Re: [patch AArch64] Do not perform a vector splat for vector initialisation if it is not useful
"Richard Earnshaw (lists)"writes: > On 16/05/18 09:37, Kyrill Tkachov wrote: >> >> On 15/05/18 10:58, Richard Biener wrote: >>> On Tue, May 15, 2018 at 10:20 AM Kyrill Tkachov >>> >>> wrote: >>> Hi all, This is a respin of James's patch from: >>> https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html The original patch was approved and committed but was later reverted >>> because of failures on big-endian. This tweaked version fixes the big-endian failures in >>> aarch64_expand_vector_init by picking the right element of VALS to move into the low part of the vector register >>> depending on endianness. The rest of the patch stays the same. I'm looking for approval on the aarch64 parts, as they >>> are the ones that have changed since the last approved version of the patch. --- In the testcase in this patch we create an SLP vector with only two elements. Our current vector initialisation code will first duplicate the first element to both lanes, then overwrite the top lane with a new value. This duplication can be clunky and wasteful. Better would be to simply use the fact that we will always be overwriting the remaining bits, and simply move the first element to the corrcet place (implicitly zeroing all other bits). This reduces the code generation for this case, and can allow more efficient addressing modes, and other second order benefits for AArch64 code which has been vectorized to V2DI mode. Note that the change is generic enough to catch the case for any vector mode, but is expected to be most useful for 2x64-bit vectorization. Unfortunately, on its own, this would cause failures in gcc.target/aarch64/load_v2vec_lanes_1.c and gcc.target/aarch64/store_v2vec_lanes.c , which expect to see many more vec_merge and vec_duplicate for their simplifications to apply. To fix this, add a special case to the AArch64 code if we are loading from two memory addresses, and use the load_pair_lanes patterns directly. We also need a new pattern in simplify-rtx.c:simplify_ternary_operation , to catch: (vec_merge:OUTER (vec_duplicate:OUTER x:INNER) (subreg:OUTER y:INNER 0) (const_int N)) And simplify it to: (vec_concat:OUTER x:INNER y:INNER) or (vec_concat y x) This is similar to the existing patterns which are tested in this function, without requiring the second operand to also be a vec_duplicate. Bootstrapped and tested on aarch64-none-linux-gnu and tested on aarch64-none-elf. Note that this requires https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html if we don't want to ICE creating broken vector zero extends. Are the non-AArch64 parts OK? >>> Is (vec_merge (subreg ..) (vec_duplicate)) canonicalized to the form >>> you handle? I see the (vec_merge (vec_duplicate...) (vec_concat)) case >>> also doesn't handle the swapped operand case. >>> >>> Otherwise the middle-end parts looks ok. >> >> I don't see any explicit canonicalisation code for it. >> I've updated the simplify-rtx part to handle the swapped operand case. >> Is the attached patch better in this regard? I couldn't think of a clean >> way to avoid >> duplicating some logic (beyond creating a new function away from the >> callsite). >> >> Thanks, >> Kyrill >> >>> Thanks, >>> Richard. >>> Thanks, James --- 2018-05-15 James Greenhalgh Kyrylo Tkachov * config/aarch64/aarch64.c (aarch64_expand_vector_init): Modify code generation for cases where splatting a value is not useful. * simplify-rtx.c (simplify_ternary_operation): Simplify vec_merge across a vec_duplicate and a paradoxical subreg >>> forming a vector mode to a vec_concat. 2018-05-15 James Greenhalgh * gcc.target/aarch64/vect-slp-dup.c: New. >> > > I'm surprised we don't seem to have a function in the compiler that > performs this check: > > + && rtx_equal_p (XEXP (x1, 0), > + plus_constant (Pmode, > + XEXP (x0, 0), > + GET_MODE_SIZE (inner_mode > > Without generating dead RTL (plus_constant will rarely be able to return > a subexpression of the original pattern). I would have thought this > sort of test was not that uncommon. FWIW, I think the way to write it without generating dead RTL is: rtx_equal_p (strip_offset (XEXP (x0, 0), _offset), strip_offset (XEXP (x1, 0), _offset)) && known_eq (x1_offset,
Re: [patch AArch64] Do not perform a vector splat for vector initialisation if it is not useful
