From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-ed1-f51.google.com (mail-ed1-f51.google.com [209.85.208.51]) by sourceware.org (Postfix) with ESMTPS id 0AB773A5300A; Wed, 5 May 2021 19:37:09 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.3.2 sourceware.org 0AB773A5300A Received: by mail-ed1-f51.google.com with SMTP id h10so3360469edt.13; Wed, 05 May 2021 12:37:08 -0700 (PDT) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20161025; h=x-gm-message-state:from:to:cc:subject:date:message-id:in-reply-to :references:mime-version:content-transfer-encoding; bh=bv+10MNwg27XR1a7T8v63GVmgJhofjH9/jqZWG2kJ8E=; b=H+KuXiLakFDjNFGVBhL+KEnta5Yr4NTi601bU5R8LoAoabn7mXj8064RjU9b90mvNN kU9gnoaSvR+4FpcJNIg0N7MYxE4tyDXn3RjwfxRMOG9v9250ZncHfXcLRzr3/M5UW/ow kEib3RUiHu8oomlKAhCbaZtZ6WnrP6ARM+RyaIfh7A654HJx+esHGCZ4DDpe0sIHn1kl kYOjPYxa8+Wl8cY08EyK5imhMDRreXdqZ09aop7nYyzrfyLr6iAavXaBwSYTxD++bkRV 2o6hxuNLywgTzIqz33pCDr4dW4ENoXrXNeHEJUMffxzkrVVhWcpCsB+TaV/m3JE7036C aGng== X-Gm-Message-State: AOAM531y03NFA/o+TlwSzwB4p0o/ZWl6NBCmLxZx8Jjx19UxLU/7WmEs rjqelk/sDi1gAWHrJhva1mBFyVKUulwEaDoM X-Google-Smtp-Source: ABdhPJxqKMNEuD1C2CIQwrxTbZFGZFZ+d0NKeHfFA0Dz0nBz+QbGGLYvCNXp60760PvqgplS74p1jg== X-Received: by 2002:aa7:de8b:: with SMTP id j11mr663447edv.363.1620243427757; Wed, 05 May 2021 12:37:07 -0700 (PDT) Received: from beast.fritz.box (62-178-148-172.cable.dynamic.surfer.at. [62.178.148.172]) by smtp.gmail.com with ESMTPSA id t7sm60531eds.26.2021.05.05.12.37.07 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Wed, 05 May 2021 12:37:07 -0700 (PDT) From: Christoph Muellner To: gcc-patches@gcc.gnu.org Cc: Jim Wilson , Kito Cheng , Christoph Muellner Subject: [PATCH v2 09/10] RISC-V: Provide programmatic implementation of CAS [PR 100266] Date: Wed, 5 May 2021 21:36:50 +0200 Message-Id: <20210505193651.2075405-10-cmuellner@gcc.gnu.org> X-Mailer: git-send-email 2.31.1 In-Reply-To: <20210505193651.2075405-1-cmuellner@gcc.gnu.org> References: <20210505193651.2075405-1-cmuellner@gcc.gnu.org> MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-Spam-Status: No, score=-11.7 required=5.0 tests=BAYES_00, FREEMAIL_ENVFROM_END_DIGIT, FREEMAIL_FORGED_FROMDOMAIN, FREEMAIL_FROM, GIT_PATCH_0, HEADER_FROM_DIFFERENT_DOMAINS, KAM_DMARC_STATUS, RCVD_IN_DNSWL_NONE, RCVD_IN_MSPIKE_H2, SPF_HELO_NONE, SPF_PASS, TXREP autolearn=ham autolearn_force=no version=3.4.2 X-Spam-Checker-Version: SpamAssassin 3.4.2 (2018-09-13) on server2.sourceware.org X-BeenThere: gcc-patches@gcc.gnu.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Gcc-patches mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Wed, 05 May 2021 19:37:10 -0000 The existing CAS implementation uses an INSN definition, which provides the core LR/SC sequence. Additionally to that, there is a follow-up code, that evaluates the results and calculates the return values. This has two drawbacks: a) an extension to sub-word CAS implementations is not possible (even if, then it would be unmaintainable), and b) the implementation is hard to maintain/improve. This patch provides a programmatic implementation of CAS, similar like many other architectures are having one. The implementation supports both, RV32 and RV64. Additionally, the implementation does not introduce data dependencies for computation of the return value. Instead, we set the return value (success state of the CAS operation) based on structural information. This approach is