From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: by sourceware.org (Postfix, from userid 1666) id 09B5938356A5; Fri, 22 Jul 2022 11:21:48 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 09B5938356A5 MIME-Version: 1.0 Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset="utf-8" From: Richard Biener To: gcc-cvs@gcc.gnu.org Subject: [gcc r11-10167] tree-optimization/103116 - SLP permutes and peeling for gaps X-Act-Checkin: gcc X-Git-Author: Richard Biener X-Git-Refname: refs/heads/releases/gcc-11 X-Git-Oldrev: 7cfef93a241e89f7e019cd1048808aa8e8c07836 X-Git-Newrev: ff408622a5f8fb395e5584a494aaaad05b3611db Message-Id: <20220722112148.09B5938356A5@sourceware.org> Date: Fri, 22 Jul 2022 11:21:48 +0000 (GMT) X-BeenThere: gcc-cvs@gcc.gnu.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Gcc-cvs mailing list List-Unsubscribe: , List-Archive: List-Help: List-Subscribe: , X-List-Received-Date: Fri, 22 Jul 2022 11:21:48 -0000 https://gcc.gnu.org/g:ff408622a5f8fb395e5584a494aaaad05b3611db commit r11-10167-gff408622a5f8fb395e5584a494aaaad05b3611db Author: Richard Biener Date: Wed May 4 10:43:07 2022 +0200 tree-optimization/103116 - SLP permutes and peeling for gaps The testcase shows that we can end up with a contiguous access across loop iterations but by means of permutations the elements accessed might only cover parts of a vector. In this case we end up with GROUP_GAP == 0 but still need to avoid accessing excess elements in the last loop iterations. Peeling for gaps is designed to cover this but a single scalar iteration might not cover all of the excess elements. The following ensures peeling for gaps is done in this situation and when that isn't sufficient because we need to peel more than one iteration (gcc.dg/vect/pr103116-2.c), fail the SLP vectorization. 2022-05-04 Richard Biener PR tree-optimization/103116 * tree-vect-stmts.c (get_group_load_store_type): Handle the case we need peeling for gaps even though GROUP_GAP is zero. * gcc.dg/vect/pr103116-1.c: New testcase. * gcc.dg/vect/pr103116-2.c: Likewise. (cherry picked from commit 52b7b86f8c72eb19e637f1e72ffd10f39b8cb829) Diff: --- gcc/testsuite/gcc.dg/vect/pr103116-1.c | 50 ++++++++++++++++++++++++++++ gcc/testsuite/gcc.dg/vect/pr103116-2.c | 59 ++++++++++++++++++++++++++++++++++ gcc/tree-vect-stmts.c | 31 ++++++++++++++++++ 3 files changed, 140 insertions(+) diff --git a/gcc/testsuite/gcc.dg/vect/pr103116-1.c b/gcc/testsuite/gcc.dg/vect/pr103116-1.c new file mode 100644 index 00000000000..d3639fc8cfd --- /dev/null +++ b/gcc/testsuite/gcc.dg/vect/pr103116-1.c @@ -0,0 +1,50 @@ +/* { dg-require-effective-target mmap } */ + +#include +#include + +#define COUNT 128 +#define MMAP_SIZE 0x20000 +#define ADDRESS 0x1122000000 +#define TYPE unsigned int + +#ifndef MAP_ANONYMOUS +#define MAP_ANONYMOUS MAP_ANON +#endif + +void __attribute__((noipa)) +loop (TYPE *restrict x, TYPE *restrict y) +{ + for (int i = 0; i < COUNT; ++i) + { + x[i * 4] = y[i * 2] + 1; + x[i * 4 + 1] = y[i * 2] + 2; + x[i * 4 + 2] = y[i * 2 + 1] + 3; + x[i * 4 + 3] = y[i * 2 + 1] + 4; + } +} + +TYPE x[COUNT * 4]; + +int +main (void) +{ + void *y; + TYPE *end_y; + + y = mmap ((void *) ADDRESS, MMAP_SIZE, PROT_READ | PROT_WRITE, + MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); + if (y == MAP_FAILED) + { + perror ("mmap"); + return 1; + } + + end_y = (TYPE *) ((char *) y + MMAP_SIZE); + + loop (x, end_y - COUNT * 2); + + return 0; +} + +/* { dg-final { scan-tree-dump "Data access with gaps requires scalar epilogue loop" "vect" { target { vect_perm && vect_int } } } } */ diff --git a/gcc/testsuite/gcc.dg/vect/pr103116-2.c b/gcc/testsuite/gcc.dg/vect/pr103116-2.c new file mode 100644 index 00000000000..2f4ed0f404c --- /dev/null +++ b/gcc/testsuite/gcc.dg/vect/pr103116-2.c @@ -0,0 +1,59 @@ +/* { dg-require-effective-target mmap } */ +/* { dg-additional-options "-mssse3" { target x86_64-*-* i?86-*-* } } */ + +#include +#include +#include "tree-vect.h" + +#define COUNT 128 +#define MMAP_SIZE 0x20000 +#define ADDRESS 0x1122000000 +#define TYPE unsigned short +#define GROUP_SIZE 2 + +#ifndef MAP_ANONYMOUS +#define MAP_ANONYMOUS MAP_ANON +#endif + +void __attribute__((noipa)) +loop (TYPE *restrict x, TYPE *restrict y) +{ + for (int i = 0; i < COUNT; ++i) + { + x[i * 8] = y[i * GROUP_SIZE] + 1; + x[i * 8 + 1] = y[i * GROUP_SIZE] + 2; + x[i * 8 + 2] = y[i * GROUP_SIZE + 1] + 3; + x[i * 8 + 3] = y[i * GROUP_SIZE + 1] + 4; + x[i * 8 + 4] = y[i * GROUP_SIZE] + 5; + x[i * 8 + 5] = y[i * GROUP_SIZE] + 6; + x[i * 8 + 6] = y[i * GROUP_SIZE + 1] + 7; + x[i * 8 + 7] = y[i * GROUP_SIZE + 1] + 8; + } +} + +TYPE x[COUNT * 4]; + +int +main (void) +{ + void *y; + TYPE *end_y; + + check_vect (); + + y = mmap ((void *) ADDRESS, MMAP_SIZE, PROT_READ | PROT_WRITE, + MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); + if (y == MAP_FAILED) + { + perror ("mmap"); + return 1; + } + + end_y = (TYPE *) ((char *) y + MMAP_SIZE); + + loop (x, end_y - COUNT * GROUP_SIZE); + + return 0; +} + +/* { dg-final { scan-tree-dump "peeling for gaps insufficient for access" "vect" { target { vect_perm_short } } } } */ diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c index 0d61b0f51d0..973c6dcea18 100644 --- a/gcc/tree-vect-stmts.c +++ b/gcc/tree-vect-stmts.c @@ -2207,6 +2207,37 @@ get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info, gcc_assert (!loop_vinfo || cmp > 0); *memory_access_type = VMAT_CONTIGUOUS; } + + /* When we have a contiguous access across loop iterations + but the access in the loop doesn't cover the full vector + we can end up with no gap recorded but still excess + elements accessed, see PR103116. Make sure we peel for + gaps if necessary and sufficient and give up if not. */ + if (loop_vinfo + && *memory_access_type == VMAT_CONTIGUOUS + && SLP_TREE_LOAD_PERMUTATION (slp_node).exists () + && !multiple_p (group_size * LOOP_VINFO_VECT_FACTOR (loop_vinfo), + nunits)) + { + unsigned HOST_WIDE_INT cnunits, cvf; + if (!can_overrun_p + || !nunits.is_constant (&cnunits) + || !LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&cvf) + /* Peeling for gaps assumes that a single scalar iteration + is enough to make sure the last vector iteration doesn't + access excess elements. + ??? Enhancements include peeling multiple iterations + or using masked loads with a static mask. */ + || (group_size * cvf) % cnunits + group_size < cnunits) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "peeling for gaps insufficient for " + "access\n"); + return false; + } + overrun_p = true; + } } } else