On 16/05/18 09:37, Kyrill Tkachov wrote: > > On 15/05/18 10:58, Richard Biener wrote: >> On Tue, May 15, 2018 at 10:20 AM Kyrill Tkachov >>>> wrote: >> >>> Hi all, >>> This is a respin of James's patch from: >> https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html >>> The original patch was approved and committed but was later reverted >> because of failures on big-endian. >>> This tweaked version fixes the big-endian failures in >> aarch64_expand_vector_init by picking the right >>> element of VALS to move into the low part of the vector register >> depending on endianness. The rest of the patch >>> stays the same. I'm looking for approval on the aarch64 parts, as they >> are the ones that have changed >>> since the last approved version of the patch. >>> --- >>> In the testcase in this patch we create an SLP vector with only two >>> elements. Our current vector initialisation code will first duplicate >>> the first element to both lanes, then overwrite the top lane with a new >>> value. >>> This duplication can be clunky and wasteful. >>> Better would be to simply use the fact that we will always be >>> overwriting >>> the remaining bits, and simply move the first element to the corrcet >>> place >>> (implicitly zeroing all other bits). >>> This reduces the code generation for this case, and can allow more >>> efficient addressing modes, and other second order benefits for AArch64 >>> code which has been vectorized to V2DI mode. >>> Note that the change is generic enough to catch the case for any vector >>> mode, but is expected to be most useful for 2x64-bit vectorization. >>> Unfortunately, on its own, this would cause failures in >>> gcc.target/aarch64/load_v2vec_lanes_1.c and >>> gcc.target/aarch64/store_v2vec_lanes.c , which expect to see many more >>> vec_merge and vec_duplicate for their simplifications to apply. To fix >>> this, >>> add a special case to the AArch64 code if we are loading from two memory >>> addresses, and use the load_pair_lanes patterns directly. >>> We also need a new pattern in simplify-rtx.c:simplify_ternary_operation >>> , to >>> catch: >>> (vec_merge:OUTER >>> (vec_duplicate:OUTER x:INNER) >>> (subreg:OUTER y:INNER 0) >>> (const_int N)) >>> And simplify it to: >>> (vec_concat:OUTER x:INNER y:INNER) or (vec_concat y x) >>> This is similar to the existing patterns which are tested in this >>> function, >>> without requiring the second operand to also be a vec_duplicate. >>> Bootstrapped and tested on aarch64-none-linux-gnu and tested on >>> aarch64-none-elf. >>> Note that this requires >>> https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html >>> if we don't want to ICE creating broken vector zero extends. >>> Are the non-AArch64 parts OK? >> Is (vec_merge (subreg ..) (vec_duplicate)) canonicalized to the form >> you handle? I see the (vec_merge (vec_duplicate...) (vec_concat)) case >> also doesn't handle the swapped operand case. >> >> Otherwise the middle-end parts looks ok. > > I don't see any explicit canonicalisation code for it. > I've updated the simplify-rtx part to handle the swapped operand case. > Is the attached patch better in this regard? I couldn't think of a clean > way to avoid > duplicating some logic (beyond creating a new function away from the > callsite). > > Thanks, > Kyrill > >> Thanks, >> Richard. >> >>> Thanks, >>> James >>> --- >>> 2018-05-15 James Greenhalgh >>> Kyrylo Tkachov >>> * config/aarch64/aarch64.c (aarch64_expand_vector_init): >>> Modify >>> code generation for cases where splatting a value is not >>> useful. >>> * simplify-rtx.c (simplify_ternary_operation): Simplify >>> vec_merge across a vec_duplicate and a paradoxical subreg >> forming a vector >>> mode to a vec_concat. >>> 2018-05-15 James Greenhalgh >>> * gcc.target/aarch64/vect-slp-dup.c: New. > I'm surprised we don't seem to have a function in the compiler that performs this check: + && rtx_equal_p (XEXP (x1, 0), + plus_constant (Pmode, +XEXP (x0, 0), +GET_MODE_SIZE (inner_mode Without generating dead RTL (plus_constant will rarely be able to return a subexpression of the original pattern). I would have thought this sort of test was not that uncommon. However, I don't think that needs to hold up this patch. OK. R. > > vec-splat.patch > > > diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c > index > a2003fe52875f1653d644347bafd7773d1f01e91..6bf6c05535b61eef1021d46bcd8448fb3a0b25f4 > 100644 > --- a/gcc/config/aarch64/aarch64.c > +++ b/gcc/config/aarch64/aarch64.c > @@ -13916,9 +13916,54 @@
Re: [patch AArch64] Do not perform a vector splat for vector initialisation if it is not useful
On Wed, May 16, 2018 at 11:10:55AM +0100, Kyrill Tkachov wrote: > On 16/05/18 10:42, Richard Biener wrote: > >Segher, do you know where canonicalization rules are documented? > >IIRC we do not actively try to canonicalize in most cases. > > The documentation we have for RTL canonicalisation is at: > https://gcc.gnu.org/onlinedocs/gccint/Insn-Canonicalizations.html#Insn-Canonicalizations > > It doesn't mention anything about vec_merge AFAICS so I couldn't convince > myself that there > is a canonicalisation that we enforce (though maybe someone can prove me > wrong). Many canonicalisations aren't documented, it's never clear which of the canonicalisations are how canonical :-/ Segher
