gollem, yes, I put that question to AMD a couple of years ago - in the form of what they were doing for compiler work for Hammer. They answered that they had a group dedicated to making optimized compilers for Hammer. I think that means that AMD developed compiler optimizations in-house and made them available to a select group of partners like Microsoft and Gnu.
...it might depend on whose compiler it is.
Target compiling is as old as the concept of software compiling itself, it goes back to the early 1960s. Remember, mainframes were multi-million dollar behemoths, custom jobs, and each one had to have compiled test code ready when the first hardware became available for first bringup. Hence the first compilations for each hardware platform was done on a previous generation.
With microprocessors the same thing was required. I'm sure the first 8086 code was compiled on an 8080 (8-bit) platform and the first 8080 code was compiled on a 4040 (4-bit) platform before it. I know I assembled 80186 (embedded) code on an 8088-based PC in the mid 1980s. In fact, my first assembler was a 6502 assembler written in FORTH (I was too poor to go out and actually buy tools, and I wanted to learn assembler programming). That assembler could have been hosted on any host processor!
Target compiling, also called cross-compiling, is so intimately a part of software development that I could not imagine anyone making a commercial or freeware x86 compiler without compiler flags to specify the destination processor.
Market Data 
Markets 
AI
