On Thu, Aug 17, 2017 at 12:35 PM, Richard Sandiford <richard.sandif...@linaro.org> wrote: > "Bin.Cheng" <amker.ch...@gmail.com> writes: >> On Wed, Aug 16, 2017 at 6:50 PM, Richard Sandiford >> <richard.sandif...@linaro.org> wrote: >>> "Bin.Cheng" <amker.ch...@gmail.com> writes: >>>> On Wed, Aug 16, 2017 at 5:00 PM, Richard Sandiford >>>> <richard.sandif...@linaro.org> wrote: >>>>> "Bin.Cheng" <amker.ch...@gmail.com> writes: >>>>>> On Wed, Aug 16, 2017 at 2:38 PM, Richard Sandiford >>>>>> <richard.sandif...@linaro.org> wrote: >>>>>>> The first loop in the testcase regressed after my recent changes to >>>>>>> dr_analyze_innermost. Previously we would treat "i" as an iv even >>>>>>> for bb analysis and end up with: >>>>>>> >>>>>>> DR_BASE_ADDRESS: p or q >>>>>>> DR_OFFSET: 0 >>>>>>> DR_INIT: 0 or 4 >>>>>>> DR_STEP: 16 >>>>>>> >>>>>>> We now always keep the step as 0 instead, so for an int "i" we'd have: >>>>>>> >>>>>>> DR_BASE_ADDRESS: p or q >>>>>>> DR_OFFSET: (intptr_t) i >>>>>>> DR_INIT: 0 or 4 >>>>>>> DR_STEP: 0 >>>>>>> >>>>>>> This is also what we'd like to have for the unsigned "i", but the >>>>>>> problem is that strip_constant_offset thinks that the "i + 1" in >>>>>>> "(intptr_t) (i + 1)" could wrap and so doesn't peel off the "+ 1". >>>>>>> The [i + 1] accesses therefore have a DR_OFFSET equal to the SSA >>>>>>> name that holds "(intptr_t) (i + 1)", meaning that the accesses no >>>>>>> longer seem to be related to the [i] ones. >>>>>> >>>>>> Didn't read the change in detail, so sorry if I mis-understood the issue. >>>>>> I made changes in scev to better fold type conversion by various sources >>>>>> of information, for example, vrp, niters, undefined overflow behavior >>>>>> etc. >>>>>> In theory these information should be available for other >>>>>> optimizers without >>>>>> querying scev. For the mentioned test, vrp should compute accurate range >>>>>> information for "i" so that we can fold (intptr_t) (i + 1) it without >>>>>> worrying >>>>>> overflow. Note we don't do it in generic folding because >>>>>> (intptr_t) (i) + 1 >>>>>> could be more expensive (especially in case of (T)(i + j)), or because >>>>>> the >>>>>> CST part is in bigger precision after conversion. >>>>>> But such folding is wanted in several places, e.g, IVOPTs. To provide >>>>>> such >>>>>> an interface, we changed tree-affine and made it do aggressive fold. I >>>>>> am >>>>>> curious if it's possible to use aff_tree to implement >>>>>> strip_constant_offset >>>>>> here since aggressive folding is wanted. After all, using additional >>>>>> chrec >>>>>> looks like a little heavy wrto the simple test. >>>>> >>>>> Yeah, using aff_tree does work here when the bounds are constant. >>>>> It doesn't look like it works for things like: >>>>> >>>>> double p[1000]; >>>>> double q[1000]; >>>>> >>>>> void >>>>> f4 (unsigned int n) >>>>> { >>>>> for (unsigned int i = 0; i < n; i += 4) >>>>> { >>>>> double a = q[i] + p[i]; >>>>> double b = q[i + 1] + p[i + 1]; >>>>> q[i] = a; >>>>> q[i + 1] = b; >>>>> } >>>>> } >>>>> >>>>> though, where the bounds on the global arrays guarantee that [i + 1] can't >>>>> overflow, even though "n" is unconstrained. The patch as posted handles >>>>> this case too. >>>> BTW is this a missed optimization in value range analysis? The range >>>> information for i should flow in a way like: array boundary -> niters >>>> -> scev/vrp. >>>> I think that's what niters/scev do in analysis. >>> >>> Yeah, maybe :-) It looks like the problem is that when SLP runs, >>> the previous VRP pass came before loop header copying, so the (single) >>> header has to cope with n == 0 case. Thus we get: >> Ah, there are several passes in between vrp and