D79794 changed INLINEASM_BR from being a termintor to not.
While reading though SelectionDAGBuilder::visitInlineAsm, I noticed a
comment regarding INLINEASM_BR being a terminator, which seemed stale.
Removing the code that has an assumption does change the placement of a
TokenFactor in one test, but does not change the final output. The
assembler generated is ultimately the same.
This means the copies to virtual register %0 and %1 are no longer guaranteed to happen before the inlineasm_br block is executed. Since I think one of those is used by the fail basic block, that seems wrong.