diff options
author | Tamar Christina <tamar.christina@arm.com> | 2023-08-04 13:49:23 +0100 |
---|---|---|
committer | Tamar Christina <tamar.christina@arm.com> | 2023-08-04 13:49:23 +0100 |
commit | 451391a6477f5b012faeca42cdba1bfb8e6eecc0 (patch) | |
tree | c30162710652747587d1e89ed3f88b7f634eaf62 | |
parent | 6b80071a4d05c9535d3bb26cddaed0895a414f59 (diff) | |
download | gcc-451391a6477f5b012faeca42cdba1bfb8e6eecc0.zip gcc-451391a6477f5b012faeca42cdba1bfb8e6eecc0.tar.gz gcc-451391a6477f5b012faeca42cdba1bfb8e6eecc0.tar.bz2 |
AArch64: Undo vec_widen_<sur>shiftl optabs [PR106346]
In GCC 11 we implemented the vectorizer optab for widening left shifts,
however this optab is only supported for uniform shift constants.
At the moment GCC still has two loop vectorization strategy (classical loop and
SLP based loop vec) and the optab is implemented as a scalar pattern.
This means that when we apply it to a non-uniform constant inside a loop we only
find out during SLP build that the constants aren't uniform. At this point it's
too late and we lose SLP entirely.
Over the years I've tried various options but none of it works well:
1. Dissolving patterns during SLP built (problematic, also dissolves them for
non-slp).
2. Optionally ignoring patterns for SLP build (problematic, ends up interfearing
with relevancy detection).
3. Relaxing contraint on SLP build to allow non-constant values and dissolving
them after SLP build using an SLP pattern. (problematic, ends up breaking
shift reassociation).
As a result we've concluded that for now this pattern should just be removed
and formed during RTL.
The plan is to move this to an SLP only pattern once we remove classical loop
vectorization support from GCC, at which time we can also properly support SVE's
Top and Bottom variants.
This removes the optab and reworks the RTL to recognize both the vector variant
and the intrinsics variant. Also just simplifies all these patterns.
gcc/ChangeLog:
PR target/106346
* config/aarch64/aarch64-simd.md (vec_widen_<sur>shiftl_lo_<mode>,
vec_widen_<sur>shiftl_hi_<mode>): Remove.
(aarch64_<sur>shll<mode>_internal): Renamed to...
(aarch64_<su>shll<mode>): .. This.
(aarch64_<sur>shll2<mode>_internal): Renamed to...
(aarch64_<su>shll2<mode>): .. This.
(aarch64_<sur>shll_n<mode>, aarch64_<sur>shll2_n<mode>): Re-use new
optabs.
* config/aarch64/constraints.md (D2, DL): New.
* config/aarch64/predicates.md (aarch64_simd_shll_imm_vec): New.
gcc/testsuite/ChangeLog:
PR target/106346
* gcc.target/aarch64/pr98772.c: Adjust assembly.
* gcc.target/aarch64/vect-widen-shift.c: New test.