pass_ch, not sure if >> any such pass depends on vrp intensively. I would suggestion reorder >> the two passes, or standalone VRP interface updating information for >> loop region after header copied? This is a non-trivial issue that >> needs to be fixed. Niters analyzer rely on >> simplify_using_initial_conditions heavily to get the same information, >> which in my opinion should be provided by VRP. Though this won't be >> able to obsolete simplify_using_initial_conditions because VRP is weak >> in symbolic range... >> >>> >>> Visiting statement: >>> i_15 = ASSERT_EXPR <i_6, i_6 < n_9(D)>; >>> Intersecting >>> [0, n_9(D) + 4294967295] EQUIVALENCES: { i_6 } (1 elements) >>> and >>> [0, 0] >>> to >>> [0, 0] EQUIVALENCES: { i_6 } (1 elements) >>> Intersecting >>> [0, 0] EQUIVALENCES: { i_6 } (1 elements) >>> and >>> [0, 1000] >>> to >>> [0, 0] EQUIVALENCES: { i_6 } (1 elements) >>> Found new range for i_15: [0, 0] >>> >>> Visiting statement: >>> _3 = i_15 + 1; >>> Match-and-simplified i_15 + 1 to 1 >>> Intersecting >>> [1, 1] >>> and >>> [0, +INF] >>> to >>> [1, 1] >>> Found new range for _3: [1, 1] >>> >>> (where _3 is the index we care about), followed by: >>> >>> Visiting statement: >>> i_15 = ASSERT_EXPR <i_6, i_6 < n_9(D)>; >>> Intersectings/aarch64-linux/trunk-orig/debug/gcc' >>> [0, n_9(D) + 4294967295] EQUIVALENCES: { i_6 } (1 elements) >>> and >>> [0, 4] >>> to >>> [0, n_9(D) + 4294967295] EQUIVALENCES: { i_6 } (1 elements) >>> Intersecting >>> [0, n_9(D) + 4294967295] EQUIVALENCES: { i_6 } (1 elements) >>> and >>> [0, 1000] >>> to >>> [0, n_9(D) + 4294967295] EQUIVALENCES: { i_6 } (1 elements) >>> Found new range for i_15: [0, n_9(D) + 4294967295] >>> >>> Visiting statement: >>> _3 = i_15 + 1; >>> Intersecting >>> [0, +INF] >>> and >>> [0, +INF] >>> to >>> [0, +INF] >>> Found new range for _3: [0, +INF] >>> >>> I guess in this case it would be better to intersect the i_15 ranges >>> to [0, 1000] rather than [0, n_9(D) + 4294967295]. >>> >>> It does work if another VRP pass runs after CH. But even then, >>> is it a good idea to rely on the range info being kept up-to-date >>> all the way through to SLP? A lot happens inbetween. >> To some extend yes. Now I understand that SCEV uses niters >> information to prove no_overflow. Niters analysis does infer better >> information from array boundary, while VRP fails to do that. I don't >> worry much about gap between vrp pass and slp, it's basically the same >> as niters. Both information are analyzed at one point and meant to be >> used by following passes. It's also not common for vrp information >> become invalid given we are on SSA? > > I'm not worried so much about vrp information becoming invalid on > an SSA name that existed when VRP was run. It's more a question > of what happens about SSA names that get introduced after VRP, > e.g. by things like dom, reassoc, PRE, etc. For induction variables in concern, these passes shouldn't aggressively introduces new variables I think. > >> Now that data address is not analyzed against loop, VRP would be the >> only information we can use unless adding back scev analysis. IMHO, >> the patch is doing so in another way than before. It requires >> additional chrec data structure. I remember the previous patch >> enables more slp vectorization, is it because of "step" information >> from scev? > > Do you mean that: > > 2017-07-03 Richard Sandiford <richard.sandif...