> Am 10.06.2025 um 15:51 schrieb Tamar Christina <tamar.christ...@arm.com>:
>
>
>>
>> -----Original Message-----
>> From: Richard Biener <rguent...@suse.de>
>> Sent: Tuesday, June 10, 2025 2:12 PM
>> To: Tamar Christina <tamar.christ...@arm.com>
>> Cc: gcc-patches@gcc.gnu.org; Richard Sandiford <richard.sandif...@arm.com>;
>> nd <n...@arm.com>
>> Subject: RE: [PATCH 1/3]middle-end: support vec_cbranch_any and
>> vec_cbranch_all [PR118974]
>>
>> On Mon, 9 Jun 2025, Tamar Christina wrote:
>>
>>>> -----Original Message-----
>>>> From: Richard Biener <rguent...@suse.de>
>>>> Sent: Monday, June 9, 2025 10:30 AM
>>>> To: Tamar Christina <tamar.christ...@arm.com>
>>>> Cc: gcc-patches@gcc.gnu.org; Richard Sandiford
>> <richard.sandif...@arm.com>;
>>>> nd <n...@arm.com>
>>>> Subject: Re: [PATCH 1/3]middle-end: support vec_cbranch_any and
>>>> vec_cbranch_all [PR118974]
>>>>
>>>>> On Mon, 9 Jun 2025, Tamar Christina wrote:
>>>>>
>>>>>> This patch introduces two new vector cbranch optabs vec_cbranch_any and
>>>>>> vec_cbranch_all.
>>>>>>
>>>>>> To explain why we need two new optabs let me explain the current cbranch
>>> and
>>>>> its
>>>>>> limitations and what I'm trying to optimize. So sorry for the long
>>>>>> email, but I
>>>>>> hope it explains why I think we want new optabs.
>>>>>>
>>>>>> Today cbranch can be used for both vector and scalar modes. In both
>>>>>> these
>>>>>> cases it's intended to compare boolean values, either scalar or vector.
>>>>>>
>>>>>> The optab documentation does not however state that it can only handle
>>>>>> comparisons against 0. So many targets have added code for the vector
>>> variant
>>>>>> that tries to deal with the case where we branch based on two non-zero
>>>>>> registers.
>>>>>>
>>>>>> However this code can't ever be reached because the cbranch expansion
>>>>>> only
>>>>> deals
>>>>>> with comparisons against 0 for vectors. This is because for vectors the
>>>>>> rest of
>>>>>> the compiler has no way to generate a non-zero comparison. e.g. the
>>> vectorizer
>>>>>> will always generate a zero comparison, and the C/C++ front-ends won't
>>> allow
>>>>>> vectors to be used in a cbranch as it expects a boolean value. ISAs
>>>>>> like SVE
>>>>>> work around this by requiring you to use an SVE PTEST intrinsics which
>>>>>> results
>>>>>> in a single scalar boolean value that represents the flag values.
>>>>>>
>>>>>> e.g. if (svptest_any (..))
>>>>>>
>>>>>> The natural question is why do we not at expand time then rewrite the
>>>>> comparison
>>>>>> to a non-zero comparison if the target supports it.
>>>>>>
>>>>>> The reason is we can't safely do so. For an ANY comparison (e.g. != b)
>>>>>> this is
>>>>>> trivial, but for an ALL comparison (e.g. == b) we would have to flip both
>>> branch
>>>>>> and invert the value being compared. i.e. we have to make it a != b
>>> comparison.
>>>>>>
>>>>>> But in emit_cmp_and_jump_insns we can't flip the branches anymore
>>> because
>>>>> they
>>>>>> have already been lowered into a fall through branch (PC) and a label,
>>>>>> ready
>>> for
>>>>>> use in an if_then_else RTL expression.
>>>>>>
>>>>>> Additionally as mentioned before, cbranches expect the values to be
>>>>>> masks,
>>> not
>>>>>> values. This kinda works out if you XOR the values, but for FP vectors
>>>>>> you'd
>>>>>> need to know what equality means for the FP format. i.e. it's possible
>>>>>> for
>>>>>> IEEE 754 values but It's not immediately obvious if this is true for all
>>>>>> formats.
>>>>>>
>>>>>> Now why does any of this matter? Well there are two optimizations we
>>>>>> want
>>> to
>>>>> be
>>>>>> able to do.
>>>>>>
>>>>>> 1. Adv. SIMD does not support a vector !=, as in there's no instruction
>>>>>> for it.
>>>>>> For both Integer and FP vectors we perform the comparisons as EQ and
>>> then
>>>>>> invert the resulting mask. Ideally we'd like to replace this with
>>>>>> just a XOR
>>>>>> and the appropriate branch.
>>>>>>
>>>>>> 2. When on an SVE enabled system we would like to use an SVE compare +
>>>>> branch
>>>>>> for the Adv. SIMD sequence which could happen due to cost modelling.
>>>>> However
>>>>>> we can only do so based on if we know that the values being compared
>>> against
>>>>>> are the boolean masks. This means we can't really use combine to do
>>>>>> this
>>>>>> because combine would have to match the entire sequence including the
>>>>>> vector comparisons because at RTL we've lost the information that
>>>>>> VECTOR_BOOLEAN_P would have given us. This sequence would be too
>>> long
>>>>> for
>>>>>> combine to match due to it having to match the compare + branch
>>>>>> sequence
>>>>>> being generated as well. It also becomes a bit messy to match ANY and
>>>>>> ALL
>>>>>> sequences.
>>>>>
>>>>> I guess I didn't really understand the above, esp. why all of this is
>>>>> not recoverable on RTL, so I'll state what I think
>>>>> cbranch is supposed to do and then ask some question on the proposal.
>>>>>
>>>
>>> Quite simply the log chain is too long. SVE has flag setting comparisons,
>>> so in a
>>> single instruction we can perform the final comparison. (which doesn't need
>>> to
>>> be restricted to EQ or NE) and then using the appropriate condition code
>>> decide
>>> if we have an ANY or ALL comparison. So the sequence to match is very long.