Re: [patch AArch64] Do not perform a vector splat for vector initialisation if it is not useful
On Wed, May 16, 2018 at 11:42:39AM +0200, Richard Biener wrote: > Works for me. Were you able to actually create such RTL from testcases? > Segher, do you know where canonicalization rules are documented? > IIRC we do not actively try to canonicalize in most cases. md.texi, node "Insn Canonicalizations"? Segher
Re: [patch AArch64] Do not perform a vector splat for vector initialisation if it is not useful
On 16/05/18 10:42, Richard Biener wrote: On Wed, May 16, 2018 at 10:37 AM Kyrill Tkachovwrote: On 15/05/18 10:58, Richard Biener wrote: On Tue, May 15, 2018 at 10:20 AM Kyrill Tkachov wrote: Hi all, This is a respin of James's patch from: https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html The original patch was approved and committed but was later reverted because of failures on big-endian. This tweaked version fixes the big-endian failures in aarch64_expand_vector_init by picking the right element of VALS to move into the low part of the vector register depending on endianness. The rest of the patch stays the same. I'm looking for approval on the aarch64 parts, as they are the ones that have changed since the last approved version of the patch. --- In the testcase in this patch we create an SLP vector with only two elements. Our current vector initialisation code will first duplicate the first element to both lanes, then overwrite the top lane with a new value. This duplication can be clunky and wasteful. Better would be to simply use the fact that we will always be overwriting the remaining bits, and simply move the first element to the corrcet place (implicitly zeroing all other bits). This reduces the code generation for this case, and can allow more efficient addressing modes, and other second order benefits for AArch64 code which has been vectorized to V2DI mode. Note that the change is generic enough to catch the case for any vector mode, but is expected to be most useful for 2x64-bit vectorization. Unfortunately, on its own, this would cause failures in gcc.target/aarch64/load_v2vec_lanes_1.c and gcc.target/aarch64/store_v2vec_lanes.c , which expect to see many more vec_merge and vec_duplicate for their simplifications to apply. To fix this, add a special case to the AArch64 code if we are loading from two memory addresses, and use the load_pair_lanes patterns directly. We also need a new pattern in simplify-rtx.c:simplify_ternary_operation , to catch: (vec_merge:OUTER (vec_duplicate:OUTER x:INNER) (subreg:OUTER y:INNER 0) (const_int N)) And simplify it to: (vec_concat:OUTER x:INNER y:INNER) or (vec_concat y x) This is similar to the existing patterns which are tested in this function, without requiring the second operand to also be a vec_duplicate. Bootstrapped and tested on aarch64-none-linux-gnu and tested on aarch64-none-elf. Note that this requires https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html if we don't want to ICE creating broken vector zero extends. Are the non-AArch64 parts OK? Is (vec_merge (subreg ..) (vec_duplicate)) canonicalized to the form you handle? I see the (vec_merge (vec_duplicate...) (vec_concat)) case also doesn't handle the swapped operand case. Otherwise the middle-end parts looks ok. I don't see any explicit canonicalisation code for it. I've updated the simplify-rtx part to handle the swapped operand case. Is the attached patch better in this regard? I couldn't think of a clean way to avoid duplicating some logic (beyond creating a new function away from the callsite). Works for me. Were you able to actually create such RTL from testcases? Segher, do you know where canonicalization rules are documented? IIRC we do not actively try to canonicalize in most cases. The documentation we have for RTL canonicalisation is at: https://gcc.gnu.org/onlinedocs/gccint/Insn-Canonicalizations.html#Insn-Canonicalizations It doesn't mention anything about vec_merge AFAICS so I couldn't convince myself that there is a canonicalisation that we enforce (though maybe someone can prove me wrong). Kyrill Richard. Thanks, Kyrill Thanks, Richard. Thanks, James --- 2018-05-15 James Greenhalgh Kyrylo Tkachov * config/aarch64/aarch64.c (aarch64_expand_vector_init): Modify code generation for cases where splatting a value is not useful. * simplify-rtx.c (simplify_ternary_operation): Simplify vec_merge across a vec_duplicate and a paradoxical subreg forming a vector mode to a vec_concat. 2018-05-15 James Greenhalgh * gcc.target/aarch64/vect-slp-dup.c: New.
