We have had a number of bugs caused by stale SSA left after RBO, both memory and local SSA. I think we should spend some effort to investigate if we can fully repair SSA after RBO. Either by rebuilding it entirely, or by doing the IR walk to find local occurrences and trying for some more surgical repairing. It seems worth it to me to pay some TP cost when this is leading to actual bad codegen bugs in the product.
A list of issues we have seen from this (purely as a list of test cases we would want to look at if we went with the repair approach):
cc @dotnet/jit-contrib
We have had a number of bugs caused by stale SSA left after RBO, both memory and local SSA. I think we should spend some effort to investigate if we can fully repair SSA after RBO. Either by rebuilding it entirely, or by doing the IR walk to find local occurrences and trying for some more surgical repairing. It seems worth it to me to pay some TP cost when this is leading to actual bad codegen bugs in the product.
A list of issues we have seen from this (purely as a list of test cases we would want to look at if we went with the repair approach):
cc @dotnet/jit-contrib