>>>
>>> That is to say, in gimple we have
>>>
>>> a = b > c
>>> if (a != 0)
>>> ...
>>> and SVE doesn't need the indirection. Conceptually this makes it
>>>
>>> if (any (b > c))
>>> ...
>>>
>>> For SVE the cbranch expands to a predicate test, which also sets the flag
>>> again.
>>> this duplication in flag setting is later then removed by various patterns
>>> in
>>> the backend. This is all fine, but where it gets difficult is trying to
>>> replace the
>>> Adv. SIMD Compare + branch with an SVE one. To do this it would have to
>>> match
>>> the entirety of
>>>
>>> a = b > c
>>> if (a != 0)
>>>
>>> which in RTL is
>>>
>>> (insn 66 65 67 4 (set (reg:V16QI 185 [ mask_patt_9.20_138 ])
>>> (neg:V16QI (gtu:V16QI (reg:V16QI 183)
>>> (reg:V16QI 184)))) "/app/example.cpp":29:7 -1
>>> (nil))
>>> (insn 67 66 68 4 (set (reg:V4SI 186)
>>> (unspec:V4SI [
>>> (subreg:V4SI (reg:V16QI 185 [ mask_patt_9.20_138 ]) 0)
>>> (subreg:V4SI (reg:V16QI 185 [ mask_patt_9.20_138 ]) 0)
>>> ] UNSPEC_UMAXV)) "/app/example.cpp":29:7 -1
>>> (nil))
>>> (insn 68 67 69 4 (set (reg:V16QI 185 [ mask_patt_9.20_138 ])
>>> (subreg:V16QI (reg:V4SI 186) 0)) "/app/example.cpp":29:7 -1
>>> (nil))
>>> (insn 69 68 70 4 (set (reg:DI 187)
>>> (subreg:DI (reg:V16QI 185 [ mask_patt_9.20_138 ]) 0))
>> "/app/example.cpp":29:7 -1
>>> (nil))
>>> (insn 70 69 71 4 (set (reg:CC 66 cc)
>>> (compare:CC (reg:DI 187)
>>> (const_int 0 [0]))) "/app/example.cpp":29:7 -1
>>> (nil))
>>> (jump_insn 71 70 72 4 (set (pc)
>>> (if_then_else (ne (reg:CC 66 cc)
>>> (const_int 0 [0]))
>>> (label_ref:DI 397)
>>> (pc))) "/app/example.cpp":29:7 19 {condjump}
>>>
>>> And any intermediate "submatch" isn't really valid. The sequence would need
>>> to
>> be matched
>>> as a whole or not at all.
>>>
>>> The second problem is the lack of Adv. SIMD is having a vector !=
>>> comparison.
>>>
>>> This means there's an additional "NOT" in inverting the mask. This
>>> operation can
>> be optimized
>>> in three ways.
>>>
>>> 1. it can be converted from a MAX reduction into a MIN reduction and then
>>> the
>> NOT can be removed.
>>> The branch would have to be inverted.
>>> 2. The data vectors can be XOR'ed together, and inequality means a
>>> non-zero bit
>> anywhere in the result.
>>> 3. Use an SVE comparison, which does have != on vectors.
>>>
>>> This chain is again quite long to match, and we can't incrementally match
>>> it since
>> it's only value if we know
>>> we're using the results in a zero or non-zero comparison.
>>>
>>>> cbranch is supposed to branch on the result of a comparison,
>>>> being it on vector or scalar operands, the result of the comparison is
>>>> a scalar, or rather in full generality, a (scalar) condition-code.
>>>> Given it uses a COMPARE RTX as comparison operator we can currently
>>>> only handle whole vector equality/inequality which map to all/any.
>>>
>>> Correct, and normally this is fine, but for SVE we have support for what's
>>> essentially "vector" condition code. This is why I also named the new
>>> optabs
>>> vec_cbranch.
>>>
>>> The biggest difference is that instead of requiring the reduction to a
>>> scalar
>>> It allows the branching on vector elements comparisons *directly*. Which
>>> makes
>>> it much easier to do the two fixes above.
>>>
>>>>
>>>>> To handle these two cases I propose the new optabs vec_cbranch_any and
>>>>> vec_cbranch_all that expect the operands to be values, not masks, and the
>>>>> comparison operation is the comparison being performed. The current
>> cbranch
>>>>> optab can't be used here because we need to be able to see both comparison
>>>>> operators (for the boolean branch and the data compare branch).
>>>>>
>>>>> The initial != 0 or == 0 is folded away into the name as _any and _all
>>>>> allowing
>>>>> the target to easily do as it wishes.
>>>>>
>>>>> I have intentionally chosen to require cbranch_optab to still be the main
>>>>> one.
>>>>> i.e. you can't have vec_cbranch_any/vec_cbranch_all without cbranch
>> because
>>>>> these two are an expand time only construct. I've also chosen them to be
>>>>> this
>>>>> such that there's no change needed to any other passes in the middle-end.
>>>>>
>>>>> With these two optabs it's trivial to implement the two optimization I
>> described
>>>>> above. A target expansion hook is also possible but optabs felt more
>>>>> natural
>>>>> for the situation.
>>>>
>>>> What's the RTL the optabs expand to for SVE/AdvSIMD? How do you
>>>> represent the any/all vector compare difference? Looking at 3/3 I see
>>>> you use UNSPECs.