@linaro.org> > > * tree-data-ref.c (dr_analyze_innermost): Replace the "nest" > parameter with a "loop" parameter and use it instead of the > loop containing DR_STMT. Don't check simple_iv when doing > BB analysis. Describe the two analysis modes in the comment. > > enabled more SLP vectorisation in bb-slp-pr65935.c? That was due > to us not doing IV analysis for BB vectorisation, and ensuring that > the step was always zero. Which means vectorizer code can handle not IV-analyzed offset, but can't for analyzed form? > >> In this patch, step information is simply discarded. I am >> wondering if possible to always analyze scev within innermost loop for >> slp while discards step information. > > Well, we don't calculate a step for bb analysis (i.e. it's not the case > the patch calculates step information and then simply discards it). > I don't see how that would work. For bb analysis, the DR_OFFSET + DR_INIT > has to give the offset for every execution of the block, not just the > first iteration of the containing loop. So if we get back a nonzero > step, we have to do something with it. Yeah. > > But: > > (a) the old simple_iv analysis is more expensive than simply calling > analyze_scev, so I don't think this is a win in terms of complexity. > > (b) for bb analysis, there's nothing particularly special about the > innermost loop. It makes more sense to analyse it in the innermost > loop for which the offset is invariant, as shown by the second > testcase in the patch. > > (c) The patch helps with loop vectorisation too, since analysing the > starting DR_OFFSET in the context of the containing loop can help > in a similar way as analysing the full offset does for SLP.
I have to admit I am not very much into this method. It complicates structure as well as code. Mostly because now dr_init are split into two different fields and one of it is lazily computed. For example: > @@ -2974,12 +2974,12 @@ vect_vfa_segment_size (struct data_refer > vect_no_alias_p (struct data_reference *a, struct data_reference *b, > tree segment_length_a, tree segment_length_b) > { > - gcc_assert (TREE_CODE (DR_INIT (a)) == INTEGER_CST > - && TREE_CODE (DR_INIT (b)) == INTEGER_CST); > - if (tree_int_cst_equal (DR_INIT (a), DR_INIT (b))) > + gcc_assert (TREE_CODE (DR_CHREC_INIT (a)) == INTEGER_CST > + && TREE_CODE (DR_CHREC_INIT (b)) == INTEGER_CST); > + if (tree_int_cst_equal (DR_CHREC_INIT (a), DR_CHREC_INIT (b))) > return false; > > - tree seg_a_min = DR_INIT (a); > + tree seg_a_min = DR_CHREC_INIT (a); > tree seg_a_max = fold_build2 (PLUS_EXPR, TREE_TYPE (seg_a_min), > seg_a_min, segment_length_a); > /* For negative step, we need to adjust address range by TYPE_SIZE_UNIT > @@ -2990,10 +2990,10 @@ vect_no_alias_p (struct data_reference * > tree unit_size = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (a))); > seg_a_min = fold_build2 (PLUS_EXPR, TREE_TYPE (seg_a_max), > seg_a_max, unit_size); > - seg_a_max = fold_build2 (PLUS_EXPR, TREE_TYPE (DR_INIT (a)), > - DR_INIT (a), unit_size); > + seg_a_max = fold_build2 (PLUS_EXPR, TREE_TYPE (DR_CHREC_INIT (a)), > + DR_CHREC_INIT (a), unit_size); > } > - tree seg_b_min = DR_INIT (b); > + tree seg_b_min = DR_CHREC_INIT (b); > tree seg_b_max = fold_build2 (PLUS_EXPR, TREE_TYPE (seg_b_min), > seg_b_min, segment_length_b); > if (tree_int_cst_compare (DR_STEP (b), size_zero_node) < 0) Use of DR_INIT is simply replaced by DR_CHREC_INIT. Is it safe to do so in case of non-ZERO DR_INIT? It worries me that I may need to think twice before referring to DR_INIT because it's not clear when DR_OFFSET is split and DR_CHREC_INIT becomes non-ZERO. It may simply because I am too dumb to handle this. I will leave this to richi. Thanks, bin > > Thanks, > Richard > >> >> Thanks, >> bin >>> >>> FWIW, the old simple_iv check that I removed for bb data-ref analysis >>> relies on SCEV analysis too, so I don't think this is more expensive >>> than what we had before. >>> >>> Thanks, >>> Richard