[llvm-commits] [llvm] r102006 - in /llvm/trunk: include/llvm/CodeGen/LinkAllCodegenComponents.h include/llvm/CodeGen/Passes.h lib/CodeGen/CMakeLists.txt lib/CodeGen/RegAllocFast.cpp

[Jakob Stoklund Olesen]

On Apr 21, 2010, at 11:43 AM, Evan Cheng wrote:

> Cool. :-) One thing to watch out for is if it generates significantly worse code and assembly time starts to be an issue (probably not especially given the integrated assembler). Have you checked if codegen is noticeably worse?

I haven't checked except I noted that the time used by the AsmPrinter is the same.

The behaviour only differs from -regalloc=local when all registers are used, and we have to spill. That almost never happens with unoptimized code. Using -regalloc=fast on optimized code is going to be bad.

But yes, depending on the speed of -integrated-as (and disk), it might actually help to do trivial coalescing and load/store folding simply to reduce the bulk of the output.

/jakob