aglew@ccvaxa.UUCP (03/23/88)
I was just rereading the floating point part of the IBM System/370 architecture manual (by the way, does anyone know how Tomasulo handled condition code dependencies in the 360/91? - it had been suggested to me that the floating point instructions didn't modify the condition codes, but this isn't true) and I came across the unnormalized operations, like ADD UNNORMALIZED. I was just wondering if any system exposes all the components of a floating point operation, eg. preshift, add, and postnormalize, and if there are some optimizations that can take advantage of this exposure? Eg. if your compiler can prove that there will be no need for normalization, why run through the normalization stage of the pipeline, etc.?