ARC: Make ARC bitops "safer" (add anti-optimization)
ARCompact/ARCv2 ISA provide that any instructions which deals with bitpos/count operand ASL, LSL, BSET, BCLR, BMSK .... will only consider lower 5 bits. i.e. auto-clamp the pos to 0-31. ARC Linux bitops exploited this fact by NOT explicitly masking out upper bits for @nr operand in general, saving a bunch of AND/BMSK instructions in generated code around bitops. While this micro-optimization has worked well over years it is NOT safe as shifting a number with a value, greater than native size is "undefined" per "C" spec. So as it turns outm EZChip ran into this eventually, in their massive muti-core SMP build with 64 cpus. There was a test_bit() inside a loop from 63 to 0 and gcc was weirdly optimizing away the first iteration (so it was really adhering to standard by implementing undefined behaviour vs. removing all the iterations which were phony i.e. (1 << [63..32]) | for i = 63 to 0 | X = ( 1 << i ) | if X == 0 | continue So fix the code to do the explicit masking at the expense of generating additional instructions. Fortunately, this can be mitigated to a large extent as gcc has SHIFT_COUNT_TRUNCATED which allows combiner to fold masking into shift operation itself. It is currently not enabled in ARC gcc backend, but could be done after a bit of testing. Fixes STAR 9000866918 ("unsafe "undefined behavior" code in kernel") Reported-by: Noam Camus <noamc@ezchip.com> Cc: <stable@vger.kernel.org> Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Showing
Please register or sign in to comment