<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0//EN" "http://www.w3.org/TR/REC-html40/strict.dtd"><html><head><meta name="qrichtext" content="1" /><style type="text/css">p, li { white-space: pre-wrap; }</style></head><body style=" font-family:'DejaVu Sans'; font-size:9pt; font-weight:400; font-style:normal;">On Sunday 05 April 2009, you wrote: > On Sat, 4 Apr 2009, Kai-Uwe Behrmann wrote: > > Ah, that is a point I did not really understand. So a library can add a > > (GPU) backend to LLVM at runtime? E.g. thus enabling GPU support in > > gallium3d while OpenGTL has no benefit? > > So a (longterm) plan could be: > > CTL/Shiva -> IL -> GLSL -> IL -> GPU/CPU assembler code ? > > It seems that the LLVM VM just needs to be ready to accept > instructions. Based on what I saw while building LLVM, it seemed like > assembly output for all supported targets is available in one build. > > Not having any experience with LLVM before, I installed it on a > FreeBSD/Xeon system and built GraphicsMagick with it. Then I ran a > benchmark suite which uses OpenMP. Usually native GCC was somewhat > faster than LLVM generated code, but for some floating-point intensive > algorithms, LLVM was faster. I would say that performance was pretty > competitive. To be honest, I wouldn't use LLVM as compiler over GCC now :) Maybe in a year or two, but I do think there is a reason apple hasn't switch for their compiler. But for now, LLVM is a very good platform when you want to write a JIT "VM". -- Cyrille Berger</body></html>