>>>>
>>>
>>> You mean in the expand RTL? That's just so I can make the iterators. The
>> expansion is:
>>>
>>> (insn 31 30 32 5 (parallel [
>>> (set (reg:VNx4BI 133 [ mask_patt_13.20_41 ])
>>> (unspec:VNx4BI [
>>> (reg:VNx4BI 134)
>>> (const_int 1 [0x1])
>>> (gtu:VNx4BI (reg:VNx4QI 112 [ vect__1.16 ])
>>> (reg:VNx4QI 128))
>>> ] UNSPEC_PRED_Z))
>>> (clobber (reg:CC_NZC 66 cc))
>>> ]) "/app/example.cpp":29:7 -1
>>> (nil))
>>> (insn 32 31 33 5 (set (reg:VNx4BI 136 [ vec_mask_and_42 ])
>>> (and:VNx4BI (reg:VNx4BI 133 [ mask_patt_13.20_41 ])
>>> (reg:VNx4BI 111 [ loop_mask_36 ]))) "/app/example.cpp":29:7 -1
>>> (nil))
>>> (insn 33 32 34 5 (set (reg:VNx16BI 137)
>>> (const_vector:VNx16BI repeat [
>>> (const_int 1 [0x1])
>>> (const_int 0 [0]) repeated x3
>>> ])) "/app/example.cpp":29:7 -1
>>> (nil))
>>> (insn 34 33 35 5 (set (reg:CC_NZC 66 cc)
>>> (unspec:CC_NZC [
>>> (reg:VNx16BI 137)
>>> (subreg:VNx4BI (reg:VNx16BI 137) 0)
>>> (const_int 1 [0x1])
>>> (reg:VNx4BI 136 [ vec_mask_and_42 ])
>>> ] UNSPEC_PTEST)) "/app/example.cpp":29:7 -1
>>> (nil))
>>> (jump_insn 35 34 36 5 (set (pc)
>>> (if_then_else (eq (reg:CC_NZC 66 cc)
>>> (const_int 0 [0]))
>>> (label_ref 41)
>>> (pc))) "/app/example.cpp":29:7 -1
>>> (int_list:REG_BR_PROB 1014686028 (nil))
>>> -> 41)
>>>
>>> Notice how insn 31 which is the vector compare sets CC_NZC.
>>
>> That's a "scalar" condition code?
>>
>
> Sure, depends on how you view it. It's set by a vector instruction.
> SVE doesn't need the reduction to scalar. As In the predicate
> result is unused and marked as dead. I think that's the core distinction
> here and one you pointed out below. This is about not needing the
> reduction for branching. Or rather the result of the compare isn't
> needed for the branch itself.
>
>>> Insn 32 is the masking of the compare, it'll be moved into insn 31 by
>>> combine replacing the all true predicate.
>>
>> So the (reg:VNx4BI 133) is essentially unused, but it's required to ...
>>
>>> Insn 34 is a flag reshaping operation. In case the predicates don't
>>> have the same element width the CC reg values needs to be
>>> reshaped (in case you're doing an eg. b.first (branch if first element
>>> matches)
>> etc).
>>
>> ... do this, which produces the actual CC_NZC we jump on.
>
> No, it's only needed because the codegen above is generated from the current
> trunk which has a different instruction doing the data comparison.
>
> 133 is not needed at all. And is only there because insn 34 is what is mapped
> to
> a cbranch. So this is the code we have today because there's no way for us to
> not emit the ptest initially ☹
>
> And our current ptest also emits CC_NZC instead of CC_Z, while ptest does set
> these
> for cbranch we don't need them. So I'm changing this to CC_Z already.
>
>>
>>> However because insn 35 only cares about 0 or !0 the shape doesn't matter.
>>> So insn 34 will be eliminated by combine.
>>>
>>> This leaves insn and 31 and 35. Essentially only the compare + branch,
>>> directly on the vector elements, which is what vec_cbranch is now
>>> representing.
>>>
>>> The optab allows us to generate the paradoxical subregs to convert
>>> Adv. SIMD registers into SVE ones, and the predicate mask to make
>>> the operation safe.
>>>
>>> Without this, we have to rely on combine which (even if it could match
>>> such long sequences) would have to not just replace the adv. simd + b
>>> with an SVE + b, but also do all the predicate optimization in the one
>>> single pattern.
>>>
>>>> What kind of compares do you expect us to generate? Any besides
>>>> equality compares?
>>>
>>> Any compare is possible. Since SVE vector comparisons take all 6
>>> comparison operators, there's also floating point versions and
>>> unordered variants. Basically in the ISA vectors are first class
>>> control flow elements.
>>>
>>>> Why would we not allow cbranch to take an
>>>> operand 0 that specifies the desired reduction?
>>>
>>> That's certainly possible, but that's just going to make cbranch itself
>>> more complicated to understand. I chose to go with _any and _all
>>> since as you say, for the branch comparison itself only != 0 and == 0
>>> makes sense. So cbranch would need to get two operators in this case
>>> one for the branch comparison and one for the data comparison.
>>>
>>> i.e. if expanding
>>>
>>> a = b > c
>>> if (a != 0)
>>>
>>> branch would get
>>> (set (reg a)
>>> (gt (reg b) (reg c))
>>> (clobber cc_nz)
>>> (ne (reg a) 0)
>>> (if_then_else ...)
>>>
>>> I basically opted for a simpler change, since you can only have != 0 and == >>> 0
>>> this is just encoded in the name "any" and "all" so we don't have to send
>>> the
>>> dummy operators along.
>>
>> Well, if we'd introduce ANY_LT_EXPR, ANY_NE_EXPR, ... ALL_LT_EXPR,
>> ALL_NE_EXPR ... then we can have on GIMPLE
>>
>> if (a ANY_GT c)
>>
>> and no need for a separate compare and compare-and-branch.
>
> Yes, new comparison codes would work here too.
> But then I'll need 12 of them.. But could do so if you prefer 😊
>
> But you're right, conceptually my new optabs are introducing an
> expand time only representation of these 12 operators.
>
>>
>>>> We could
>>>> have (any:CC (eq:V4BI (reg:V4SI) (reg:V4SI))) or so, or alternatively
>>>> (more intrusive) have (any_eq:CC (reg:V4SI) (reg:V4SI))? That
>>>> would also allow you to have a proper RTL representation rather
>>>> than more and more UNSEPCs.
>>>
>>> We do have proper RTL representation though..
