Alloca addr space value depends on target triple. No matter what alloca
addr space value is used, alloca instruction is still an alloca instruction
and should be treated equally.
Currently in ValueTracking.cpp, in function isKnownNonZero, only alloca
instruction with addr space 0 is treated as nonnull, which causes less
performant ISA for target amdgcn---amdgiz since its alloca addr space
is 5.
This patch fixes that.
What was the original comment about? Is there a chance alloca can be used for malloc?