Re: [patch AArch64] Do not perform a vector splat for vector initialisation if it is not useful
On Wed, May 16, 2018 at 10:37 AM Kyrill Tkachovwrote: > On 15/05/18 10:58, Richard Biener wrote: > > On Tue, May 15, 2018 at 10:20 AM Kyrill Tkachov > > > > wrote: > > > >> Hi all, > >> This is a respin of James's patch from: > > https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html > >> The original patch was approved and committed but was later reverted > > because of failures on big-endian. > >> This tweaked version fixes the big-endian failures in > > aarch64_expand_vector_init by picking the right > >> element of VALS to move into the low part of the vector register > > depending on endianness. The rest of the patch > >> stays the same. I'm looking for approval on the aarch64 parts, as they > > are the ones that have changed > >> since the last approved version of the patch. > >> --- > >> In the testcase in this patch we create an SLP vector with only two > >> elements. Our current vector initialisation code will first duplicate > >> the first element to both lanes, then overwrite the top lane with a new > >> value. > >> This duplication can be clunky and wasteful. > >> Better would be to simply use the fact that we will always be > >> overwriting > >> the remaining bits, and simply move the first element to the corrcet > >> place > >> (implicitly zeroing all other bits). > >> This reduces the code generation for this case, and can allow more > >> efficient addressing modes, and other second order benefits for AArch64 > >> code which has been vectorized to V2DI mode. > >> Note that the change is generic enough to catch the case for any vector > >> mode, but is expected to be most useful for 2x64-bit vectorization. > >> Unfortunately, on its own, this would cause failures in > >> gcc.target/aarch64/load_v2vec_lanes_1.c and > >> gcc.target/aarch64/store_v2vec_lanes.c , which expect to see many more > >> vec_merge and vec_duplicate for their simplifications to apply. To fix > >> this, > >> add a special case to the AArch64 code if we are loading from two memory > >> addresses, and use the load_pair_lanes patterns directly. > >> We also need a new pattern in simplify-rtx.c:simplify_ternary_operation > >> , to > >> catch: > >> (vec_merge:OUTER > >> (vec_duplicate:OUTER x:INNER) > >> (subreg:OUTER y:INNER 0) > >> (const_int N)) > >> And simplify it to: > >> (vec_concat:OUTER x:INNER y:INNER) or (vec_concat y x) > >> This is similar to the existing patterns which are tested in this > >> function, > >> without requiring the second operand to also be a vec_duplicate. > >> Bootstrapped and tested on aarch64-none-linux-gnu and tested on > >> aarch64-none-elf. > >> Note that this requires > >> https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html > >> if we don't want to ICE creating broken vector zero extends. > >> Are the non-AArch64 parts OK? > > Is (vec_merge (subreg ..) (vec_duplicate)) canonicalized to the form > > you handle? I see the (vec_merge (vec_duplicate...) (vec_concat)) case > > also doesn't handle the swapped operand case. > > > > Otherwise the middle-end parts looks ok. > I don't see any explicit canonicalisation code for it. > I've updated the simplify-rtx part to handle the swapped operand case. > Is the attached patch better in this regard? I couldn't think of a clean way to avoid > duplicating some logic (beyond creating a new function away from the callsite). Works for me. Were you able to actually create such RTL from testcases? Segher, do you know where canonicalization rules are documented? IIRC we do not actively try to canonicalize in most cases. Richard. > Thanks, > Kyrill > > Thanks, > > Richard. > > > >> Thanks, > >> James > >> --- > >> 2018-05-15 James Greenhalgh > >>Kyrylo Tkachov > >>* config/aarch64/aarch64.c (aarch64_expand_vector_init): Modify > >>code generation for cases where splatting a value is not useful. > >>* simplify-rtx.c (simplify_ternary_operation): Simplify > >>vec_merge across a vec_duplicate and a paradoxical subreg > > forming a vector > >>mode to a vec_concat. > >> 2018-05-15 James Greenhalgh > >>* gcc.target/aarch64/vect-slp-dup.c: New.