>>
>> You have
>>
>>> (insn 31 30 32 5 (parallel [
>>> (set (reg:VNx4BI 133 [ mask_patt_13.20_41 ])
>>> (unspec:VNx4BI [
>>> (reg:VNx4BI 134)
>>> (const_int 1 [0x1])
>>> (gtu:VNx4BI (reg:VNx4QI 112 [ vect__1.16 ])
>>> (reg:VNx4QI 128))
>>> ] UNSPEC_PRED_Z))
>>> (clobber (reg:CC_NZC 66 cc))
>>
>> that's not
>>
>> (set (reg:CC_NZC 66 cc) (any_gtu:CC_NZC (reg:VNx4QI 112 [ vect__1.16 ])
>> ((reg:VNx4QI 128)))
>>
>> so nothing combine or simplify-rtx can work with (because of the UNSPEC).
>>
>
> Sure, I don’t think any_gtu provides any value here though. this information
> "any" or "all" is a property of the use of the result not the comparison in
> RTL.
But that was the gist of the question- so the compare does _not_ set a CC that
you can directly branch on but there’s a use insn of <whatever> that either
all- or any-reduces <whatever> to a CC? I’m now confused…
> So the "none" is here
>
> (jump_insn 35 34 36 5 (set (pc)
> (if_then_else (eq (reg:CC_NZC 66 cc)
> (const_int 0 [0]))
> (label_ref 41)
> (pc))) "/app/example.cpp":29:7 -1
> (int_list:REG_BR_PROB 1014686028 (nil))
>
> So I don’t think that the ANY or ALL was ever a property of the compare,
> But rather what you do with it.
> Which is why today in gimple we have the != 0 or == 0 in the if no?
>
>> That said, I wanted to know whether SVE/NEON can do {any,all}_<compare>
>> ops on vectors, setting the CC flag register you can then do a
>> conditional jump on. And that seems to be the case and what you
>> are after.
>>
>
> Yes, SVE can indeed for integer comparisons. Floating point needs the
> ptest.
>
> Also other operations can do this too, for instance the MATCH instruction
> being worked on now etc
>
>> But that's exactly what the cbranch optab is supposed to facilitate.
>> What we're lacking is communication of the all vs. any. Your choice
>> is to do that by using different optabs, 'vec_cbranch_any' vs.
>> 'vec_cbranch_all'. I see how thats "easy", you just have to adjust
>> RTL expansion.
>>
>
> Indeed, and the benefit of this is that we can expand directly to the
> optimal sequence, rather than relying on later passes to eliminate
> the unneeded predicate operations.
>
>> On GIMPLE you could argue we have everything, we can support
>>
>> if (_1 ==/!= {-1, -1, ... })
>>
>> to match vec_cbranch_all/any for an inverted compare?
>>
>
> Yes, so my patch at expand time just checks if _1 is a compare
> And if so, reads the arguments of _1 and based on ==/!= it
> chooses vec_cbranch_all/any and passes the argument of the
> compare to the optab. It can't handle {-1,-1,-1,...} today because
> the expansion happens quite low during expand, so we can no
> longer flip the branches to make them a normal != 0 compare.
>
> I hadn't worried about this since it would just fall back to cbranch
> and we no longer generate this representation.
>
>> I see one of your issues is that the AND does not set flags, right?
>> You do need the ptest once you have that?
>>
>
> The AND is just because the middle-end can only generate an unpredicated
> compare. And so we rely on combine to push the predicate inside eventually.
>
> So we don't need the AND itself to set flags, because for early exits all we
> care
> About is CC_Z, and the AND can have an effect on N and C but not Z.
>
> So for pure SVE I can eliminate it (see patch 2/3) but for the case where we
> have
> to generate an SVE compare for Adv. SIMD (as it does the flag setting) means
> that
> we have to match everything from the compare to the if_then_else. You need
> the compare since you have to replace it with the SVE variant, and you need
> the
> if_then_else to realize if you have any or all.
>
> The optab allows us to avoid this by just emitting the SVE commit directly,
> and
> completely avoid the AND by generating a predicated compare.
>
> So I am trying to fix some of the optimization in RTL, but for some cases
> expanding
> correctly is the most reliable way.
>
> I would have though that the two optabs would be preferrable to 12 new
> comparison
> operators, but could make those work too 😊
>
> Thanks,
> Tamar
>
>>
>> +@cindex @code{cbranch_any@var{mode}4} instruction pattern
>> +@item @samp{cbranch_any@var{mode}4}
>> +Conditional branch instruction combined with a compare instruction on
>> vectors
>> +where it is required that at least one of the elementwise comparisons of
>> the
>> +two input vectors is true.
>>
>> That sentence is a bit odd. Operand 0 is a comparison operator that
>> is applied to all vector elements. When at least one comparison evaluates
>> to true the branch is taken.
>>
>> +Operand 0 is a comparison operator. Operand 1 and operand 2 are the
>> +first and second operands of the comparison, respectively. Operand 3
>> +is the @code{code_label} to jump to.
>>
>> Tamar Christina
>>
>> AttachmentsJun 9, 2025, 8:03 AM (1 day ago)
>>
>> to gcc-patches, nd, rguenther
>> This patch introduces two new vector cbranch optabs vec_cbranch_any and
>> vec_cbranch_all.
>>
>> To explain why we need two new optabs let me explain the current cbranch
>> and its
>> limitations and what I'm trying to optimize. So sorry for the long email,
>> but I
>> hope it explains why I think we want new optabs.
>>
>> Today cbranch can be used for both vector and scalar modes. In both these
>> cases it's intended to compare boolean values, either scalar or vector.
>>
>> The optab documentation does not however state that it can only handle
>> comparisons against 0. So many targets have added code for the vector
>> variant
>> that tries to deal with the case where we branch based on two non-zero
>> registers.