-rw-r--r-- | gcc/config/aarch64/aarch64-simd.md | 118 | ||||
-rw-r--r-- | gcc/config/aarch64/constraints.md | 14 | ||||
-rw-r--r-- | gcc/config/aarch64/predicates.md | 5 | ||||
-rw-r--r-- | gcc/testsuite/gcc.target/aarch64/pr98772.c | 2 | ||||
-rw-r--r-- | gcc/testsuite/gcc.target/aarch64/vect-widen-shift.c | 50 |
5 files changed, 110 insertions, 79 deletions
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md index d953941..f67eb70 100644 --- a/gcc/config/aarch64/aarch64-simd.md +++ b/gcc/config/aarch64/aarch64-simd.md @@ -6387,105 +6387,67 @@ [(set_attr "type" "neon_sat_shift_reg<q>")] ) -(define_expand "vec_widen_<sur>shiftl_lo_<mode>" - [(set (match_operand:<VWIDE> 0 "register_operand" "=w") - (unspec:<VWIDE> [(match_operand:VQW 1 "register_operand" "w") - (match_operand:SI 2 - "aarch64_simd_shift_imm_bitsize_<ve_mode>" "i")] - VSHLL))] - "TARGET_SIMD" - { - rtx p = aarch64_simd_vect_par_cnst_half (<MODE>mode, <nunits>, false); - emit_insn (gen_aarch64_<sur>shll<mode>_internal (operands[0], operands[1], - p, operands[2])); - DONE; - } -) - -(define_expand "vec_widen_<sur>shiftl_hi_<mode>" - [(set (match_operand:<VWIDE> 0 "register_operand") - (unspec:<VWIDE> [(match_operand:VQW 1 "register_operand" "w") - (match_operand:SI 2 - "immediate_operand" "i")] - VSHLL))] - "TARGET_SIMD" - { - rtx p = aarch64_simd_vect_par_cnst_half (<MODE>mode, <nunits>, true); - emit_insn (gen_aarch64_<sur>shll2<mode>_internal (operands[0], operands[1], - p, operands[2])); - DONE; - } -) - ;; vshll_n -(define_insn "aarch64_<sur>shll<mode>_internal" - [(set (match_operand:<VWIDE> 0 "register_operand" "=w") - (unspec:<VWIDE> [(vec_select:<VHALF> - (match_operand:VQW 1 "register_operand" "w") - (match_operand:VQW 2 "vect_par_cnst_lo_half" "")) - (match_operand:SI 3 - "aarch64_simd_shift_imm_bitsize_<ve_mode>" "i")] - VSHLL))] +(define_insn "aarch64_<su>shll<mode>" + [(set (match_operand:<VWIDE> 0 "register_operand") + (ashift:<VWIDE> (ANY_EXTEND:<VWIDE> + (match_operand:VD_BHSI 1 "register_operand")) + (match_operand:<VWIDE> 2 + "aarch64_simd_shll_imm_vec")))] "TARGET_SIMD" - { - if (INTVAL (operands[3]) == GET_MODE_UNIT_BITSIZE (<MODE>mode)) - return "shll\\t%0.<Vwtype>, %1.<Vhalftype>, %3"; - else - return "<sur>shll\\t%0.<Vwtype>, %1.<Vhalftype>, %3"; + {@ [cons: =0, 1, 2] + [w, w, D2] shll\t%0.<Vwtype>, %1.<Vtype>, %I2 + [w, w, DL] <su>shll\t%0.<Vwtype>, %1.<Vtype>, %I2 } [(set_attr "type" "neon_shift_imm_long")] ) -(define_insn "aarch64_<sur>shll2<mode>_internal" - [(set (match_operand:<VWIDE> 0 "register_operand" "=w") - (unspec:<VWIDE> [(vec_select:<VHALF> - (match_operand:VQW 1 "register_operand" "w") - (match_operand:VQW 2 "vect_par_cnst_hi_half" "")) - (match_operand:SI 3 - "aarch64_simd_shift_imm_bitsize_<ve_mode>" "i")] +(define_expand "aarch64_<sur>shll_n<mode>" + [(set (match_operand:<VWIDE> 0 "register_operand") + (unspec:<VWIDE> [(match_operand:VD_BHSI 1 "register_operand") + (match_operand:SI 2 + "aarch64_simd_shift_imm_bitsize_<ve_mode>")] VSHLL))] "TARGET_SIMD" { - if (INTVAL (operands[3]) == GET_MODE_UNIT_BITSIZE (<MODE>mode)) - return "shll2\\t%0.