Re: [patch AArch64] Do not perform a vector splat for vector initialisation if it is not useful
On 15/05/18 10:58, Richard Biener wrote: On Tue, May 15, 2018 at 10:20 AM Kyrill Tkachovwrote: Hi all, This is a respin of James's patch from: https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html The original patch was approved and committed but was later reverted because of failures on big-endian. This tweaked version fixes the big-endian failures in aarch64_expand_vector_init by picking the right element of VALS to move into the low part of the vector register depending on endianness. The rest of the patch stays the same. I'm looking for approval on the aarch64 parts, as they are the ones that have changed since the last approved version of the patch. --- In the testcase in this patch we create an SLP vector with only two elements. Our current vector initialisation code will first duplicate the first element to both lanes, then overwrite the top lane with a new value. This duplication can be clunky and wasteful. Better would be to simply use the fact that we will always be overwriting the remaining bits, and simply move the first element to the corrcet place (implicitly zeroing all other bits). This reduces the code generation for this case, and can allow more efficient addressing modes, and other second order benefits for AArch64 code which has been vectorized to V2DI mode. Note that the change is generic enough to catch the case for any vector mode, but is expected to be most useful for 2x64-bit vectorization. Unfortunately, on its own, this would cause failures in gcc.target/aarch64/load_v2vec_lanes_1.c and gcc.target/aarch64/store_v2vec_lanes.c , which expect to see many more vec_merge and vec_duplicate for their simplifications to apply. To fix this, add a special case to the AArch64 code if we are loading from two memory addresses, and use the load_pair_lanes patterns directly. We also need a new pattern in simplify-rtx.c:simplify_ternary_operation , to catch: (vec_merge:OUTER (vec_duplicate:OUTER x:INNER) (subreg:OUTER y:INNER 0) (const_int N)) And simplify it to: (vec_concat:OUTER x:INNER y:INNER) or (vec_concat y x) This is similar to the existing patterns which are tested in this function, without requiring the second operand to also be a vec_duplicate. Bootstrapped and tested on aarch64-none-linux-gnu and tested on aarch64-none-elf. Note that this requires https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html if we don't want to ICE creating broken vector zero extends. Are the non-AArch64 parts OK? Is (vec_merge (subreg ..) (vec_duplicate)) canonicalized to the form you handle? I see the (vec_merge (vec_duplicate...) (vec_concat)) case also doesn't handle the swapped operand case. Otherwise the middle-end parts looks ok. I don't see any explicit canonicalisation code for it. I've updated the simplify-rtx part to handle the swapped operand case. Is the attached patch better in this regard? I couldn't think of a clean way to avoid duplicating some logic (beyond creating a new function away from the callsite). Thanks, Kyrill Thanks, Richard. Thanks, James --- 2018-05-15 James Greenhalgh Kyrylo Tkachov * config/aarch64/aarch64.c (aarch64_expand_vector_init): Modify code generation for cases where splatting a value is not useful. * simplify-rtx.c (simplify_ternary_operation): Simplify vec_merge across a vec_duplicate and a paradoxical subreg forming a vector mode to a vec_concat. 