Comments on: Playing with LLVM http://blog.reverberate.org/2009/01/12/playing-with-llvm/ parsing, performance, minimalism with C99 Mon, 06 Feb 2012 23:44:51 +0000 hourly 1 http://wordpress.org/?v=3.3.1 By: Andrew Pennebaker http://blog.reverberate.org/2009/01/12/playing-with-llvm/comment-page-1/#comment-1606 Andrew Pennebaker Sun, 06 Mar 2011 10:50:21 +0000 http://blog.reverberate.org/2009/01/12/playing-with-llvm/#comment-1606 Oh, THAT'S how to do a for loop in LLVM. I've chosen LLVM as my first assembly language. Tutorials like this are a dire necessity; Assembly is hard enough without SSA. Hello LLVM http://www.yellosoft.us/hello-llvm Oh, THAT’S how to do a for loop in LLVM. I’ve chosen LLVM as my first assembly language. Tutorials like this are a dire necessity; Assembly is hard enough without SSA.

Hello LLVM
http://www.yellosoft.us/hello-llvm

]]> By: Anton Korobeynikov http://blog.reverberate.org/2009/01/12/playing-with-llvm/comment-page-1/#comment-1479 Anton Korobeynikov Sat, 28 Nov 2009 14:35:56 +0000 http://blog.reverberate.org/2009/01/12/playing-with-llvm/#comment-1479 As mentioned previously - you'll get SSA 'for free'. Write everything in load/store manner, then run opt -std-compile-opts and it will try to optimize your bytecode hard (e.g. convert to ssa form with phi's, etc). As mentioned previously – you’ll get SSA ‘for free’. Write everything in load/store manner, then run opt -std-compile-opts and it will try to optimize your bytecode hard (e.g. convert to ssa form with phi’s, etc).

]]> By: Alex http://blog.reverberate.org/2009/01/12/playing-with-llvm/comment-page-1/#comment-1276 Alex Tue, 03 Mar 2009 00:50:05 +0000 http://blog.reverberate.org/2009/01/12/playing-with-llvm/#comment-1276 The pass Eugene mentioned is called 'mem2reg'. http://llvm.org/docs/tutorial/LangImpl7.html Actually using this method saves a lot of efforts because it's not trivial to find where to insert these phi nodes. The pass Eugene mentioned is called ‘mem2reg’.

http://llvm.org/docs/tutorial/LangImpl7.html

Actually using this method saves a lot of efforts because it’s not trivial to find where to insert these phi nodes.

]]> By: Eugene O'neil http://blog.reverberate.org/2009/01/12/playing-with-llvm/comment-page-1/#comment-1169 Eugene O'neil Wed, 28 Jan 2009 15:43:46 +0000 http://blog.reverberate.org/2009/01/12/playing-with-llvm/#comment-1169 Actually, LLVM has an infinite supply of registers: for example, current_i_value is a register. They just aren't rewritable. That's what "static single assignment" means. An easy way to get around this limitation is to use loads and stores, as you have discovered, but the "correct" way is to use the mysterious and inscrutable phi instruction: http://llvm.org/docs/LangRef.html#i_phi This documentation even provides a simple example of how to write a counting loop, without loads and stores. There is an optimization pass that tries to convert the load-and-store method to the phi method, but it isn't in the register allocation phase. Actually, LLVM has an infinite supply of registers: for example, current_i_value is a register. They just aren’t rewritable. That’s what “static single assignment” means.

An easy way to get around this limitation is to use loads and stores, as you have discovered, but the “correct” way is to use the mysterious and inscrutable phi instruction:

http://llvm.org/docs/LangRef.html#i_phi

This documentation even provides a simple example of how to write a counting loop, without loads and stores.

There is an optimization pass that tries to convert the load-and-store method to the phi method, but it isn’t in the register allocation phase.

]]>