>>
>> However this code can't ever be reached because the cbranch expansion only
>> deals
>> with comparisons against 0 for vectors. This is because for vectors the
>> rest of
>> the compiler has no way to generate a non-zero comparison. e.g. the
>> vectorizer
>> will always generate a zero comparison, and the C/C++ front-ends won't
>> allow
>> vectors to be used in a cbranch as it expects a boolean value. ISAs like
>> SVE
>> work around this by requiring you to use an SVE PTEST intrinsics which
>> results
>> in a single scalar boolean value that represents the flag values.
>>
>> e.g. if (svptest_any (..))
>>
>> The natural question is why do we not at expand time then rewrite the
>> comparison
>> to a non-zero comparison if the target supports it.
>>
>> The reason is we can't safely do so. For an ANY comparison (e.g. != b)
>> this is
>> trivial, but for an ALL comparison (e.g. == b) we would have to flip both
>> branch
>> and invert the value being compared. i.e. we have to make it a != b
>> comparison.
>>
>> But in emit_cmp_and_jump_insns we can't flip the branches anymore because
>> they
>> have already been lowered into a fall through branch (PC) and a label,
>> ready for
>> use in an if_then_else RTL expression.
>>
>> Additionally as mentioned before, cbranches expect the values to be masks,
>> not
>> values. This kinda works out if you XOR the values, but for FP vectors
>> you'd
>> need to know what equality means for the FP format. i.e. it's possible
>> for
>> IEEE 754 values but It's not immediately obvious if this is true for all
>> formats.
>>
>> Now why does any of this matter? Well there are two optimizations we want
>> to be
>> able to do.
>>
>> 1. Adv. SIMD does not support a vector !=, as in there's no instruction
>> for it.
>> For both Integer and FP vectors we perform the comparisons as EQ and
>> then
>> invert the resulting mask. Ideally we'd like to replace this with just
>> a XOR
>> and the appropriate branch.
>>
>> 2. When on an SVE enabled system we would like to use an SVE compare +
>> branch
>> for the Adv. SIMD sequence which could happen due to cost modelling.
>> However
>> we can only do so based on if we know that the values being compared
>> against
>> are the boolean masks. This means we can't really use combine to do
>> this
>> because combine would have to match the entire sequence including the
>> vector comparisons because at RTL we've lost the information that
>> VECTOR_BOOLEAN_P would have given us. This sequence would be too long
>> for
>> combine to match due to it having to match the compare + branch
>> sequence
>> being generated as well. It also becomes a bit messy to match ANY and
>> ALL
>> sequences.
>>
>> To handle these two cases I propose the new optabs vec_cbranch_any and
>> vec_cbranch_all that expect the operands to be values, not masks, and the
>> comparison operation is the comparison being performed. The current
>> cbranch
>> optab can't be used here because we need to be able to see both comparison
>> operators (for the boolean branch and the data compare branch).
>>
>> The initial != 0 or == 0 is folded away into the name as _any and _all
>> allowing
>> the target to easily do as it wishes.
>>
>> I have intentionally chosen to require cbranch_optab to still be the main
>> one.
>> i.e. you can't have vec_cbranch_any/vec_cbranch_all without cbranch
>> because
>> these two are an expand time only construct. I've also chosen them to be
>> this
>> such that there's no change needed to any other passes in the middle-end.
>>
>> With these two optabs it's trivial to implement the two optimization I
>> described
>> above. A target expansion hook is also possible but optabs felt more
>> natural
>> for the situation.
>>
>> I.e. with them we can now generate
>>
>> .L2:
>> ldr q31, [x1, x2]
>> add v29.4s, v29.4s, v25.4s
>> add v28.4s, v28.4s, v26.4s
>> add v31.4s, v31.4s, v30.4s
>> str q31, [x1, x2]
>> add x1, x1, 16
>> cmp x1, 2560
>> beq .L1
>> .L6:
>> ldr q30, [x3, x1]
>> cmpeq p15.s, p7/z, z30.s, z27.s
>> b.none .L2
>>
>> and easily prove it correct.
>>
>> Bootstrapped Regtested on aarch64-none-linux-gnu,
>> arm-none-linux-gnueabihf, x86_64-pc-linux-gnu
>> -m32, -m64 and no issues.
>>
>> Ok for master?
>>
>> Thanks,
>> Tamar
>>
>> gcc/ChangeLog:
>>
>> PR target/118974
>> * optabs.cc (prepare_cmp_insn): Refactor to take optab to check
>> for
>> instead of hardcoded cbranch.
>> (emit_cmp_and_jump_insns): Try to emit a vec_cbranch if supported.
>> * optabs.def (vec_cbranch_any_optab, vec_cbranch_all_optab): New.
>> * doc/md.texi (cbranch_any@var{mode}4, cbranch_any@var{mode}4):
>> Document
>> them.
>>
>> ---
>> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
>> index
>> f6314af46923beee0100a1410f089efd34d7566d..7dbfbbe1609a196b0834d458
>> b26f61904eaf5b24
>> 100644
>> --- a/gcc/doc/md.texi
>> +++ b/gcc/doc/md.texi
>> @@ -7622,6 +7622,24 @@ Operand 0 is a comparison operator. Operand 1 and
>> operand 2 are the
>> first and second operands of the comparison, respectively. Operand 3
>> is the @code{code_label} to jump to.
>>
>> +@cindex @code{cbranch_any@var{mode}4} instruction pattern
>> +@item @samp{cbranch_any@var{mode}4}
>> +Conditional branch instruction combined with a compare instruction on
>> vectors
>> +where it is required that at least one of the elementwise comparisons of
>> the
>> +two input vectors is true.
>> +Operand 0 is a comparison operator. Operand 1 and operand 2 are the
>> +first and second operands of the comparison, respectively. Operand 3
>> +is the @code{code_label} to jump to.
>> +
>> +@cindex @code{cbranch_all@var{mode}4} instruction pattern
>> +@item @samp{cbranch_all@var{mode}4}
>> +Conditional branch instruction combined with a compare instruction on
>> vectors
>> +where it is required that at all of the elementwise comparisons of the
>> +two input vectors are true.
>> +Operand 0 is a comparison operator. Operand 1 and operand 2 are the
>> +first and second operands of the comparison, respectively. Operand 3
>> +is the @code{code_label} to jump to.