2018-05-15 James Greenhalgh * gcc.target/aarch64/vect-slp-dup.c: New. diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c index a2003fe52875f1653d644347bafd7773d1f01e91..6bf6c05535b61eef1021d46bcd8448fb3a0b25f4 100644 --- a/gcc/config/aarch64/aarch64.c +++ b/gcc/config/aarch64/aarch64.c @@ -13916,9 +13916,54 @@ aarch64_expand_vector_init (rtx target, rtx vals) maxv = matches[i][1]; } - /* Create a duplicate of the most common element. */ - rtx x = copy_to_mode_reg (inner_mode, XVECEXP (vals, 0, maxelement)); - aarch64_emit_move (target, gen_vec_duplicate (mode, x)); + /* Create a duplicate of the most common element, unless all elements + are equally useless to us, in which case just immediately set the + vector register using the first element. */ + + if (maxv == 1) + { + /* For vectors of two 64-bit elements, we can do even better. */ + if (n_elts == 2 + && (inner_mode == E_DImode + || inner_mode == E_DFmode)) + + { + rtx x0 = XVECEXP (vals, 0, 0); + rtx x1 = XVECEXP (vals, 0, 1); + /* Combine can pick up this case, but handling it directly + here leaves clearer RTL. + + This is load_pair_lanes, and also gives us a clean-up + for store_pair_lanes. */ + if (memory_operand (x0,
Re: [patch AArch64] Do not perform a vector splat for vector initialisation if it is not useful
On Tue, May 15, 2018 at 10:20 AM Kyrill Tkachovwrote: > Hi all, > This is a respin of James's patch from: https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html > The original patch was approved and committed but was later reverted because of failures on big-endian. > This tweaked version fixes the big-endian failures in aarch64_expand_vector_init by picking the right > element of VALS to move into the low part of the vector register depending on endianness. The rest of the patch > stays the same. I'm looking for approval on the aarch64 parts, as they are the ones that have changed > since the last approved version of the patch. > --- > In the testcase in this patch we create an SLP vector with only two > elements. Our current vector initialisation code will first duplicate > the first element to both lanes, then overwrite the top lane with a new > value. > This duplication can be clunky and wasteful. > Better would be to simply use the fact that we will always be > overwriting > the remaining bits, and simply move the first element to the corrcet > place > (implicitly zeroing all other bits). > This reduces the code generation for this case, and can allow more > efficient addressing modes, and other second order benefits for AArch64 > code which has been vectorized to V2DI mode. > Note that the change is generic enough to catch the case for any vector > mode, but is expected to be most useful for 2x64-bit vectorization. > Unfortunately, on its own, this would cause failures in > gcc.target/aarch64/load_v2vec_lanes_1.c and > gcc.target/aarch64/store_v2vec_lanes.c , which expect to see many more > vec_merge and vec_duplicate for their simplifications to apply. To fix > this, > add a special case to the AArch64 code if we are loading from two memory > addresses, and use the load_pair_lanes patterns directly. > We also need a new pattern in simplify-rtx.c:simplify_ternary_operation > , to > catch: > (vec_merge:OUTER >(vec_duplicate:OUTER x:INNER) >(subreg:OUTER y:INNER 0) >(const_int N)) > And simplify it to: > (vec_concat:OUTER x:INNER y:INNER) or (vec_concat y x) > This is similar to the existing patterns which are tested in this > function, > without requiring the second operand to also be a vec_duplicate. > Bootstrapped and tested on aarch64-none-linux-gnu and tested on > aarch64-none-elf. > Note that this requires > https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html > if we don't want to ICE creating broken vector zero extends. > Are the non-AArch64 parts OK? Is (vec_merge (subreg ..) (vec_duplicate)) canonicalized to the form you handle? I see the (vec_merge (vec_duplicate...) (vec_concat)) case also doesn't handle the swapped operand case. Otherwise the middle-end parts looks ok. Thanks, Richard. > Thanks, > James > --- > 2018-05-15 James Greenhalgh > Kyrylo Tkachov > * config/aarch64/aarch64.c (aarch64_expand_vector_init): Modify > code generation for cases where splatting a value is not useful. > * simplify-rtx.c (simplify_ternary_operation): Simplify > vec_merge across a vec_duplicate and a paradoxical subreg forming a vector > mode to a vec_concat. > 2018-05-15 James Greenhalgh > * gcc.target/aarch64/vect-slp-dup.c: New.