also shown in the the RISC-V unpriv spec (as part of the sample code for a compare-and-swap function using LR/SC). The cost of this implementation is a single LI instruction on top, which is actually not required in case of success (it will be overwritten in the success case later). The resulting sequence requires 9 instructions in the success case. The previous implementation required 11 instructions in the succcess case (including a taken branch) and had a "subw;seqz;beqz" sequence, with direct dependencies. Below is the generated code of a 32-bit CAS sequence with the old implementation and the new implementation (ignore the ANDIs below). Old: f00: 419c lw a5,0(a1) f02: 1005272f lr.w a4,(a0) f06: 00f71563 bne a4,a5,f10 f0a: 18c526af sc.w a3,a2,(a0) f0e: faf5 bnez a3,f02 f10: 40f707bb subw a5,a4,a5 f14: 0017b513 seqz a0,a5 f18: c391 beqz a5,f1c f1a: c198 sw a4,0(a1) f1c: 8905 andi a0,a0,1 f1e: 8082 ret New: e28: 4194 lw a3,0(a1) e2a: 4701 li a4,0 e2c: 1005282f lr.w a6,(a0) e30: 00d81963 bne a6,a3,e42 e34: 18c527af sc.w a5,a2,(a0) e38: fbf5 bnez a5,e2c e3a: 4705 li a4,1 e3c: 00177513 andi a0,a4,1 e40: 8082 ret e42: 0105a023 sw a6,0(a1) e46: 00177513 andi a0,a4,1 e4a: 8082 ret gcc/ PR 100266 * config/riscv/riscv-protos.h (riscv_expand_compare_and_swap): New. * config/riscv/riscv.c (riscv_emit_unlikely_jump): New. * config/rsicv/riscv.c (riscv_expand_compare_and_swap): New. * config/rsicv/sync.md (atomic_cas_value_strong): Removed. * config/rsicv/sync.md (atomic_compare_and_swap): Call riscv_expand_compare_and_swap. --- gcc/config/riscv/riscv-protos.h | 1 + gcc/config/riscv/riscv.c | 75 +++++++++++++++++++++++++++++++++ gcc/config/riscv/sync.md | 35 +-------------- 3 files changed, 77 insertions(+), 34 deletions(-) diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h index 43d7224d6941..eb7e67d3b95a 100644 --- a/gcc/config/riscv/riscv-protos.h +++ b/gcc/config/riscv/riscv-protos.h @@ -59,6 +59,7 @@ extern void riscv_expand_int_scc (rtx, enum rtx_code, rtx, rtx); extern void riscv_expand_float_scc (rtx, enum rtx_code, rtx, rtx); extern void riscv_expand_conditional_branch (rtx, enum rtx_code, rtx, rtx); extern void riscv_expand_conditional_move (rtx, rtx, rtx, rtx_code, rtx, rtx); +extern void riscv_expand_compare_and_swap (rtx[]); #endif extern rtx riscv_legitimize_call_address (rtx); extern void riscv_set_return_address (rtx, rtx); diff --git a/gcc/config/riscv/riscv.c b/gcc/config/riscv/riscv.c index 5fe65776e608..a7b18d650daa 100644 --- a/gcc/config/riscv/riscv.c +++ b/gcc/config/riscv/riscv.c @@ -2496,6 +2496,81 @@ riscv_expand_conditional_move (rtx dest, rtx cons, rtx alt, rtx_code code, cons, alt))); } +/* Mark the previous jump instruction as unlikely. */ + +static void +riscv_emit_unlikely_jump (rtx insn) +{ + rtx_insn *jump = emit_jump_insn (insn); + add_reg_br_prob_note (jump, profile_probability::very_unlikely ()); +} + +/* Expand code to perform a compare-and-swap. */ + +extern void +riscv_expand_compare_and_swap (rtx operands[]) +{ + rtx bval, oldval, mem, expval, newval, mod_s, mod_f, scratch, cond1, cond2; + machine_mode mode; + rtx_code_label *begin_label, *end_label; + + bval = operands[0]; + oldval = operands[1]; + mem = operands[2]; + expval = operands[3]; + newval = operands[4]; + mod_s = operands[6]; + mod_f = operands[7]; + mode = GET_MODE (mem); + scratch = gen_reg_rtx (mode); + begin_label = gen_label_rtx (); + end_label = gen_label_rtx (); + + /* No support for sub-word CAS. */ + if (mode == QImode || mode == HImode) + gcc_unreachable (); + + /* We use mod_f for LR and mod_s for SC below, but + RV does not have any guarantees for LR.rl and SC.aq. */ + if (is_mm_acquire (memmodel_base (INTVAL (mod_s))) + && is_mm_relaxed (memmodel_base (INTVAL (mod_f)))) + { + mod_f = GEN_INT (MEMMODEL_ACQUIRE); + mod_s = GEN_INT (MEMMODEL_RELAXED); + } + + /* Since we want to maintain a branch-free good-case, but also want + to not have two branches in the bad-case, we set bval to FALSE + on top of the sequence. In the bad case, we simply jump over + the assignment of bval to TRUE at the end of the sequence. */ + + emit_insn (gen_rtx_SET (bval, gen_rtx_CONST_INT (SImode, FALSE))); + + /* Make sure the scheduler does not move INSNs beyond here. */ + emit_insn (gen_blockage ()); + + emit_label (begin_label); + + emit_insn (gen_riscv_load_reserved (mode, oldval, mem, mod_f)); + + cond1 = gen_rtx_NE (mode, oldval, expval); + riscv_emit_unlikely_jump (gen_cbranch4 (Pmode, cond1, oldval, expval, + end_label)); + + emit_insn (gen_riscv_store_conditional (mode, scratch, mem, newval, mod_s)); + + /* Make sure the scheduler does not move INSNs beyond here. */ + emit_insn (gen_blockage ()); + + cond2 = gen_rtx_NE (mode, scratch, const0_rtx); + riscv_emit_unlikely_jump (gen_cbranch4 (Pmode, cond2, scratch, const0_rtx, + begin_label)); + + emit_insn (gen_rtx_SET (bval, gen_rtx_CONST_INT (SImode, TRUE))); + + emit_label (end_label); +} + /* Implement TARGET_FUNCTION_ARG_BOUNDARY. Every parameter gets at least PARM_BOUNDARY bits of alignment, but will be given anything up to PREFERRED_STACK_BOUNDARY bits if the type requires it. */ diff --git a/gcc/config/riscv/sync.md b/gcc/config/riscv/sync.md index 49b860da8ef0..da8dbf698163 100644 --- a/gcc/config/riscv/sync.md +++ b/gcc/config/riscv/sync.md @@ -207,20 +207,6 @@ "amoswap.%A3 %0,%z2,%1" ) -(define_insn "atomic_cas_value_strong" - [(set (match_operand:GPR 0 "register_operand" "=&r") - (match_operand:GPR 1 "memory_operand" "+A")) - (set (match_dup 1) - (unspec_volatile:GPR [(match_operand:GPR 2 "reg_or_0_operand" "rJ") - (match_operand:GPR 3 "reg_or_0_operand" "rJ") - (match_operand:SI 4 "const_int_operand") ;; mod_s - (match_operand:SI 5 "const_int_operand")] ;; mod_f - UNSPEC_COMPARE_AND_SWAP)) - (clobber (match_scratch:GPR 6 "=&r"))] - "TARGET_ATOMIC" - "1: lr.%A5 %0,%1; bne %0,%z2,1f; sc.%A4 %6,%z3,%1; bnez %6,1b; 1:" - [(set (attr "length") (const_int 20))]) - (define_expand "atomic_compare_and_swap" [(match_operand:SI 0 "register_operand" "") ;; bool output (match_operand:GPR 1 "register_operand" "") ;; val output @@ -232,26 +218,7 @@ (match_operand:SI 7 "const_int_operand" "")] ;; mod_f "TARGET_ATOMIC" { - emit_insn (gen_atomic_cas_value_strong (operands[1], operands[2], - operands[3], operands[4], - operands[6], operands[7])); - - rtx compare = operands[1]; - if (operands[3] != const0_rtx) - { - rtx difference = gen_rtx_MINUS (mode, operands[1], operands[3]); - compare = gen_reg_rtx (mode); - emit_insn (gen_rtx_SET (compare, difference)); - } - - if (word_mode != mode) - { - rtx reg = gen_reg_rtx (word_mode); - emit_insn (gen_rtx_SET (reg, gen_rtx_SIGN_EXTEND (word_mode, compare))); - compare = reg; - } - - emit_insn (gen_rtx_SET (operands[0], gen_rtx_EQ (SImode, compare, const0_rtx))); + riscv_expand_compare_and_swap (operands); DONE; }) -- 2.31.1