From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: by sourceware.org (Postfix, from userid 7921) id 1C7713858D1E; Tue, 25 Apr 2023 17:09:20 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 1C7713858D1E DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gcc.gnu.org; s=default; t=1682442560; bh=YHos7/ysMHtKEZjLpyYL2bP+XsY90pH61kGBFN+Xb4w=; h=From:To:Subject:Date:From; b=x9n+1Fpx0Qgwmc9hMwWkC64nhRTTWDNdBPOuLoO+VQDToNQOx7zKuci2xPYfYj0rq AX7U6aw3/npfNXF5O2sxD5T9pybfDYxiB/W4zYBzN7dtHhQzBHRSG+af5b+QKJQx85 YXVYrfvysd9G6Y45HY6ZwtoVBLyK5k6ytBhvEE4M= MIME-Version: 1.0 Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset="utf-8" From: Vineet Gupta To: gcc-cvs@gcc.gnu.org Subject: [gcc r14-227] riscv: relax splitter restrictions for creating pseudos X-Act-Checkin: gcc X-Git-Author: Vineet Gupta X-Git-Refname: refs/heads/master X-Git-Oldrev: e262cdf49c52d2c6d6aedb30e9ce67e68a9f62a9 X-Git-Newrev: 0530254413f8d356805442f659c6d2921739284d Message-Id: <20230425170920.1C7713858D1E@sourceware.org> Date: Tue, 25 Apr 2023 17:09:20 +0000 (GMT) List-Id: https://gcc.gnu.org/g:0530254413f8d356805442f659c6d2921739284d commit r14-227-g0530254413f8d356805442f659c6d2921739284d Author: Vineet Gupta Date: Tue Apr 11 14:05:29 2023 -0700 riscv: relax splitter restrictions for creating pseudos [partial addressing of PR/109279] RISCV splitters have restrictions to not create pesudos due to a combine limitatation. And despite this being a split-during-combine limitation, all split passes take the hit due to way define*_split are used in gcc. With the original combine issue being fixed 61bee6aed2 ("combine: Don't record for UNDO_MODE pointers into regno_reg_rtx array [PR104985]") the RV splitters can now be relaxed. This improves the codegen in general. e.g. long long f(void) { return 0x0101010101010101ull; } Before li a0,0x01010000 addi a0,0x0101 slli a0,a0,16 addi a0,a0,0x0101 slli a0,a0,16 addi a0,a0,0x0101 ret With patch li a5,0x01010000 addi a5,a5,0x0101 mv a0,a5 slli a5,a5,32 add a0,a5,a0 ret This reduces the qemu icounts, even if slightly, across SPEC2017. 500.perlbench_r 0 1235310737733 1231742384460 0.29% 1 744489708820 743515759958 2 714072106766 712875768625 0.17% 502.gcc_r 0 197365353269 197178223030 1 235614445254 235465240341 2 226769189971 226604663947 3 188315686133 188123584015 4 289372107644 289187945424 503.bwaves_r 0 326291538768 326291539697 1 515809487294 515809488863 2 401647004144 401647005463 3 488750661035 488750662484 505.mcf_r 0 681926695281 681925418147 507.cactuBSSN_r 0 3832240965352 3832226068734 508.namd_r 0 1919838790866 1919832527292 510.parest_r 0 3515999635520 3515878553435 511.povray_r 0 3073889223775 3074758622749 519.lbm_r 0 1194077464296 1194077464041 520.omnetpp_r 0 1014144252460 1011530791131 0.26% 521.wrf_r 0 3966715533120 3966265425092 523.xalancbmk_r 0 1064914296949 1064506711802 525.x264_r 0 509290028335 509258131632 1 2001424246635 2001677767181 2 1914660798226 1914869407575 526.blender_r 0 1726083839515 1725974286174 527.cam4_r 0 2336526136415 2333656336419 531.deepsjeng_r 0 1689007489539 1686541299243 0.15% 538.imagick_r 0 