>> +
>> @cindex @code{jump} instruction pattern
>> @item @samp{jump}
>> A jump inside a function; an unconditional branch. Operand 0 is the
>> diff --git a/gcc/optabs.cc b/gcc/optabs.cc
>> index
>> 0a14b1eef8a5795e6fd24ade6da55841696315b8..77d5e6ee5d26ccda3d39126
>> 5bc45fd454530cc67
>> 100644
>> --- a/gcc/optabs.cc
>> +++ b/gcc/optabs.cc
>> @@ -4418,6 +4418,9 @@ can_vec_extract_var_idx_p (machine_mode vec_mode,
>> machine_mode extr_mode)
>>
>> *PMODE is the mode of the inputs (in case they are const_int).
>>
>> + *OPTAB is the optab to check for OPTAB_DIRECT support. Defaults to
>> + cbranch_optab.
>> +
>> This function performs all the setup necessary so that the caller only
>> has
>> to emit a single comparison insn. This setup can involve doing a
>> BLKmode
>> comparison or emitting a library call to perform the comparison if no
>> insn
>> @@ -4429,7 +4432,7 @@ can_vec_extract_var_idx_p (machine_mode vec_mode,
>> machine_mode extr_mode)
>> static void
>> prepare_cmp_insn (rtx x, rtx y, enum rtx_code comparison, rtx size,
>> int unsignedp, enum optab_methods methods,
>> - rtx *ptest, machine_mode *pmode)
>> + rtx *ptest, machine_mode *pmode, optab
>> optab=cbranch_optab)
>> {
>> machine_mode mode = *pmode;
>> rtx libfunc, test;
>> @@ -4547,7 +4550,7 @@ prepare_cmp_insn (rtx x, rtx y, enum rtx_code
>> comparison, rtx size,
>> FOR_EACH_WIDER_MODE_FROM (cmp_mode, mode)
>> {
>> enum insn_code icode;
>> - icode = optab_handler (cbranch_optab, cmp_mode);
>> + icode = optab_handler (optab, cmp_mode);
>> if (icode != CODE_FOR_nothing
>> && insn_operand_matches (icode, 0, test))
>> {
>> @@ -4580,7 +4583,7 @@ prepare_cmp_insn (rtx x, rtx y, enum rtx_code
>> comparison, rtx size,
>> if (comparison == UNORDERED && rtx_equal_p (x, y))
>> {
>> prepare_cmp_insn (x, y, UNLT, NULL_RTX, unsignedp, OPTAB_WIDEN,
>> - ptest, pmode);
>> + ptest, pmode, optab);
>> if (*ptest)
>> return;
>> }
>> @@ -4632,7 +4635,7 @@ prepare_cmp_insn (rtx x, rtx y, enum rtx_code
>> comparison, rtx size,
>>
>> *pmode = ret_mode;
>> prepare_cmp_insn (x, y, comparison, NULL_RTX, unsignedp, methods,
>> - ptest, pmode);
>> + ptest, pmode, optab);
>> }
>>
>> return;
>> @@ -4816,12 +4819,68 @@ emit_cmp_and_jump_insns (rtx x, rtx y, enum
>> rtx_code comparison, rtx size,
>> the target supports tbranch. */
>> machine_mode tmode = mode;
>> direct_optab optab;
>> - if (op1 == CONST0_RTX (GET_MODE (op1))
>> - && validate_test_and_branch (val, &test, &tmode,
>> - &optab) != CODE_FOR_nothing)
>> + if (op1 == CONST0_RTX (GET_MODE (op1)))
>> {
>> - emit_cmp_and_jump_insn_1 (test, tmode, label, optab, prob, true);
>> - return;
>> + if (validate_test_and_branch (val, &test, &tmode,
>> + &optab) != CODE_FOR_nothing)
>> + {
>> + emit_cmp_and_jump_insn_1 (test, tmode, label, optab, prob,
>> true);
>> + return;
>> + }
>> +
>> + /* If we are comparing equality with 0, check if VAL is another
>> equality
>> + comparison and if the target supports it directly. */
>> + if (val && TREE_CODE (val) == SSA_NAME
>> + && VECTOR_BOOLEAN_TYPE_P (TREE_TYPE (val))
>> + && (comparison == NE || comparison == EQ))
>> + {
>> + auto def_stmt = SSA_NAME_DEF_STMT (val);
>>
>> I think you may only look at the definition via get_def_for_expr (aka
>> TER), and the result may be NULL.
>>
>> + if ((icode = optab_handler (optab, mode2))
>> + != CODE_FOR_nothing
>>
>> can we perform this check before expanding the operands please?
>> Can we use vector_compare_rtx ().
>>
>> Somehow I feel this belongs in a separate function like
>> validate_test_and_branch.
>>
>> Richard.
>>
>>> I think in the patch you've missed that the RTL from the expand is
>>> overwritten.
>>>
>>> Basically the new optabs are doing essentially the same thing that you
>>> are suggesting, but without needing the dummy second comparison_operator.
>>>
>>> Notice how in the patch the comparison operator is the general
>>>
>>> aarch64_comparison_operator
>>>
>>> and not the stricter aarch64_equality_operator.
>>>
>>> Thanks,
>>> Tamar
>>>>
>>>> Richard.
>>>>
>>>>> I.e. with them we can now generate
>>>>>
>>>>> .L2:
>>>>> ldr q31, [x1, x2]
>>>>> add v29.4s, v29.4s, v25.4s
>>>>> add v28.4s, v28.4s, v26.4s
>>>>> add v31.4s, v31.4s, v30.4s
>>>>> str q31, [x1, x2]
>>>>> add x1, x1, 16
>>>>> cmp x1, 2560
>>>>> beq .L1
>>>>> .L6:
>>>>> ldr q30, [x3, x1]
>>>>> cmpeq p15.s, p7/z, z30.s, z27.s
>>>>> b.none .L2
>>>>>
>>>>> and easily prove it correct.
