From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: by sourceware.org (Postfix, from userid 2017) id 296783857355; Mon, 29 Aug 2022 09:18:10 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 296783857355 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gcc.gnu.org; s=default; t=1661764690; bh=exLGU0YV+nb0xiJBhTJtqlSjMV3fUI1vILybBY0dtSA=; h=From:To:Subject:Date:From; b=APtBK0UbHXd6dXsgcYDNeHBPVKrqnrb97rm5tSbj90fXyIzxqRLRUyIMM4JJrNlfk wysM9kmLhomsfDiBPjiJ/lTRoBMszugH0zziQ5RNh6CTjmJSSrDg/hjy/HqlkpjNsb v5g/lLejgkscjd24vtpEjbL+Cy/UZG81bzaWi0J8= MIME-Version: 1.0 Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset="utf-8" From: Robin Dapp To: gcc-cvs@gcc.gnu.org Subject: [gcc r13-2250] s390: Recognize reverse/element swap permute patterns. X-Act-Checkin: gcc X-Git-Author: Robin Dapp X-Git-Refname: refs/heads/master X-Git-Oldrev: bde28df1933b972e080613a00b6819413aefa2e6 X-Git-Newrev: 34895b6606561a36ffded81efee2904feb9c1bfd Message-Id: <20220829091810.296783857355@sourceware.org> Date: Mon, 29 Aug 2022 09:18:10 +0000 (GMT) List-Id: https://gcc.gnu.org/g:34895b6606561a36ffded81efee2904feb9c1bfd commit r13-2250-g34895b6606561a36ffded81efee2904feb9c1bfd Author: Robin Dapp Date: Fri Jun 24 15:17:08 2022 +0200 s390: Recognize reverse/element swap permute patterns. This adds functions to recognize reverse/element swap permute patterns for vler, vster as well as vpdi and rotate. gcc/ChangeLog: * config/s390/s390.cc (expand_perm_with_vpdi): Recognize swap pattern. (is_reverse_perm_mask): New function. (expand_perm_with_rot): Recognize reverse pattern. (expand_perm_with_vstbrq): New function. (expand_perm_with_vster): Use vler/vster for element reversal on z15. (vectorize_vec_perm_const_1): Use. (s390_vectorize_vec_perm_const): Add expand functions. * config/s390/vx-builtins.md: Prefer vster over vler. gcc/testsuite/ChangeLog: * gcc.target/s390/vector/vperm-rev-z14.c: New test. * gcc.target/s390/vector/vperm-rev-z15.c: New test. * gcc.target/s390/zvector/vec-reve-store-byte.c: Adjust test expectation. Diff: --- gcc/config/s390/s390.cc | 119 ++++++++++++++++++++- gcc/config/s390/vx-builtins.md | 21 ++++ .../gcc.target/s390/vector/vperm-rev-z14.c | 87 +++++++++++++++ .../gcc.target/s390/vector/vperm-rev-z15.c | 118 ++++++++++++++++++++ .../gcc.target/s390/zvector/vec-reve-store-byte.c | 6 +- 5 files changed, 345 insertions(+), 6 deletions(-) diff --git a/gcc/config/s390/s390.cc b/gcc/config/s390/s390.cc index 71c0063f716..a10e9a7b957 100644 --- a/gcc/config/s390/s390.cc +++ b/gcc/config/s390/s390.cc @@ -17124,6 +17124,8 @@ struct expand_vec_perm_d machine_mode vmode; unsigned char nelt; bool testing_p; + bool only_op0; + bool only_op1; }; /* Try to expand the vector permute operation described by D using the @@ -17191,7 +17193,9 @@ expand_perm_with_vpdi (const struct expand_vec_perm_d &d) if (d.perm[0] == 0 && d.perm[1] == 3) vpdi1_p = true; - if (d.perm[0] == 1 && d.perm[1] == 2) + if ((d.perm[0] == 1 && d.perm[1] == 2) + || (d.perm[0] == 1 && d.perm[1] == 0) + || (d.perm[0] == 3 && d.perm[1] == 2)) vpdi4_p = true; if (!vpdi1_p && !vpdi4_p) @@ -17203,15 +17207,107 @@ expand_perm_with_vpdi (const struct expand_vec_perm_d &d) op0_reg = force_reg (GET_MODE (d.op0), d.op0); op1_reg = force_reg (GET_MODE (d.op1), d.op1); + /* If we only reference either of the operands in + the permute mask, just use one of them. */ + if (d.only_op0) + op1_reg = op0_reg; + else if (d.only_op1) + op0_reg = op1_reg; + if (vpdi1_p) emit_insn (gen_vpdi1 (d.vmode, d.target, op0_reg, op1_reg)); - if (vpdi4_p) emit_insn (gen_vpdi4 (d.vmode, d.target, op0_reg, op1_reg)); return true; } +/* Helper that checks if a vector permutation mask D + represents a reversal of the vector's elements. */ +static inline bool +is_reverse_perm_mask (const struct expand_vec_perm_d &d) +{ + for (int i = 0; i < d.nelt; i++) + if (d.perm[i] != d.nelt - i - 1) + return false; + return true; +} + +/* The case of reversing a four-element vector [0, 1, 2, 3] + can be handled by first permuting the doublewords + [2, 3, 0, 1] and subsequently rotating them by 32 bits. */ +static bool +expand_perm_with_rot (const struct expand_vec_perm_d &d) +{ + if (d.nelt != 4) + return false; + + if (d.op0 == d.op1 && is_reverse_perm_mask (d)) + { + if (d.testing_p) + return true; + + rtx tmp = gen_reg_rtx (d.vmode); + rtx op0_reg = force_reg (GET_MODE (d.op0), d.op0); + + emit_insn (gen_vpdi4_2 (d.vmode, tmp, op0_reg, op0_reg)); + if (d.vmode == V4SImode) + emit_insn (gen_rotlv4si3_di (d.target, tmp)); + else if (d.vmode == V4SFmode) + emit_insn (gen_rotlv4sf3_di (d.target, tmp)); + + return true; + } + + return false; +} + +/* If we just reverse the elements, emit an eltswap if we have + vler/vster. */ +static bool +expand_perm_with_vster (const struct expand_vec_perm_d &d) +{ + if (TARGET_VXE2 && d.op0 == d.op1 && is_reverse_perm_mask (d) + && (d.vmode == V2DImode || d.vmode == V2DFmode + || d.vmode == V4SImode || d.vmode == V4SFmode + || d.vmode == V8HImode)) + { + if (d.testing_p) + return true; + + if (d.vmode == V2DImode) + emit_insn (gen_eltswapv2di (d.target, d.op0)); + else if (d.vmode == V2DFmode) + emit_insn (gen_eltswapv2df (d.target, d.op0)); + else if (d.vmode == V4SImode) + emit_insn (gen_eltswapv4si (d.target, d.op0)); + else if (d.vmode == V4SFmode) + emit_insn (gen_eltswapv4sf (d.target, d.op0)); + else if (d.vmode == V8HImode) + emit_insn (gen_eltswapv8hi (d.target, d.op0)); + return true; + } + return false; +} + +/* If we reverse a byte-vector this is the same as + byte reversing it which can be done with vstbrq. */ +static bool +expand_perm_with_vstbrq (const struct expand_vec_perm_d &d) +{ + if (TARGET_VXE2 && d.op0 == d.op1 && is_reverse_perm_mask (d) + && d.vmode == V16QImode) + { + if (d.testing_p) + return true; + + emit_insn (gen_eltswapv16qi (d.target, d.op0)); + return true; + } + return false; +} + + /* Try to find the best sequence for the vector permute operation described by D. Return true if the operation could be expanded. */ @@ -17221,9 +17317,18 @@ vectorize_vec_perm_const_1 (const struct expand_vec_perm_d &d) if (expand_perm_with_merge (d)) return true; + if (expand_perm_with_vster (d)) + return true; + + if (expand_perm_with_vstbrq (d)) + return true; + if (expand_perm_with_vpdi (d)) return true; + if (expand_perm_with_rot (d)) + return true; + return false; } @@ -17253,17 +17358,27 @@ s390_vectorize_vec_perm_const (machine_mode vmode, machine_mode op_mode, gcc_assert (VECTOR_MODE_P (d.vmode)); d.nelt = nelt = GET_MODE_NUNITS (d.vmode); d.testing_p = target == NULL_RTX; + d.only_op0 = false; + d.only_op1 = false; gcc_assert (target == NULL_RTX || REG_P (target)); gcc_assert (sel.length () == nelt); + unsigned int highest = 0, lowest = 2 * nelt - 1; for (i = 0; i < nelt; i++) { unsigned char e = sel[i]; + lowest = MIN (lowest, e); + highest = MAX (highest, e); gcc_assert (e < 2 * nelt); d.perm[i] = e; } + if (lowest < nelt && highest < nelt) + d.only_op0 = true; + else if (lowest >= nelt && highest >= nelt) + d.only_op1 = true; + return vectorize_vec_perm_const_1 (d); } diff --git a/gcc/config/s390/vx-builtins.md b/gcc/config/s390/vx-builtins.md index d20771dbe8a..e4bb514556c 100644 --- a/gcc/config/s390/vx-builtins.md +++ b/gcc/config/s390/vx-builtins.md @@ -2186,6 +2186,27 @@ vster\t%v1,%v0" [(set_attr "op_type" "*,VRX,VRX")]) +; The emulation pattern below will also accept +; vst (eltswap (vl)) +; i.e. both operands in memory, which reload needs to fix. +; Split into +; vl +; vster (=vst (eltswap)) +; since we prefer vster over vler as long as the latter +; does not support alignment hints. +(define_split + [(set (match_operand:VEC_HW 0 "memory_operand" "") + (unspec:VEC_HW [(match_operand:VEC_HW 1 "memory_operand" "")] + UNSPEC_VEC_ELTSWAP))] + "TARGET_VXE2 && can_create_pseudo_p ()" + [(set (match_dup 2) (match_dup 1)) + (set (match_dup 0) + (unspec:VEC_HW [(match_dup 2)] UNSPEC_VEC_ELTSWAP))] +{ + operands[2] = gen_reg_rtx (mode); +}) + + ; Swapping v2df/v2di can be done via vpdi on z13 and z14. (define_split [(set (match_operand:V_HW_2 0 "register_operand" "") diff --git a/gcc/testsuite/gcc.target/s390/vector/vperm-rev-z14.c b/gcc/testsuite/gcc.target/s390/vector/vperm-rev-z14.c new file mode 100644 index 00000000000..5c64fac4646 --- /dev/null +++ b/gcc/testsuite/gcc.target/s390/vector/vperm-rev-z14.c @@ -0,0 +1,87 @@ +/* Make sure that the reverse permute patterns are optimized + correctly. */ +/* { dg-do run { target { s390*-*-* } } } */ +/* { dg-options "-O2 -march=z14 -mzarch -fno-unroll-loops" } */ + +/* { dg-final { scan-assembler-times "vpdi\t" 4 } } */ +/* { dg-final { scan-assembler-times "verllg\t" 2 } } */ + +#include + +__attribute__((noipa)) +void reversel (long long *restrict a, long long *restrict b, int n) +{ + for (int i = 0; i < n; i += 2) + { + a[i + 1] = b[i + 0]; + a[i + 0] = b[i + 1]; + } +} + +__attribute__((noipa)) +void reversed (double *restrict a, double *restrict b, int n) +{ + for (int i = 0; i < n; i += 2) + { + a[i + 1] = b[i + 0]; + a[i + 0] = b[i + 1]; + } +} + +__attribute__((noipa)) +void reversei (unsigned int *restrict a, unsigned int *restrict b, int n) +{ + for (int i = 0; i < n; i += 4) + { + a[i + 3] = b[i + 0]; + a[i + 2] = b[i + 1]; + a[i + 1] = b[i + 2]; + a[i + 0] = b[i + 3]; + } +} + +__attribute__((noipa)) +void reversef (float *restrict a, float *restrict b, int n) +{ + for (int i = 0; i < n; i += 4) + { + a[i + 3] = b[i + 0]; + a[i + 2] = b[i + 1]; + a[i + 1] = b[i + 2]; + a[i + 0] = b[i + 3]; + } +} + +int main() +{ + const int n = 1024; + unsigned int u[n], u2[n]; + long long l[n], l2[n]; + double d[n], d2[n]; + float f[n], f2[n]; + + for (int i = 0; i < n; i++) + { + u[i] = i; + l[i] = i; + d[i] = i; + f[i] = i; + u2[i] = i; + l2[i] = i; + d2[i] = i; + f2[i] = i; + } + + reversei (u2, u, n); + reversel (l2, l, n); + reversed (d2, d, n); + reversef (f2, f, n); + + for (int i = 0; i < n - 16; i++) + { + assert (u[i] == u2[i / (16 / sizeof (u[0])) * (16 / sizeof (u[0])) + 16 / sizeof (u[0]) - 1 - i % (16 / sizeof (u[0]))]); + assert (l[i] == l2[i / (16 / sizeof (l[0])) * (16 / sizeof (l[0])) + 16 / sizeof (l[0]) - 1 - i % (16 / sizeof (l[0]))]); + assert (d[i] == d2[i / (16 / sizeof (d[0])) * (16 / sizeof (d[0])) + 16 / sizeof (d[0]) - 1 - i % (16 / sizeof (d[0]))]); + assert (f[i] == f2[i / (16 / sizeof (f[0])) * (16 / sizeof (f[0])) + 16 / sizeof (f[0]) - 1 - i % (16 / sizeof (f[0]))]); + } +} diff --git a/gcc/testsuite/gcc.target/s390/vector/vperm-rev-z15.c b/gcc/testsuite/gcc.target/s390/vector/vperm-rev-z15.c new file mode 100644 index 00000000000..bff52406fa9 --- /dev/null +++ b/gcc/testsuite/gcc.target/s390/vector/vperm-rev-z15.c @@ -0,0 +1,118 @@ +/* Make sure that the reverse permute patterns are optimized + correctly. */ +/* { dg-do run { target { s390*-*-* } } } */ +/* { dg-options "-O2 -march=z15 -mzarch -fno-unroll-loops" } */ + +/* { dg-final { scan-assembler-times "vsterg\t" 2 } } */ +/* { dg-final { scan-assembler-times "vsterf" 2 } } */ +/* { dg-final { scan-assembler-times "vstbrq\t" 1 } } */ +/* { dg-final { scan-assembler-times "vperm" 0 } } */ + +#include + +__attribute__((noipa)) +void reversec (char *restrict a, char *restrict b, int n) +{ + for (int i = 0; i < n; i += 16) + { + a[i + 0] = b[i + 15]; + a[i + 1] = b[i + 14]; + a[i + 2] = b[i + 13]; + a[i + 3] = b[i + 12]; + a[i + 4] = b[i + 11]; + a[i + 5] = b[i + 10]; + a[i + 6] = b[i + 9]; + a[i + 7] = b[i + 8]; + a[i + 8] = b[i + 7]; + a[i + 