions.

I can imagine not always wanting to do this, for example if
the expression is duplicated, maybe I do want to pay for the load/store.
And maybe sometimes the semantics are such that I need to reuse it.

Is there a formal algorithm or classical technique that describes this?

Hayden
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20141117/1616f804/attachment.html>