Re: [patch AArch64] Do not perform a vector splat for vector initialisation if it is not useful
On 19 December 2017 at 00:36, Jeff Lawwrote: > On 12/11/2017 08:44 AM, James Greenhalgh wrote: >> Hi, >> >> In the testcase in this patch we create an SLP vector with only two >> elements. Our current vector initialisation code will first duplicate >> the first element to both lanes, then overwrite the top lane with a new >> value. >> >> This duplication can be clunky and wasteful. >> >> Better would be to simply use the fact that we will always be overwriting >> the remaining bits, and simply move the first element to the corrcet place >> (implicitly zeroing all other bits). >> >> This reduces the code generation for this case, and can allow more >> efficient addressing modes, and other second order benefits for AArch64 >> code which has been vectorized to V2DI mode. >> >> Note that the change is generic enough to catch the case for any vector >> mode, but is expected to be most useful for 2x64-bit vectorization. >> >> Unfortunately, on its own, this would cause failures in >> gcc.target/aarch64/load_v2vec_lanes_1.c and >> gcc.target/aarch64/store_v2vec_lanes.c , which expect to see many more >> vec_merge and vec_duplicate for their simplifications to apply. To fix this, >> add a special case to the AArch64 code if we are loading from two memory >> addresses, and use the load_pair_lanes patterns directly. >> >> We also need a new pattern in simplify-rtx.c:simplify_ternary_operation , to >> catch: >> >> (vec_merge:OUTER >> (vec_duplicate:OUTER x:INNER) >> (subreg:OUTER y:INNER 0) >> (const_int N)) >> >> And simplify it to: >> >> (vec_concat:OUTER x:INNER y:INNER) or (vec_concat y x) >> >> This is similar to the existing patterns which are tested in this function, >> without requiring the second operand to also be a vec_duplicate. >> >> Bootstrapped and tested on aarch64-none-linux-gnu and tested on >> aarch64-none-elf. >> >> Note that this requires >> https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html >> if we don't want to ICE creating broken vector zero extends. >> >> Are the non-AArch64 parts OK? >> >> Thanks, >> James >> >> --- >> 2017-12-11 James Greenhalgh >> >> * config/aarch64/aarch64.c (aarch64_expand_vector_init): Modify code >> generation for cases where splatting a value is not useful. >> * simplify-rtx.c (simplify_ternary_operation): Simplify vec_merge >> across a vec_duplicate and a paradoxical subreg forming a vector >> mode to a vec_concat. >> >> 2017-12-11 James Greenhalgh >> >> * gcc.target/aarch64/vect-slp-dup.c: New. >> >> >> 0001-patch-AArch64-Do-not-perform-a-vector-splat-for-vect.patch >> >> > >> diff --git a/gcc/simplify-rtx.c b/gcc/simplify-rtx.c >> index 806c309..ed16f70 100644 >> --- a/gcc/simplify-rtx.c >> +++ b/gcc/simplify-rtx.c >> @@ -5785,6 +5785,36 @@ simplify_ternary_operation (enum rtx_code code, >> machine_mode mode, >> return simplify_gen_binary (VEC_CONCAT, mode, newop0, newop1); >> } >> >> + /* Replace: >> + >> + (vec_merge:outer (vec_duplicate:outer x:inner) >> +(subreg:outer y:inner 0) >> +(const_int N)) >> + >> + with (vec_concat:outer x:inner y:inner) if N == 1, >> + or (vec_concat:outer y:inner x:inner) if N == 2. >> + >> + Implicitly, this means we have a paradoxical subreg, but such >> + a check is cheap, so make it anyway. > I'm going to assume that N == 1 and N == 3 are handled elsewhere and do > not show up here in practice. > > > So is it advisable to handle the case where the VEC_DUPLICATE and SUBREG > show up in the opposite order? Or is there some canonicalization that > prevents that? > > simplify-rtx bits are OK as-is if we're certain we're not going to get > the alternate ordering of the VEC_MERGE operands. ALso OK if you either > generalize this chunk of code or duplicate & twiddle it to handle the > alternate order. > > I didn't look at the aarch64 specific bits. > Hi James, Your patch (r255946) introduced regressions on aarch64_be: http://people.linaro.org/~christophe.lyon/cross-validation/gcc/trunk/255946/report-build-info.html I filed https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83663 to track this. Thanks, Christophe > jeff >