9] = b[i + 6]; + a[i + 10] = b[i + 5]; + a[i + 11] = b[i + 4]; + a[i + 12] = b[i + 3]; + a[i + 13] = b[i + 2]; + a[i + 14] = b[i + 1]; + a[i + 15] = b[i + 0]; + } +} + +__attribute__((noipa)) +void reversel (long long *restrict a, long long *restrict b, int n) +{ + for (int i = 0; i < n; i += 2) + { + a[i + 1] = b[i + 0]; + a[i + 0] = b[i + 1]; + } +} + +__attribute__((noipa)) +void reversed (double *restrict a, double *restrict b, int n) +{ + for (int i = 0; i < n; i += 2) + { + a[i + 1] = b[i + 0]; + a[i + 0] = b[i + 1]; + } +} + +__attribute__((noipa)) +void reversei (unsigned int *restrict a, unsigned int *restrict b, int n) +{ + for (int i = 0; i < n; i += 4) + { + a[i + 3] = b[i + 0]; + a[i + 2] = b[i + 1]; + a[i + 1] = b[i + 2]; + a[i + 0] = b[i + 3]; + } +} + +__attribute__((noipa)) +void reversef (float *restrict a, float *restrict b, int n) +{ + for (int i = 0; i < n; i += 4) + { + a[i + 3] = b[i + 0]; + a[i + 2] = b[i + 1]; + a[i + 1] = b[i + 2]; + a[i + 0] = b[i + 3]; + } +} + +int main() +{ + const int n = 1024; + char c[n], c2[n]; + unsigned int u[n], u2[n]; + long long l[n], l2[n]; + double d[n], d2[n]; + float f[n], f2[n]; + + for (int i = 0; i < n; i++) + { + c[i] = i; + u[i] = i; + l[i] = i; + d[i] = i; + f[i] = i; + c2[i] = i; + u2[i] = i; + l2[i] = i; + d2[i] = i; + f2[i] = i; + } + + reversec (c2, c, n); + reversei (u2, u, n); + reversel (l2, l, n); + reversed (d2, d, n); + reversef (f2, f, n); + + for (int i = 0; i < n - 16; i++) + { + assert (c[i] == c2[i / (16 / sizeof (c[0])) * (16 / sizeof (c[0])) + 16 / sizeof (c[0]) - 1 - i % (16 / sizeof (c[0]))]); + assert (u[i] == u2[i / (16 / sizeof (u[0])) * (16 / sizeof (u[0])) + 16 / sizeof (u[0]) - 1 - i % (16 / sizeof (u[0]))]); + assert (l[i] == l2[i / (16 / sizeof (l[0])) * (16 / sizeof (l[0])) + 16 / sizeof (l[0]) - 1 - i % (16 / sizeof (l[0]))]); + assert (d[i] == d2[i / (16 / sizeof (d[0])) * (16 / sizeof (d[0])) + 16 / sizeof (d[0]) - 1 - i % (16 / sizeof (d[0]))]); + assert (f[i] == f2[i / (16 / sizeof (f[0])) * (16 / sizeof (f[0])) + 16 / sizeof (f[0]) - 1 - i % (16 / sizeof (f[0]))]); + } +} diff --git a/gcc/testsuite/gcc.target/s390/zvector/vec-reve-store-byte.c b/gcc/testsuite/gcc.target/s390/zvector/vec-reve-store-byte.c index db8284b1f8f..6c061c69fea 100644 --- a/gcc/testsuite/gcc.target/s390/zvector/vec-reve-store-byte.c +++ b/gcc/testsuite/gcc.target/s390/zvector/vec-reve-store-byte.c @@ -16,13 +16,11 @@ bar (signed char *target, vector signed char x) vec_xst (vec_reve (x), 0, target); } -/* { dg-final { scan-assembler-times "vstbrq\t" 2 } } */ - -/* mem -> mem: This becomes vlbrq + vst */ +/* mem -> mem: This becomes vl + vstbrq */ void baz (vector signed char *target, vector signed char *x) { *target = vec_reve (*x); } -/* { dg-final { scan-assembler-times "vlbrq\t" 1 } } */ +/* { dg-final { scan-assembler-times "vstbrq\t" 3 } } */