>>>>>
>>>>> Bootstrapped Regtested on aarch64-none-linux-gnu,
>>>>> arm-none-linux-gnueabihf, x86_64-pc-linux-gnu
>>>>> -m32, -m64 and no issues.
>>>>>
>>>>> Ok for master?
>>>>>
>>>>> Thanks,
>>>>> Tamar
>>>>>
>>>>> gcc/ChangeLog:
>>>>>
>>>>> PR target/118974
>>>>> * optabs.cc (prepare_cmp_insn): Refactor to take optab to check for
>>>>> instead of hardcoded cbranch.
>>>>> (emit_cmp_and_jump_insns): Try to emit a vec_cbranch if supported.
>>>>> * optabs.def (vec_cbranch_any_optab, vec_cbranch_all_optab): New.
>>>>> * doc/md.texi (cbranch_any@var{mode}4, cbranch_any@var{mode}4):
>>>> Document
>>>>> them.
>>>>>
>>>>> ---
>>>>> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
>>>>> index
>>>>
>> f6314af46923beee0100a1410f089efd34d7566d..7dbfbbe1609a196b0834d458
>>>> b26f61904eaf5b24 100644
>>>>> --- a/gcc/doc/md.texi
>>>>> +++ b/gcc/doc/md.texi
>>>>> @@ -7622,6 +7622,24 @@ Operand 0 is a comparison operator. Operand 1
>>>> and operand 2 are the
>>>>> first and second operands of the comparison, respectively. Operand 3
>>>>> is the @code{code_label} to jump to.
>>>>>
>>>>> +@cindex @code{cbranch_any@var{mode}4} instruction pattern
>>>>> +@item @samp{cbranch_any@var{mode}4}
>>>>> +Conditional branch instruction combined with a compare instruction on
>> vectors
>>>>> +where it is required that at least one of the elementwise comparisons of
>>>>> the
>>>>> +two input vectors is true.
>>>>> +Operand 0 is a comparison operator. Operand 1 and operand 2 are the
>>>>> +first and second operands of the comparison, respectively. Operand 3
>>>>> +is the @code{code_label} to jump to.
>>>>> +
>>>>> +@cindex @code{cbranch_all@var{mode}4} instruction pattern
>>>>> +@item @samp{cbranch_all@var{mode}4}
>>>>> +Conditional branch instruction combined with a compare instruction on
>> vectors
>>>>> +where it is required that at all of the elementwise comparisons of the
>>>>> +two input vectors are true.
>>>>> +Operand 0 is a comparison operator. Operand 1 and operand 2 are the
>>>>> +first and second operands of the comparison, respectively. Operand 3
>>>>> +is the @code{code_label} to jump to.
>>>>> +
>>>>> @cindex @code{jump} instruction pattern
>>>>> @item @samp{jump}
>>>>> A jump inside a function; an unconditional branch. Operand 0 is the
>>>>> diff --git a/gcc/optabs.cc b/gcc/optabs.cc
>>>>> index
>>>>
>> 0a14b1eef8a5795e6fd24ade6da55841696315b8..77d5e6ee5d26ccda3d39126
>>>> 5bc45fd454530cc67 100644
>>>>> --- a/gcc/optabs.cc
>>>>> +++ b/gcc/optabs.cc
>>>>> @@ -4418,6 +4418,9 @@ can_vec_extract_var_idx_p (machine_mode
>>>> vec_mode, machine_mode extr_mode)
>>>>>
>>>>> *PMODE is the mode of the inputs (in case they are const_int).
>>>>>
>>>>> + *OPTAB is the optab to check for OPTAB_DIRECT support. Defaults to
>>>>> + cbranch_optab.
>>>>> +
>>>>> This function performs all the setup necessary so that the caller only
>>>>> has
>>>>> to emit a single comparison insn. This setup can involve doing a
>>>>> BLKmode
>>>>> comparison or emitting a library call to perform the comparison if no
>>>>> insn
>>>>> @@ -4429,7 +4432,7 @@ can_vec_extract_var_idx_p (machine_mode
>>>> vec_mode, machine_mode extr_mode)
>>>>> static void
>>>>> prepare_cmp_insn (rtx x, rtx y, enum rtx_code comparison, rtx size,
>>>>> int unsignedp, enum optab_methods methods,
>>>>> - rtx *ptest, machine_mode *pmode)
>>>>> + rtx *ptest, machine_mode *pmode, optab
>>>> optab=cbranch_optab)
>>>>> {
>>>>> machine_mode mode = *pmode;
>>>>> rtx libfunc, test;
>>>>> @@ -4547,7 +4550,7 @@ prepare_cmp_insn (rtx x, rtx y, enum rtx_code
>>>> comparison, rtx size,
>>>>> FOR_EACH_WIDER_MODE_FROM (cmp_mode, mode)
>>>>> {
>>>>> enum insn_code icode;
>>>>> - icode = optab_handler (cbranch_optab, cmp_mode);
>>>>> + icode = optab_handler (optab, cmp_mode);
>>>>> if (icode != CODE_FOR_nothing
>>>>> && insn_operand_matches (icode, 0, test))
>>>>> {
>>>>> @@ -4580,7 +4583,7 @@ prepare_cmp_insn (rtx x, rtx y, enum rtx_code
>>>> comparison, rtx size,
>>>>> if (comparison == UNORDERED && rtx_equal_p (x, y))
>>>>> {
>>>>> prepare_cmp_insn (x, y, UNLT, NULL_RTX, unsignedp, OPTAB_WIDEN,
>>>>> - ptest, pmode);
>>>>> + ptest, pmode, optab);
>>>>> if (*ptest)
>>>>> return;