Re: [patch AArch64] Do not perform a vector splat for vector initialisation if it is not useful
On 12/11/2017 08:44 AM, James Greenhalgh wrote: > Hi, > > In the testcase in this patch we create an SLP vector with only two > elements. Our current vector initialisation code will first duplicate > the first element to both lanes, then overwrite the top lane with a new > value. > > This duplication can be clunky and wasteful. > > Better would be to simply use the fact that we will always be overwriting > the remaining bits, and simply move the first element to the corrcet place > (implicitly zeroing all other bits). > > This reduces the code generation for this case, and can allow more > efficient addressing modes, and other second order benefits for AArch64 > code which has been vectorized to V2DI mode. > > Note that the change is generic enough to catch the case for any vector > mode, but is expected to be most useful for 2x64-bit vectorization. > > Unfortunately, on its own, this would cause failures in > gcc.target/aarch64/load_v2vec_lanes_1.c and > gcc.target/aarch64/store_v2vec_lanes.c , which expect to see many more > vec_merge and vec_duplicate for their simplifications to apply. To fix this, > add a special case to the AArch64 code if we are loading from two memory > addresses, and use the load_pair_lanes patterns directly. > > We also need a new pattern in simplify-rtx.c:simplify_ternary_operation , to > catch: > > (vec_merge:OUTER > (vec_duplicate:OUTER x:INNER) > (subreg:OUTER y:INNER 0) > (const_int N)) > > And simplify it to: > > (vec_concat:OUTER x:INNER y:INNER) or (vec_concat y x) > > This is similar to the existing patterns which are tested in this function, > without requiring the second operand to also be a vec_duplicate. > > Bootstrapped and tested on aarch64-none-linux-gnu and tested on > aarch64-none-elf. > > Note that this requires > https://gcc.gnu.org/ml/gcc-patches/2017-12/msg00614.html > if we don't want to ICE creating broken vector zero extends. > > Are the non-AArch64 parts OK? > > Thanks, > James > > --- > 2017-12-11 James Greenhalgh> > * config/aarch64/aarch64.c (aarch64_expand_vector_init): Modify code > generation for cases where splatting a value is not useful. > * simplify-rtx.c (simplify_ternary_operation): Simplify vec_merge > across a vec_duplicate and a paradoxical subreg forming a vector > mode to a vec_concat. > > 2017-12-11 James Greenhalgh > > * gcc.target/aarch64/vect-slp-dup.c: New. > > > 0001-patch-AArch64-Do-not-perform-a-vector-splat-for-vect.patch > > > diff --git a/gcc/simplify-rtx.c b/gcc/simplify-rtx.c > index 806c309..ed16f70 100644 > --- a/gcc/simplify-rtx.c > +++ b/gcc/simplify-rtx.c > @@ -5785,6 +5785,36 @@ simplify_ternary_operation (enum rtx_code code, > machine_mode mode, > return simplify_gen_binary (VEC_CONCAT, mode, newop0, newop1); > } > > + /* Replace: > + > + (vec_merge:outer (vec_duplicate:outer x:inner) > +(subreg:outer y:inner 0) > +(const_int N)) > + > + with (vec_concat:outer x:inner y:inner) if N == 1, > + or (vec_concat:outer y:inner x:inner) if N == 2. > + > + Implicitly, this means we have a paradoxical subreg, but such > + a check is cheap, so make it anyway. I'm going to assume that N == 1 and N == 3 are handled elsewhere and do not show up here in practice. So is it advisable to handle the case where the VEC_DUPLICATE and SUBREG show up in the opposite order? Or is there some canonicalization that prevents that? simplify-rtx bits are OK as-is if we're certain we're not going to get the alternate ordering of the VEC_MERGE operands. ALso OK if you either generalize this chunk of code or duplicate & twiddle it to handle the alternate order. I didn't look at the aarch64 specific bits. jeff