Re: [PATCH] optabs: Fix up expand_doubleword_shift_condmove for shift_mask == 0 [PR108803]

[Richard Sandiford <richard.sandiford@arm.com>]

Jakub Jelinek <jakub@redhat.com> writes:
> On Mon, Feb 27, 2023 at 07:51:21PM +0000, Richard Sandiford wrote:
>> I think RTL and gimple are different in that respect.
>> SHIFT_COUNT_TRUNCATED's effect on shifts is IMO a bit like
>> CTZ_DEFINED_VALUE_AT_ZERO's effect on CTZ: it enumerates common
>> target-specific behaviour, but doesn't turn invalid/should-not-be-evaluated
>> values into valid values.  Not defining SHIFT_COUNT_TRUNCATED is like
>> defining CTZ_DEFINED_VALUE_AT_ZERO to 0.
>> 
>> The docs say:
>> 
>>   Note that regardless of this macro the ``definedness'' of @code{clz}
>>   and @code{ctz} at zero do @emph{not} extend to the builtin functions
>>   visible to the user.  Thus one may be free to adjust the value at will
>>   to match the target expansion of these operations without fear of
>>   breaking the API@.
>> 
>> So for CTZ this really is an RTL thing, which can leak into gimple
>> through ifns.  I'd argue that the same is true for SHIFT_COUNT_TRUNCATED
>> and conditional shifts like COND_SHL: normal gimple shifts aren't guaranteed
>> to honour SHIFT_COUNT_TRUNCATED, but COND_SHL should be.
>
> I understand that if SHIFT_COUNT_TRUNCATED 1 is defined, then formerly
> out of bounds shift is well defined on RTL. after all, for
> SHIFT_COUNT_TRUNCATED the generic code removes shift count masking as
> redundant, so code without UB in the source could otherwise appear to have
> UB on RTL.
> The question is what happens with SHIFT_COUNT_TRUNCATED 0 or
> C?Z_DEFINED_VALUE_AT_ZERO 0, if encountering the RTL with invalid operand(s)
> is undefined behavior, or simply undefined value but no other side-effects.
> There are many RTL expressions which invoke on invalid values really
> undefined behavior, it can crash the program etc.  The question is if
> out of bounds shifts are like that too or not.  Ditto for CLZ/CTZ.

My argument was that !SHIFT_COUNT_TRUNCATED and
C?Z_DEFINED_VALUE_AT_ZERO==0 mean that the behaviour is undefined
only in the sense that target-independent code doesn't know what
the behaviour is.  !SHIFT_COUNT_TRUNCATED doesn't mean that
target-independent code can assume that out-of-range shift values
invoke program UB (and therefore target-independent code can optimise
shifts on the principle that all shifts are in-range).  Similarly
CTZ_DEFINED_VALUE_AT_ZERO==0 doesn't mean the corresponding thing for CTZ.

If !SHIFT_COUNT_TRUNCATED meant that all out-of-range shifts are UB then:

	    wide_int wop1 = pop1;
	    if (SHIFT_COUNT_TRUNCATED)
	      wop1 = wi::umod_trunc (wop1, GET_MODE_PRECISION (int_mode));
	    else if (wi::geu_p (wop1, GET_MODE_PRECISION (int_mode)))
	      return NULL_RTX;

in simplify_const_binary_operation wouldn't be necessary.  We could
just fold constant shifts in the SHIFT_COUNT_TRUNCATED way for all
values, like wide_int_binop folds all nonnegative shifts on trees.

As I say, arm_emit_coreregs_64bit_shift relies on being able to create
RTL shifts whose counts might be out-of-range (to a certain degree),
because the arm port knows how arm shifts behave.  Treating the
out-of-range shifts as UB would break the DI shift expansions.

Thanks,
Richard