>>>>> }
>>>>> @@ -4632,7 +4635,7 @@ prepare_cmp_insn (rtx x, rtx y, enum rtx_code
>>>> comparison, rtx size,
>>>>>
>>>>> *pmode = ret_mode;
>>>>> prepare_cmp_insn (x, y, comparison, NULL_RTX, unsignedp, methods,
>>>>> - ptest, pmode);
>>>>> + ptest, pmode, optab);
>>>>> }
>>>>>
>>>>> return;
>>>>> @@ -4816,12 +4819,68 @@ emit_cmp_and_jump_insns (rtx x, rtx y, enum
>>>> rtx_code comparison, rtx size,
>>>>> the target supports tbranch. */
>>>>> machine_mode tmode = mode;
>>>>> direct_optab optab;
>>>>> - if (op1 == CONST0_RTX (GET_MODE (op1))
>>>>> - && validate_test_and_branch (val, &test, &tmode,
>>>>> - &optab) != CODE_FOR_nothing)
>>>>> + if (op1 == CONST0_RTX (GET_MODE (op1)))
>>>>> {
>>>>> - emit_cmp_and_jump_insn_1 (test, tmode, label, optab, prob, true);
>>>>> - return;
>>>>> + if (validate_test_and_branch (val, &test, &tmode,
>>>>> + &optab) != CODE_FOR_nothing)
>>>>> + {
>>>>> + emit_cmp_and_jump_insn_1 (test, tmode, label, optab, prob, true);
>>>>> + return;
>>>>> + }
>>>>> +
>>>>> + /* If we are comparing equality with 0, check if VAL is another
>>>>> equality
>>>>> + comparison and if the target supports it directly. */
>>>>> + if (val && TREE_CODE (val) == SSA_NAME
>>>>> + && VECTOR_BOOLEAN_TYPE_P (TREE_TYPE (val))
>>>>> + && (comparison == NE || comparison == EQ))
>>>>> + {
>>>>> + auto def_stmt = SSA_NAME_DEF_STMT (val);
>>>>> + enum insn_code icode;
>>>>> + if (is_gimple_assign (def_stmt)
>>>>> + && TREE_CODE_CLASS (gimple_assign_rhs_code (def_stmt))
>>>>> + == tcc_comparison)
>>>>> + {
>>>>> + class expand_operand ops[2];
>>>>> + rtx_insn *tmp = NULL;
>>>>> + start_sequence ();
>>>>> + rtx op0c = expand_normal (gimple_assign_rhs1 (def_stmt));
>>>>> + rtx op1c = expand_normal (gimple_assign_rhs2 (def_stmt));
>>>>> + machine_mode mode2 = GET_MODE (op0c);
>>>>> + create_input_operand (&ops[0], op0c, mode2);
>>>>> + create_input_operand (&ops[1], op1c, mode2);
>>>>> +
>>>>> + int unsignedp2 = TYPE_UNSIGNED (TREE_TYPE (val));
>>>>> + auto inner_code = gimple_assign_rhs_code (def_stmt);
>>>>> + rtx test2 = NULL_RTX;
>>>>> +
>>>>> + enum rtx_code comparison2 = get_rtx_code (inner_code,
>>>>> unsignedp2);
>>>>> + if (unsignedp2)
>>>>> + comparison2 = unsigned_condition (comparison2);
>>>>> + if (comparison == NE)
>>>>> + optab = vec_cbranch_any_optab;
>>>>> + else
>>>>> + optab = vec_cbranch_all_optab;
>>>>> +
>>>>> + if ((icode = optab_handler (optab, mode2))
>>>>> + != CODE_FOR_nothing
>>>>> + && maybe_legitimize_operands (icode, 1, 2, ops))
>>>>> + {
>>>>> + prepare_cmp_insn (ops[0].value, ops[1].value, comparison2,
>>>>> + size, unsignedp2, OPTAB_DIRECT, &test2,
>>>>> + &mode2, optab);
>>>>> + emit_cmp_and_jump_insn_1 (test2, mode2, label,
>>>>> + optab, prob, false);
>>>>> + tmp = get_insns ();
>>>>> + }
>>>>> +
>>>>> + end_sequence ();
>>>>> + if (tmp)
>>>>> + {
>>>>> + emit_insn (tmp);
>>>>> + return;
>>>>> + }
>>>>> + }
>>>>> + }
>>>>> }
>>>>>
>>>>> emit_cmp_and_jump_insn_1 (test, mode, label, cbranch_optab, prob,
>> false);
>>>>> diff --git a/gcc/optabs.def b/gcc/optabs.def
>>>>> index
>>>>
>> 23f792352388dd5f8de8a1999643179328214abf..a600938fb9e4480ff2d78c9b
>>>> 0be344e4856539bb 100644
>>>>> --- a/gcc/optabs.def
>>>>> +++ b/gcc/optabs.def
>>>>> @@ -424,6 +424,8 @@ OPTAB_D (smulhrs_optab, "smulhrs$a3")
>>>>> OPTAB_D (umulhs_optab, "umulhs$a3")
>>>>> OPTAB_D (umulhrs_optab, "umulhrs$a3")
>>>>> OPTAB_D (sdiv_pow2_optab, "sdiv_pow2$a3")
>>>>> +OPTAB_D (vec_cbranch_any_optab, "vec_cbranch_any$a4")
>>>>> +OPTAB_D (vec_cbranch_all_optab, "vec_cbranch_all$a4")
>>>>> OPTAB_D (vec_pack_sfix_trunc_optab, "vec_pack_sfix_trunc_$a")
>>>>> OPTAB_D (vec_pack_ssat_optab, "vec_pack_ssat_$a")
>>>>> OPTAB_D (vec_pack_trunc_optab, "vec_pack_trunc_$a")
>>>>>
>>>>>
>>>>>
>>>>
>>>> --
>>>> Richard Biener <rguent...@suse.de>
>>>> SUSE Software Solutions Germany GmbH,
>>>> Frankenstrasse 146, 90461 Nuernberg, Germany;
>>>> GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG
>> Nuernberg)
>>>
>>
>> --
>> Richard Biener <rguent...@suse.de>
>> SUSE Software Solutions Germany GmbH,
>> Frankenstrasse 146, 90461 Nuernberg, Germany;
>> GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)