From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: by sourceware.org (Postfix, from userid 1923) id 75712384F6D0; Thu, 17 Nov 2022 22:27:10 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 75712384F6D0 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gcc.gnu.org; s=default; t=1668724030; bh=bVetkwgDv/haEZRQq6gkYYCkC4RjZUxSkP5gWlN/mTg=; h=From:To:Subject:Date:From; b=jJpDju4t+sxQsWtRFPQK6hD2cjaeuPm0xYVgSoxAd2Ibxaa8n7+5xE+PSKeMylSyb dbMpJTF6c4TlXPu50CmysRdD3Zq/T7FE3Tz9nubpC5EAB6kEl8cycl0W4pKKaKWdnT FuSoXhNr3sB3NxTRBMxhx4NCGMSgYo09sURwFXGw= MIME-Version: 1.0 Content-Transfer-Encoding: 8bit Content-Type: text/plain; charset="utf-8" From: Philipp Tomsich To: gcc-cvs@gcc.gnu.org Subject: [gcc(refs/vendors/vrull/heads/for-upstream)] riscv: Use by-pieces to do overlapping accesses in block_move_straight X-Act-Checkin: gcc X-Git-Author: =?utf-8?q?Christoph_M=C3=BCllner?= X-Git-Refname: refs/vendors/vrull/heads/for-upstream X-Git-Oldrev: 7c3c415a159f182d896927acd7e481d18d89453d X-Git-Newrev: f206e3f545a5ad1358fb69b25b21a5e2b5810110 Message-Id: <20221117222710.75712384F6D0@sourceware.org> Date: Thu, 17 Nov 2022 22:27:10 +0000 (GMT) List-Id: https://gcc.gnu.org/g:f206e3f545a5ad1358fb69b25b21a5e2b5810110 commit f206e3f545a5ad1358fb69b25b21a5e2b5810110 Author: Christoph Müllner Date: Wed Oct 5 02:10:30 2022 +0200 riscv: Use by-pieces to do overlapping accesses in block_move_straight The current implementation of riscv_block_move_straight() emits a couple of load-store pairs with maximum width (e.g. 8-byte for RV64). The remainder is handed over to move_by_pieces(), which emits code based target settings like slow_unaligned_access and overlap_op_by_pieces. move_by_pieces() will emit overlapping memory accesses with maximum width only if the given length exceeds the size of one access (e.g. 15-bytes for 8-byte accesses). This patch changes the implementation of riscv_block_move_straight() such, that it preserves a remainder within the interval [delta..2*delta) instead of [0..delta), so that overlapping memory access may be emitted (if the requirements for them are given). gcc/ChangeLog: * config/riscv/riscv-string.c (riscv_block_move_straight): Adjust range for emitted load/store pairs. (riscv_expand_block_move): Fix duplicated extraction of length. gcc/testsuite/ChangeLog: * gcc.target/riscv/memcpy-overlapping.c: Adjust test for the improved expected instruction sequence. Signed-off-by: Christoph Müllner Diff: --- gcc/config/riscv/riscv-string.cc | 8 ++++---- gcc/testsuite/gcc.target/riscv/memcpy-overlapping.c | 19 ++++++++----------- 2 files changed, 12 insertions(+), 15 deletions(-) diff --git a/gcc/config/riscv/riscv-string.cc b/gcc/config/riscv/riscv-string.cc index 6882f0be269..1137df475be 100644 --- a/gcc/config/riscv/riscv-string.cc +++ b/gcc/config/riscv/riscv-string.cc @@ -57,18 +57,18 @@ riscv_block_move_straight (rtx dest, rtx src, unsigned HOST_WIDE_INT length) delta = bits / BITS_PER_UNIT; /* Allocate a buffer for the temporary registers. */ - regs = XALLOCAVEC (rtx, length / delta); + regs = XALLOCAVEC (rtx, length / delta - 1); /* Load as many BITS-sized chunks as possible. Use a normal load if the source has enough alignment, otherwise use left/right pairs. */ - for (offset = 0, i = 0; offset + delta <= length; offset += delta, i++) + for (offset = 0, i = 0; offset + 2 * delta <= length; offset += delta, i++) { regs[i] = gen_reg_rtx (mode); riscv_emit_move (regs[i], adjust_address (src, mode, offset)); } /* Copy the chunks to the destination. */ - for (offset = 0, i = 0; offset + delta <= length; offset += delta, i++) + for (offset = 0, i = 0; offset + 2 * delta <= length; offset += delta, i++) riscv_emit_move (adjust_address (dest, mode, offset), regs[i]); /* Mop up any left-over bytes. */ @@ -166,7 +166,7 @@ riscv_expand_block_move (rtx dest, rtx src, rtx length) if (hwi_length <= (RISCV_MAX_MOVE_BYTES_STRAIGHT / factor)) { - riscv_block_move_straight (dest, src, INTVAL (length)); + riscv_block_move_straight (dest, src, hwi_length); return true; } else if (optimize && align >= BITS_PER_WORD) diff --git a/gcc/testsuite/gcc.target/riscv/memcpy-overlapping.c b/gcc/testsuite/gcc.target/riscv/memcpy-overlapping.c index ffb7248bfd1..ef95bfb879b 100644 --- a/gcc/testsuite/gcc.target/riscv/memcpy-overlapping.c +++ b/gcc/testsuite/gcc.target/riscv/memcpy-overlapping.c @@ -25,26 +25,23 @@ COPY_N(15) /* Emits 2x {ld,sd} and 1x {lw,sw}. */ COPY_N(19) -/* Emits 3x ld and 3x sd. */ +/* Emits 3x {ld,sd}. */ COPY_N(23) /* The by-pieces infrastructure handles up to 24 bytes. So the code below is emitted via cpymemsi/block_move_straight. */ -/* Emits 3x {ld,sd} and 1x {lhu,lbu,sh,sb}. */ +/* Emits 3x {ld,sd} and 1x {lw,sw}. */ COPY_N(27) -/* Emits 3x {ld,sd} and 1x {lw,lbu,sw,sb}. */ +/* Emits 4x {ld,sd}. */ COPY_N(29) -/* Emits 3x {ld,sd} and 2x {lw,sw}. */ +/* Emits 4x {ld,sd}. */ COPY_N(31) -/* { dg-final { scan-assembler-times "ld\t" 21 } } */ -/* { dg-final { scan-assembler-times "sd\t" 21 } } */ +/* { dg-final { scan-assembler-times "ld\t" 23 } } */ +/* { dg-final { scan-assembler-times "sd\t" 23 } } */ -/* { dg-final { scan-assembler-times "lw\t" 5 } } */ -/* { dg-final { scan-assembler-times "sw\t" 5 } } */ - -/* { dg-final { scan-assembler-times "lbu\t" 2 } } */ -/* { dg-final { scan-assembler-times "sb\t" 2 } } */ +/* { dg-final { scan-assembler-times "lw\t" 3 } } */ +/* { dg-final { scan-assembler-times "sw\t" 3 } } */