<Vwtype>, %1.<Vtype>, %3"; - else - return "<sur>shll2\\t%0.<Vwtype>, %1.<Vtype>, %3"; + rtx shft = gen_const_vec_duplicate (<VWIDE>mode, operands[2]); + emit_insn (gen_aarch64_<sur>shll<mode> (operands[0], operands[1], shft)); + DONE; } - [(set_attr "type" "neon_shift_imm_long")] ) -(define_insn "aarch64_<sur>shll_n<mode>" - [(set (match_operand:<VWIDE> 0 "register_operand" "=w") - (unspec:<VWIDE> [(match_operand:VD_BHSI 1 "register_operand" "w") - (match_operand:SI 2 - "aarch64_simd_shift_imm_bitsize_<ve_mode>" "i")] - VSHLL))] +;; vshll_high_n + +(define_insn "aarch64_<su>shll2<mode>" + [(set (match_operand:<VWIDE> 0 "register_operand") + (ashift:<VWIDE> (ANY_EXTEND:<VWIDE> + (vec_select:<VHALF> + (match_operand:VQW 1 "register_operand") + (match_operand:VQW 2 "vect_par_cnst_hi_half"))) + (match_operand:<VWIDE> 3 + "aarch64_simd_shll_imm_vec")))] "TARGET_SIMD" - { - if (INTVAL (operands[2]) == GET_MODE_UNIT_BITSIZE (<MODE>mode)) - return "shll\\t%0.<Vwtype>, %1.<Vtype>, %2"; - else - return "<sur>shll\\t%0.<Vwtype>, %1.<Vtype>, %2"; + {@ [cons: =0, 1, 2, 3] + [w, w, , D2] shll2\t%0.<Vwtype>, %1.<Vtype>, %I3 + [w, w, , DL] <su>shll2\t%0.<Vwtype>, %1.<Vtype>, %I3 } [(set_attr "type" "neon_shift_imm_long")] ) -;; vshll_high_n - -(define_insn "aarch64_<sur>shll2_n<mode>" - [(set (match_operand:<VWIDE> 0 "register_operand" "=w") - (unspec:<VWIDE> [(match_operand:VQW 1 "register_operand" "w") - (match_operand:SI 2 "immediate_operand" "i")] - VSHLL))] +(define_expand "aarch64_<sur>shll2_n<mode>" + [(set (match_operand:<VWIDE> 0 "register_operand") + (unspec:<VWIDE> [(match_operand:VQW 1 "register_operand") + (match_operand:SI 2 + "aarch64_simd_shift_imm_bitsize_<ve_mode>")] + VSHLL))] "TARGET_SIMD" { - if (INTVAL (operands[2]) == GET_MODE_UNIT_BITSIZE (<MODE>mode)) - return "shll2\\t%0.<Vwtype>, %1.<Vtype>, %2"; - else - return "<sur>shll2\\t%0.<Vwtype>, %1.<Vtype>, %2"; + rtx shft = gen_const_vec_duplicate (<VWIDE>mode, operands[2]); + rtx p = aarch64_simd_vect_par_cnst_half (<MODE>mode, <nunits>, true); + emit_insn (gen_aarch64_<sur>shll2<mode> (operands[0], operands[1], p, shft)); + DONE; } - [(set_attr "type" "neon_shift_imm_long")] ) ;; vrshr_n diff --git a/gcc/config/aarch64/constraints.md b/gcc/config/aarch64/constraints.md index 6df1dbe..371a008 100644 --- a/gcc/config/aarch64/constraints.md +++ b/gcc/config/aarch64/constraints.md @@ -468,6 +468,20 @@ GET_MODE_UNIT_BITSIZE (mode) - 1, GET_MODE_UNIT_BITSIZE (mode) - 1)"))) +(define_constraint "D2" + "@internal + A constraint that matches vector of immediates that is bits(mode)/2." + (and (match_code "const,const_vector") + (match_test "aarch64_simd_shift_imm_vec_exact_top (op, mode)"))) + +(define_constraint "DL" + "@internal + A constraint that matches vector of immediates for left shift long. + That is immediates between 0 to (bits(mode)/2)-1." + (and (match_code "const,const_vector") + (match_test "aarch64_const_vec_all_same_in_range_p (op, 0, + (GET_MODE_UNIT_BITSIZE (mode) / 2) - 1)"))) + (define_constraint "Dr" "@internal A constraint that matches vector of immediates for right shifts." diff --git a/gcc/config/aarch64/predicates.md b/gcc/config/aarch64/predicates.md index d5a4a1c..2d8d1fe 100644 --- a/gcc/config/aarch64/predicates.md +++ b/gcc/config/aarch64/predicates.md @@ -638,6 +638,11 @@ HOST_WIDE_INT_1U << (GET_MODE_UNIT_BITSIZE (mode) / 2 - 1))"))) +(define_predicate "aarch64_simd_shll_imm_vec" + (and (match_code "const_vector") + (match_test "aarch64_const_vec_all_same_in_range_p (op, 0, + GET_MODE_UNIT_BITSIZE (mode) / 2)"))) + (define_predicate "aarch64_simd_shift_imm_bitsize_qi" (and (match_code "const_int") (match_test "IN_RANGE (INTVAL (op), 0, 8)"))) diff --git a/gcc/testsuite/gcc.target/aarch64/pr98772.c b/gcc/testsuite/gcc.target/aarch64/pr98772.c index 8259251..52ad012 100644 --- a/gcc/testsuite/gcc.target/aarch64/pr98772.c +++ b/gcc/testsuite/gcc.target/aarch64/pr98772.c @@ -155,4 +155,4 @@ int main () /* { dg-final { scan-assembler-times "uaddl\\tv" 2 } } */ /* { dg-final { scan-assembler-times "usubl\\tv" 2 } } */ /* { dg-final { scan-assembler-times "umull\\tv" 2 } } */ -/* { dg-final { scan-assembler-times "shl\\tv" 2 } } */ +/* { dg-final { scan-assembler-times "shll\\tv" 2 } } */ diff --git a/gcc/testsuite/gcc.target/aarch64/vect-widen-shift.c b/gcc/testsuite/gcc.target/aarch64/vect-widen-shift.c new file mode 100644 index 0000000..6ee41f6 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/vect-widen-shift.c @@ -0,0 +1,50 @@ +/* { dg-do compile } */ +/* { dg-options "-O3 -save-temps" } */ +/* { dg-final { check-function-bodies "**" "" "" } } */ +#include <stdint.h> +#include <string.h> + +#pragma GCC target "+nosve" + +#define ARR_SIZE 1024 + +/* Should produce an shll,shll2 pair*/ +/* +** sshll_opt1: +** ... +** shll v[0-9]+.4s, v[0-9]+.4h, 16 +** shll2 v[0-9]+.4s, v[0-9]+.8h, 16 +** ... +*/ +void sshll_opt1 (int32_t *foo, int16_t *a, int16_t *b) +{ + for( int i = 0; i < ARR_SIZE - 3;i=i+4) + { + foo[i] = a[i] << 16; + foo[i+1] = a[i+1] << 16; + foo[i+2] = a[i+2] << 16; + foo[i+3] = a[i+3] << 16; + } +} + +/* +** sshll_opt2: +** ... +** sxtl v[0-9]+.4s, v[0-9]+.4h +** sxtl2 v[0-9]+.4s, v[0-9]+.8h +** sshl v[0-9]+.4s, v[0-9]+.4s, v[0-9]+.4s +** sshl v[0-9]+.4s, v[0-9]+.4s, v[0-9]+.4s +** ... +*/ +void sshll_opt2 (int32_t *foo, int16_t *a, int16_t *b) +{ + for( int i = 0; i < ARR_SIZE - 3;i=i+4) + { + foo[i] = a[i] << 16; + foo[i+1] = a[i+1] << 15; + foo[i+2] = a[i+2] << 14; + foo[i+3] = a[i+3] << 17; + } +} + + |