3247960667520 3247942048723 541.leela_r 0 2072315300365 2070248271250 544.nab_r 0 1527909091282 1527906483039 548.exchange2_r 0 2086120304280 2086314757502 549.fotonik3d_r 0 2261694058444 2261670330720 554.roms_r 0 2640547903140 2640512733483 557.xz_r 0 388736881767 386880875636 0.48% 1 959356981818 959993132842 2 547643353034 546374038310 0.23% 997.specrand_fr 0 512881578 512599641 999.specrand_ir 0 512881578 512599641 This is testsuite clean, no regression w/ patch. ========= Summary of gcc testsuite ========= | # of unexpected case / # of unique unexpected case | gcc | g++ | gfortran | rv64imafdc/ lp64d/ medlow | 2 / 2 | 1 / 1 | 6 / 1 | rv64imac/ lp64/ medlow | 3 / 3 | 1 / 1 | 43 / 8 | rv32imafdc/ ilp32d/ medlow | 1 / 1 | 3 / 2 | 6 / 1 | rv32imac/ ilp32/ medlow | 1 / 1 | 3 / 2 | 43 / 8 | This came up as part of IRC chat on PR/109279 and was suggested by Andrew Pinski. gcc/ChangeLog: * config/riscv/riscv.md: riscv_move_integer() drop in_splitter arg. riscv_split_symbol() drop in_splitter arg. * config/riscv/riscv.cc: riscv_move_integer() drop in_splitter arg. riscv_split_symbol() drop in_splitter arg. riscv_force_temporary() drop in_splitter arg. * config/riscv/riscv-protos.h: riscv_move_integer() drop in_splitter arg. riscv_split_symbol() drop in_splitter arg. Signed-off-by: Vineet Gupta Diff: --- gcc/config/riscv/riscv-protos.h | 4 ++-- gcc/config/riscv/riscv.cc | 46 ++++++++++++++++------------------------- gcc/config/riscv/riscv.md | 8 +++---- 3 files changed, 24 insertions(+), 34 deletions(-) diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h index 5244e8dcbf0..607ff6ea697 100644 --- a/gcc/config/riscv/riscv-protos.h +++ b/gcc/config/riscv/riscv-protos.h @@ -44,10 +44,10 @@ extern int riscv_const_insns (rtx); extern int riscv_split_const_insns (rtx); extern int riscv_load_store_insns (rtx, rtx_insn *); extern rtx riscv_emit_move (rtx, rtx); -extern bool riscv_split_symbol (rtx, rtx, machine_mode, rtx *, bool); +extern bool riscv_split_symbol (rtx, rtx, machine_mode, rtx *); extern bool riscv_split_symbol_type (enum riscv_symbol_type); extern rtx riscv_unspec_address (rtx, enum riscv_symbol_type); -extern void riscv_move_integer (rtx, rtx, HOST_WIDE_INT, machine_mode, bool); +extern void riscv_move_integer (rtx, rtx, HOST_WIDE_INT, machine_mode); extern bool riscv_legitimize_move (machine_mode, rtx, rtx); extern rtx riscv_subword (rtx, bool); extern bool riscv_split_64bit_move_p (rtx, rtx); diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc index 5d2550871c7..ac8e4420896 100644 --- a/gcc/config/riscv/riscv.cc +++ b/gcc/config/riscv/riscv.cc @@ -689,8 +689,8 @@ riscv_split_integer (HOST_WIDE_INT val, machine_mode mode) unsigned HOST_WIDE_INT hival = sext_hwi ((val - loval) >> 32, 32); rtx hi = gen_reg_rtx (mode), lo = gen_reg_rtx (mode); - riscv_move_integer (hi, hi, hival, mode, FALSE); - riscv_move_integer (lo, lo, loval, mode, FALSE); + riscv_move_integer (hi, hi, hival, mode); + riscv_move_integer (lo, lo, loval, mode); hi = gen_rtx_fmt_ee (ASHIFT, mode, hi, GEN_INT (32)); hi = force_reg (mode, hi); @@ -1342,12 +1342,9 @@ riscv_swap_instruction (rtx inst) are allowed, copy it into a new register, otherwise use DEST. */ static rtx -riscv_force_temporary (rtx dest, rtx value, bool in_splitter) +riscv_force_temporary (rtx dest, rtx value) { - /* We can't call gen_reg_rtx from a splitter, because this might realloc - the regno_reg_rtx array, which would invalidate reg rtx pointers in the - combine undo buffer. */ - if (can_create_pseudo_p () && !in_splitter) + if (can_create_pseudo_p ()) return force_reg (Pmode, value); else { @@ -1406,7 +1403,7 @@ static rtx riscv_unspec_offset_high (rtx temp, rtx addr, enum riscv_symbol_type symbol_type) { addr = gen_rtx_HIGH (Pmode, riscv_unspec_address (addr, symbol_type)); - return riscv_force_temporary (temp, addr, FALSE); + return riscv_force_temporary (temp, addr); } /* Load an entry from the GOT for a TLS GD access. */ @@ -1454,8 +1451,7 @@ static rtx riscv_tls_add_tp_le (rtx dest, rtx base, rtx sym) is guaranteed to be a legitimate address for mode MODE. */ bool -riscv_split_symbol (rtx temp, rtx addr, machine_mode mode, rtx *low_out, - bool in_splitter) +riscv_split_symbol (rtx temp, rtx addr, machine_mode mode, rtx *low_out) { enum riscv_symbol_type symbol_type; @@ -1471,7 +1467,7 @@ riscv_split_symbol (rtx temp, rtx addr, machine_mode mode, rtx *low_out, case SYMBOL_ABSOLUTE: { rtx high = gen_rtx_HIGH (Pmode, copy_rtx (addr)); - high = riscv_force_temporary (temp, high, in_splitter); + high = riscv_force_temporary (temp, high); *low_out = gen_rtx_LO_SUM (Pmode, high, addr); } break; @@ -1530,9 +1526,8 @@ riscv_add_offset (rtx temp, rtx reg, HOST_WIDE_INT offset) overflow, so we need to force a sign-extension check. */ high = gen_int_mode (CONST_HIGH_PART (offset), Pmode); offset = CONST_LOW_PART (offset); - high = riscv_force_temporary (temp, high, FALSE); - reg = riscv_force_temporary (temp, gen_rtx_PLUS (Pmode, high, reg), - FALSE); + high = riscv_force_temporary (temp, high); + reg = riscv_force_temporary (temp, gen_rtx_PLUS (Pmode, high, reg)); } return plus_constant (Pmode, reg, offset); } @@ -1662,7 +1657,7 @@ riscv_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED, return riscv_legitimize_tls_address (x); /* See if the address can split into a high part and a LO_SUM. */ - if (riscv_split_symbol (NULL, x, mode, &addr, FALSE)) + if (riscv_split_symbol (NULL, x, mode, &addr)) return riscv_force_address (addr, mode); /* Handle BASE + OFFSET. */ @@ -1693,24 +1688,19 @@ riscv_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED, void riscv_move_integer (rtx temp, rtx dest, HOST_WIDE_INT value, - machine_mode orig_mode, bool in_splitter) + machine_mode orig_mode) { struct riscv_integer_op codes[RISCV_MAX_INTEGER_OPS]; machine_mode mode; int i, num_ops; rtx x; - /* We can't call gen_reg_rtx from a splitter, because this might realloc - the regno_reg_rtx array, which would invalidate reg rtx pointers in the - combine undo buffer. */ - bool can_create_pseudo = can_create_pseudo_p () && ! in_splitter; - mode = GET_MODE (dest); /* We use the original mode for the riscv_build_integer call, because HImode values are given special treatment. */ num_ops = riscv_build_integer (codes, value, orig_mode); - if (can_create_pseudo && num_ops > 2 /* not a simple constant */ + if (can_create_pseudo_p () && num_ops > 2 /* not a simple constant */ && num_ops >= riscv_split_integer_cost (value)) x = riscv_split_integer (value, mode); else @@ -1721,7 +1711,7 @@ riscv_move_integer (rtx temp, rtx dest, HOST_WIDE_INT value, for (i = 1; i < num_ops; i++) { - if (!can_create_pseudo) + if (!can_create_pseudo_p ()) x = riscv_emit_set (temp, x); else x = force_reg (mode, x); @@ -1745,12 +1735,12 @@ riscv_legitimize_const_move (machine_mode mode, rtx dest, rtx src) /* Split moves of big integers into smaller pieces. */ if (splittable_const_int_operand (src, mode)) { - riscv_move_integer (dest, dest, INTVAL (src), mode, FALSE); + riscv_move_integer (dest, dest, INTVAL (src), mode); return; } /* Split moves of symbolic constants into high/low pairs. */ - if (riscv_split_symbol (dest, src, MAX_MACHINE_MODE, &src, FALSE)) + if (riscv_split_symbol (dest, src, MAX_MACHINE_MODE, &src)) { riscv_emit_set (dest, src); return; @@ -1771,7 +1761,7 @@ riscv_legitimize_const_move (machine_mode mode, rtx dest, rtx src) if (offset != const0_rtx && (targetm.cannot_force_const_mem (mode, src) || can_create_pseudo_p ())) { - base = riscv_force_temporary (dest, base, FALSE); + base = riscv_force_temporary (dest, base); riscv_emit_move (dest, riscv_add_offset (NULL, base, INTVAL (offset))); return; } @@ -1780,7 +1770,7 @@ riscv_legitimize_const_move (machine_mode mode, rtx dest, rtx src) /* When using explicit relocs, constant pool references are sometimes not legitimate addresses. */ - riscv_split_symbol (dest, XEXP (src, 0), mode, &XEXP (src, 0), FALSE); + riscv_split_symbol (dest, XEXP (src, 0), mode, &XEXP (src, 0)); riscv_emit_move (dest, src); } @@ -2129,7 +2119,7 @@ riscv_legitimize_move (machine_mode mode, rtx dest, rtx src) if (splittable_const_int_operand (src, mode)) { reg = gen_reg_rtx (promoted_mode); - riscv_move_integer (reg, reg, INTVAL (src), mode, FALSE); + riscv_move_integer (reg, reg, INTVAL (src), mode); } else reg = force_reg (promoted_mode, src); diff --git a/gcc/config/riscv/riscv.md b/gcc/config/riscv/riscv.md index 0c69407ac09..dd845cc1ed3 100644 --- a/gcc/config/riscv/riscv.md +++ b/gcc/config/riscv/riscv.md @@ -1669,7 +1669,7 @@ [(const_int 0)] { riscv_move_integer (operands[2], operands[0], INTVAL (operands[1]), - mode, TRUE); + mode); DONE; }) @@ -1678,11 +1678,11 @@ [(set (match_operand:P 0 "register_operand") (match_operand:P 1)) (clobber (match_operand:P 2 "register_operand"))] - "riscv_split_symbol (operands[2], operands[1], MAX_MACHINE_MODE, NULL, TRUE)" + "riscv_split_symbol (operands[2], operands[1], MAX_MACHINE_MODE, NULL)" [(set (match_dup 0) (match_dup 3))] { riscv_split_symbol (operands[2], operands[1], - MAX_MACHINE_MODE, &operands[3], TRUE); + MAX_MACHINE_MODE, &operands[3]); }) ;; Pretend to have the ability to load complex const_int in order to get @@ -1699,7 +1699,7 @@ [(const_int 0)] { riscv_move_integer (operands[0], operands[0], INTVAL (operands[1]), - mode, TRUE); + mode); DONE; })