On 09/06/2017 10:43 AM, Richard Henderson wrote:
> On 08/29/2017 05:36 PM, Michael Clark wrote:
>> We’re investigating an issue with redundant sign-extension instructions
>> being emitted with the riscv backend. Firstly I would like to state that
>> riscv is possibly a unique backend with respect to its canonical
>> sign-extended register form due to the following:
>>
>> - POINTERS_EXTEND_UNSIGNED is false, and the canonical form after 32-bit
>> operations on RV64 is sign-extended not zero-extended.
>>
>> - PROMOTE_MODE is defined on RV64 such that SI mode values are promoted to
>> signed DI mode values holding SI mode subregs
>>
>> - RISC-V does not have register aliases for these different modes, rather
>> different instructions are selected to operate on different register widths.
>
> This is identical to Alpha.
>
>> - The 32-bit instructions sign-extend results. e.g. all shifts, add, sub,
>> etc.
>>
>> - Some instructions such as logical ops only have full word width variants
>> (AND, OR, XOR) but these instructions maintain canonical form as there is no
>> horizontal movement of bits.
>
> Alpha only had 32-bit instructions for for add/sub/mul that also sign-extend,
> so there is a little less scope for optimization.
>
>> testcase 1:
>>
>> $ cat bswap.c
>> unsigned bswap(unsigned p) {
>> return ((p >> 24) & 0x000000ff) | ((p << 8 ) & 0x00ff0000) |
>> ((p >> 8 ) & 0x0000ff00) | ((p << 24) & 0xff000000);
>> }
>> $ cat bswap.s
>> bswap2:
>> sllw a3,a0,24
>> srlw a5,a0,24
>> sllw a4,a0,8
>> or a5,a5,a3
>> li a3,16711680
>> and a4,a4,a3
>> or a5,a5,a4
>> li a4,65536
>> add a4,a4,-256
>> srlw a0,a0,8
>> and a0,a0,a4
>> or a0,a5,a0
>> sext.w a0,a0 # redundant
>> ret
>
> I think the easiest solution to this is for combine to notice when IOR has
> operands with non-zero-bits that do not overlap, convert the operation to ADD.
> That allows the final two insns to fold to "addw" and the compiler need do no
> further analysis.
I thought we had combine support for that. I don't think it's ever been
particularly good though. With your alpha background you're probably
more familiar with it than anyone -- IIRC it fell out of removal of low
order bit masking to deal with alignment issues on the Alpha.
I wrote some match.pd patterns to do it in gimple as part of a larger
problem. The work as a whole on that larger problem ultimately didn't
pan out (generated even worse code than what we have on the trunk). But
it might be possible to resurrect those patterns and see if they are
useful independently. My recollection was I was looking at low order
bits only, but the concepts were general enough.
>
>> $ cat rshift.c
>> unsigned rs24(unsigned rs1) { return rs1 >> 24; }
>> $ cat rshift.s
>> rs24:
>> sllw a0,a0,24
>> sext.w a0,a0 # redundant
>> ret
>
> That seems like a missing check somewhere (combine? simplify-rtx? both?) for
> SUBREG_PROMOTED_SIGNED_P. Since Alpha didn't have a 32-bit shift you're in
> new
> territory for this one.
Yea. I'd also expect zero/nonzero bits tracking in combine to catch
this. Shouldn't the sign bit be known to be zero after the shift which
makes the extension redundant regardless of the SUBREG_PROMOTED flag?
jeff
