From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-ej1-x62a.google.com (mail-ej1-x62a.google.com [IPv6:2a00:1450:4864:20::62a]) by sourceware.org (Postfix) with ESMTPS id 3595F394EC2A for ; Tue, 2 Mar 2021 12:30:12 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.3.2 sourceware.org 3595F394EC2A Received: by mail-ej1-x62a.google.com with SMTP id mj10so14807846ejb.5 for ; Tue, 02 Mar 2021 04:30:12 -0800 (PST) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20161025; h=x-gm-message-state:mime-version:references:in-reply-to:from:date :message-id:subject:to:cc; bh=v+Y5sp3lUvlQ6gsdHSa8L/2bB1TsfsUI3VScGOUJlA8=; b=AAmez7G2z+G2Kc6zZuqjt96LruQOTkQzppgz5zyh8kMMaTtkQPXgD9RZbS0EG8u9J9 mj0R+DNYPAjJXZ4DhaEzNSZsRgtEVz5Z6CJofWsBkutUBNeeDE1LHEglBTmAOEAzq5Bw kemNzVzJFFum5qGWLN/6HTevuQZHLPO6zIUJFOQ6UjSsNrQFh8LQ5E+3zmbIM8BX+Dub XmWJEygrEopbUaBjKJr0iQa6WJJpl7P6wvtnSjA/Cxz66KKAWupEtCkVdSri717uMfBm nnMQlNxgl4u+IZ/nL+yRa4LmKnhLMzEPvS7tPCZOxw87YSXQEkYNrNG9/Ed1m+EkNpbi FZrA== X-Gm-Message-State: AOAM533SICEt/grqFeEACHc+U/ZxHRMwlip+YhFUl28q/fbyFclQToXG hBItqqG2u81h97L2WHRTQtKWnp9Su8WP9IrJMPI= X-Google-Smtp-Source: ABdhPJwoeyIXVJHDThMl23IjVx/hWXnkGZlJ7YYJ6G3YmYedoJppcDcq7eEak6KUVkz3q4wamNbKmCbegZITBfh/2QA= X-Received: by 2002:a17:906:8474:: with SMTP id hx20mr20814143ejc.129.1614688210732; Tue, 02 Mar 2021 04:30:10 -0800 (PST) MIME-Version: 1.0 References: <20210208124716.487709-1-stefansf@linux.ibm.com> In-Reply-To: From: Richard Biener Date: Tue, 2 Mar 2021 13:29:59 +0100 Message-ID: Subject: Re: [RFC] ldist: Recognize rawmemchr loop patterns To: Stefan Schulze Frielinghaus Cc: GCC Patches Content-Type: text/plain; charset="UTF-8" X-Spam-Status: No, score=-9.1 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, FREEMAIL_FROM, GIT_PATCH_0, RCVD_IN_DNSWL_NONE, 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: Tue, 02 Mar 2021 12:30:15 -0000 On Sun, Feb 14, 2021 at 11:27 AM Stefan Schulze Frielinghaus wrote: > > On Tue, Feb 09, 2021 at 09:57:58AM +0100, Richard Biener wrote: > > On Mon, Feb 8, 2021 at 3:11 PM Stefan Schulze Frielinghaus via > > Gcc-patches wrote: > > > > > > This patch adds support for recognizing loops which mimic the behaviour > > > of function rawmemchr, and replaces those with an internal function call > > > in case a target provides them. In contrast to the original rawmemchr > > > function, this patch also supports different instances where the memory > > > pointed to and the pattern are interpreted as 8, 16, and 32 bit sized, > > > respectively. > > > > > > This patch is not final and I'm looking for some feedback: > > > > > > Previously, only loops which mimic the behaviours of functions memset, > > > memcpy, and memmove have been detected and replaced by corresponding > > > function calls. One characteristic of those loops/partitions is that > > > they don't have a reduction. In contrast, loops which mimic the > > > behaviour of rawmemchr compute a result and therefore have a reduction. > > > My current attempt is to ensure that the reduction statement is not used > > > in any other partition and only in that case ignore the reduction and > > > replace the loop by a function call. We then only need to replace the > > > reduction variable of the loop which contained the loop result by the > > > variable of the lhs of the internal function call. This should ensure > > > that the transformation is correct independently of how partitions are > > > fused/distributed in the end. Any thoughts about this? > > > > Currently we're forcing reduction partitions last (and force to have a single > > one by fusing all partitions containing a reduction) because code-generation > > does not properly update SSA form for the reduction results. ISTR that > > might be just because we do not copy the LC PHI nodes or do not adjust > > them when copying. That might not be an issue in case you replace the > > partition with a call. I guess you can try to have a testcase with > > two rawmemchr patterns and a regular loop part that has to be scheduled > > inbetween both for correctness. > > Ah ok, in that case I updated my patch by removing the constraint that > the reduction statement must be in precisely one partition. Please find > attached the testcases I came up so far. Since transforming a loop into > a rawmemchr function call is backend dependend, I planned to include > those only in my backend patch. I wasn't able to come up with any > testcase where a loop is distributed into multiple partitions and where > one is classified as a rawmemchr builtin. The latter boils down to a > for loop with an empty body only in which case I suspect that loop > distribution shouldn't be done anyway. > > > > Furthermore, I simply added two new members (pattern, fn) to structure > > > builtin_info which I consider rather hacky. For the long run I thought > > > about to split up structure builtin_info into a union where each member > > > is a structure for a particular builtin of a partition, i.e., something > > > like this: > > > > > > union builtin_info > > > { > > > struct binfo_memset *memset; > > > struct binfo_memcpymove *memcpymove; > > > struct binfo_rawmemchr *rawmemchr; > > > }; > > > > > > Such that a structure for one builtin does not get "polluted" by a > > > different one. Any thoughts about this? > > > > Probably makes sense if the list of recognized patterns grow further. > > > > I see you use internal functions rather than builtin functions. I guess > > that's OK. But you use new target hooks for expansion where I think > > new optab entries similar to cmpmem would be more appropriate > > where the distinction between 8, 16 or 32 bits can be encoded in > > the modes. > > The optab implementation is really nice which allows me to use iterators > in the backend which in the end saves me some boiler plate code compared > to the previous implementation :) > > While using optabs now, I only require one additional member (pattern) > in the builtin_info struct. Thus I didn't want to overcomplicate things > and kept the single struct approach as is. > > For the long run, should I resubmit this patch once stage 1 opens or how > would you propose to proceed? Yes, and sorry for the delay. Few comments on the patch given I had a quick look: +void +expand_RAWMEMCHR (internal_fn, gcall *stmt) +{ + expand_operand ops[3]; + + tree lhs = gimple_call_lhs (stmt); I think that give we have people testing with -fno-tree-dce you should try to handle a NULL LHS gracefully. I suppose by simply doing nothing. + tree result_old = build_fold_addr_expr (DR_REF (dr)); + tree result_new = copy_ssa_name (result_old); I think you simply want tree result = make_ssa_name (ptr_type_node); most definitely + imm_use_iterator iter; + gimple *stmt; + use_operand_p use_p; + FOR_EACH_IMM_USE_STMT (stmt, iter, result_old) + { + FOR_EACH_IMM_USE_ON_STMT (use_p, iter) + SET_USE (use_p, result_new); + + update_stmt (stmt); + } isn't going to work. It might work for the very specific case that the IV used for the read is a pointer IV and thus built_fold_addr_expr (...) results in the pointer. But you can't really rely on any such thing. (try a loop operating on a decl - the patch misses testcases at least) Instead you have to record the reduction result during matching. I see you're doing the same there, I think you should try to do sth more general to handle cases like extern char s[]; int foo () { int i; for (i = 1; s[i-1]; ++i) ; return i; } which compute the number of chars including the terminating '\0', because we might want to eventually recognize this as strlen(). Correlating the reduction IV evolution with the rawmemchr result using SCEV shouldn't be difficult and a pointer subtraction would yield the reduction result. If you think that's too complicated for the start then please re-formulate the + tree dr_access_base = build_fold_addr_expr (dr_ref); as if (TREE_CODE (dr_ref) != MEM_REF || !integer_zerop (TREE_OPERAND (dr_ref, 0))) return false; dr_access_base = TREE_OPERAND (dr_ref, 0); + /* A limitation of the current implementation is that we only support + constant patterns. */ + gcond *cond_stmt = as_a (last_stmt (loop->header)); the last stmt in the loop header might not be a gcond. I think you want sth like edge e = single_exit (loop); if (!e) return false; gcond *cond_stmt = safe_dyn_cast (last_stmt (e->src)); if (!cond_stmt) return false; here. + /* Bail out if direction of memory accesses is not growing. */ + if (get_range_pos_neg (iv.step) != 1) + return false; that looks redundant with the access size check. But I'd add a check that the access is of integral, mode-precision type. if (!INTEGRAL_TYPE_P (TREE_TYPE (DR_REF (dr))) || !type_has_mode_precision_p (TREE_TYPE (DR_REF (dr)))) return false; + /* Bail out if target does not provide rawmemchr for a certain mode. */ + machine_mode mode; + switch (TREE_INT_CST_LOW (iv.step)) + { + case 1: mode = QImode; break; + case 2: mode = HImode; break; + case 4: mode = SImode; break; + default: return false; then this simply becomes mode = TYPE_MODE (TREE_TYPE (DR_REF (dr))); For testcases try sth like extern char s[], r[]; char *foo () { int i = 0; char *p = r; while (*p) { ++p; s[i++] = '\0'; } return p; } to end up with multiple partitions (and slightly off IVs because we apply header copying). > Thanks for your review so far! > > Cheers, > Stefan > > > > > Richard. > > > > > Cheers, > > > Stefan > > > --- > > > gcc/internal-fn.c | 42 ++++++ > > > gcc/internal-fn.def | 3 + > > > gcc/target-insns.def | 3 + > > > gcc/tree-loop-distribution.c | 257 ++++++++++++++++++++++++++++++----- > > > 4 files changed, 272 insertions(+), 33 deletions(-) > > > > > > diff --git a/gcc/internal-fn.c b/gcc/internal-fn.c > > > index dd7173126fb..9cd62544a1a 100644 > > > --- a/gcc/internal-fn.c > > > +++ b/gcc/internal-fn.c > > > @@ -2917,6 +2917,48 @@ expand_VEC_CONVERT (internal_fn, gcall *) > > > gcc_unreachable (); > > > } > > > > > > +static void > > > +expand_RAWMEMCHR8 (internal_fn, gcall *stmt) > > > +{ > > > + if (targetm.have_rawmemchr8 ()) > > > + { > > > + rtx result = expand_expr (gimple_call_lhs (stmt), NULL_RTX, VOIDmode, EXPAND_WRITE); > > > + rtx start = expand_normal (gimple_call_arg (stmt, 0)); > > > + rtx pattern = expand_normal (gimple_call_arg (stmt, 1)); > > > + emit_insn (targetm.gen_rawmemchr8 (result, start, pattern)); > > > + } > > > + else > > > + gcc_unreachable(); > > > +} > > > + > > > +static void > > > +expand_RAWMEMCHR16 (internal_fn, gcall *stmt) > > > +{ > > > + if (targetm.have_rawmemchr16 ()) > > > + { > > > + rtx result = expand_expr (gimple_call_lhs (stmt), NULL_RTX, VOIDmode, EXPAND_WRITE); > > > + rtx start = expand_normal (gimple_call_arg (stmt, 0)); > > > + rtx pattern = expand_normal (gimple_call_arg (stmt, 1)); > > > + emit_insn (targetm.gen_rawmemchr16 (result, start, pattern)); > > > + } > > > + else > > > + gcc_unreachable(); > > > +} > > > + > > > +static void > > > +expand_RAWMEMCHR32 (internal_fn, gcall *stmt) > > > +{ > > > + if (targetm.have_rawmemchr32 ()) > > > + { > > > + rtx result = expand_expr (gimple_call_lhs (stmt), NULL_RTX, VOIDmode, EXPAND_WRITE); > > > + rtx start = expand_normal (gimple_call_arg (stmt, 0)); > > > + rtx pattern = expand_normal (gimple_call_arg (stmt, 1)); > > > + emit_insn (targetm.gen_rawmemchr32 (result, start, pattern)); > > > + } > > > + else > > > + gcc_unreachable(); > > > +} > > > + > > > /* Expand the IFN_UNIQUE function according to its first argument. */ > > > > > > static void > > > diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def > > > index daeace7a34e..34247859704 100644 > > > --- a/gcc/internal-fn.def > > > +++ b/gcc/internal-fn.def > > > @@ -348,6 +348,9 @@ DEF_INTERNAL_FN (MUL_OVERFLOW, ECF_CONST | ECF_LEAF | ECF_NOTHROW, NULL) > > > DEF_INTERNAL_FN (TSAN_FUNC_EXIT, ECF_NOVOPS | ECF_LEAF | ECF_NOTHROW, NULL) > > > DEF_INTERNAL_FN (VA_ARG, ECF_NOTHROW | ECF_LEAF, NULL) > > > DEF_INTERNAL_FN (VEC_CONVERT, ECF_CONST | ECF_LEAF | ECF_NOTHROW, NULL) > > > +DEF_INTERNAL_FN (RAWMEMCHR8, ECF_PURE | ECF_LEAF | ECF_NOTHROW, NULL) > > > +DEF_INTERNAL_FN (RAWMEMCHR16, ECF_PURE | ECF_LEAF | ECF_NOTHROW, NULL) > > > +DEF_INTERNAL_FN (RAWMEMCHR32, ECF_PURE | ECF_LEAF | ECF_NOTHROW, NULL) > > > > > > /* An unduplicable, uncombinable function. Generally used to preserve > > > a CFG property in the face of jump threading, tail merging or > > > diff --git a/gcc/target-insns.def b/gcc/target-insns.def > > > index 672c35698d7..9248554cbf3 100644 > > > --- a/gcc/target-insns.def > > > +++ b/gcc/target-insns.def > > > @@ -106,3 +106,6 @@ DEF_TARGET_INSN (trap, (void)) > > > DEF_TARGET_INSN (unique, (void)) > > > DEF_TARGET_INSN (untyped_call, (rtx x0, rtx x1, rtx x2)) > > > DEF_TARGET_INSN (untyped_return, (rtx x0, rtx x1)) > > > +DEF_TARGET_INSN (rawmemchr8, (rtx x0, rtx x1, rtx x2)) > > > +DEF_TARGET_INSN (rawmemchr16, (rtx x0, rtx x1, rtx x2)) > > > +DEF_TARGET_INSN (rawmemchr32, (rtx x0, rtx x1, rtx x2)) > > > diff --git a/gcc/tree-loop-distribution.c b/gcc/tree-loop-distribution.c > > > index 7ee19fc8677..f5b24bf53bc 100644 > > > --- a/gcc/tree-loop-distribution.c > > > +++ b/gcc/tree-loop-distribution.c > > > @@ -218,7 +218,7 @@ enum partition_kind { > > > be unnecessary and removed once distributed memset can be understood > > > and analyzed in data reference analysis. See PR82604 for more. */ > > > PKIND_PARTIAL_MEMSET, > > > - PKIND_MEMSET, PKIND_MEMCPY, PKIND_MEMMOVE > > > + PKIND_MEMSET, PKIND_MEMCPY, PKIND_MEMMOVE, PKIND_RAWMEMCHR > > > }; > > > > > > /* Type of distributed loop. */ > > > @@ -244,6 +244,8 @@ struct builtin_info > > > is only used in memset builtin distribution for now. */ > > > tree dst_base_base; > > > unsigned HOST_WIDE_INT dst_base_offset; > > > + tree pattern; > > > + internal_fn fn; > > > }; > > > > > > /* Partition for loop distribution. */ > > > @@ -588,7 +590,8 @@ class loop_distribution > > > bool > > > classify_partition (loop_p loop, > > > struct graph *rdg, partition *partition, > > > - bitmap stmt_in_all_partitions); > > > + bitmap stmt_in_all_partitions, > > > + vec *partitions); > > > > > > > > > /* Returns true when PARTITION1 and PARTITION2 access the same memory > > > @@ -1232,6 +1235,67 @@ generate_memcpy_builtin (class loop *loop, partition *partition) > > > } > > > } > > > > > > +/* Generate a call to rawmemchr{8,16,32} for PARTITION in LOOP. */ > > > + > > > +static void > > > +generate_rawmemchr_builtin (class loop *loop, partition *partition) > > > +{ > > > + gimple_stmt_iterator gsi; > > > + tree mem, pattern; > > > + struct builtin_info *builtin = partition->builtin; > > > + gimple *fn_call; > > > + > > > + data_reference_p dr = builtin->src_dr; > > > + tree base = builtin->src_base; > > > + > > > + tree result_old = TREE_OPERAND (DR_REF (dr), 0); > > > + tree result_new = copy_ssa_name (result_old); > > > + > > > + /* The new statements will be placed before LOOP. */ > > > + gsi = gsi_last_bb (loop_preheader_edge (loop)->src); > > > + > > > + mem = force_gimple_operand_gsi (&gsi, base, true, NULL_TREE, false, GSI_CONTINUE_LINKING); > > > + pattern = builtin->pattern; > > > + if (TREE_CODE (pattern) == INTEGER_CST) > > > + pattern = fold_convert (integer_type_node, pattern); > > > + fn_call = gimple_build_call_internal (builtin->fn, 2, mem, pattern); > > > + gimple_call_set_lhs (fn_call, result_new); > > > + gimple_set_location (fn_call, partition->loc); > > > + gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); > > > + > > > + imm_use_iterator iter; > > > + gimple *stmt; > > > + use_operand_p use_p; > > > + FOR_EACH_IMM_USE_STMT (stmt, iter, result_old) > > > + { > > > + FOR_EACH_IMM_USE_ON_STMT (use_p, iter) > > > + SET_USE (use_p, result_new); > > > + > > > + update_stmt (stmt); > > > + } > > > + > > > + fold_stmt (&gsi); > > > + > > > + if (dump_file && (dump_flags & TDF_DETAILS)) > > > + switch (builtin->fn) > > > + { > > > + case IFN_RAWMEMCHR8: > > > + fprintf (dump_file, "generated rawmemchr8\n"); > > > + break; > > > + > > > + case IFN_RAWMEMCHR16: > > > + fprintf (dump_file, "generated rawmemchr16\n"); > > > + break; > > > + > > > + case IFN_RAWMEMCHR32: > > > + fprintf (dump_file, "generated rawmemchr32\n"); > > > + break; > > > + > > > + default: > > > + gcc_unreachable (); > > > + } > > > +} > > > + > > > /* Remove and destroy the loop LOOP. */ > > > > > > static void > > > @@ -1334,6 +1398,10 @@ generate_code_for_partition (class loop *loop, > > > generate_memcpy_builtin (loop, partition); > > > break; > > > > > > + case PKIND_RAWMEMCHR: > > > + generate_rawmemchr_builtin (loop, partition); > > > + break; > > > + > > > default: > > > gcc_unreachable (); > > > } > > > @@ -1525,44 +1593,53 @@ find_single_drs (class loop *loop, struct graph *rdg, partition *partition, > > > } > > > } > > > > > > - if (!single_st) > > > + if (!single_ld && !single_st) > > > return false; > > > > > > - /* Bail out if this is a bitfield memory reference. */ > > > - if (TREE_CODE (DR_REF (single_st)) == COMPONENT_REF > > > - && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_st), 1))) > > > - return false; > > > - > > > - /* Data reference must be executed exactly once per iteration of each > > > - loop in the loop nest. We only need to check dominance information > > > - against the outermost one in a perfect loop nest because a bb can't > > > - dominate outermost loop's latch without dominating inner loop's. */ > > > - basic_block bb_st = gimple_bb (DR_STMT (single_st)); > > > - if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_st)) > > > - return false; > > > + basic_block bb_ld = NULL; > > > + basic_block bb_st = NULL; > > > > > > if (single_ld) > > > { > > > - gimple *store = DR_STMT (single_st), *load = DR_STMT (single_ld); > > > - /* Direct aggregate copy or via an SSA name temporary. */ > > > - if (load != store > > > - && gimple_assign_lhs (load) != gimple_assign_rhs1 (store)) > > > - return false; > > > - > > > /* Bail out if this is a bitfield memory reference. */ > > > if (TREE_CODE (DR_REF (single_ld)) == COMPONENT_REF > > > && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_ld), 1))) > > > return false; > > > > > > - /* Load and store must be in the same loop nest. */ > > > - basic_block bb_ld = gimple_bb (DR_STMT (single_ld)); > > > - if (bb_st->loop_father != bb_ld->loop_father) > > > + /* Data reference must be executed exactly once per iteration of each > > > + loop in the loop nest. We only need to check dominance information > > > + against the outermost one in a perfect loop nest because a bb can't > > > + dominate outermost loop's latch without dominating inner loop's. */ > > > + bb_ld = gimple_bb (DR_STMT (single_ld)); > > > + if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_ld)) > > > + return false; > > > + } > > > + > > > + if (single_st) > > > + { > > > + /* Bail out if this is a bitfield memory reference. */ > > > + if (TREE_CODE (DR_REF (single_st)) == COMPONENT_REF > > > + && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_st), 1))) > > > return false; > > > > > > /* Data reference must be executed exactly once per iteration. > > > - Same as single_st, we only need to check against the outermost > > > + Same as single_ld, we only need to check against the outermost > > > loop. */ > > > - if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_ld)) > > > + bb_st = gimple_bb (DR_STMT (single_st)); > > > + if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_st)) > > > + return false; > > > + } > > > + > > > + if (single_ld && single_st) > > > + { > > > + gimple *store = DR_STMT (single_st), *load = DR_STMT (single_ld); > > > + /* Direct aggregate copy or via an SSA name temporary. */ > > > + if (load != store > > > + && gimple_assign_lhs (load) != gimple_assign_rhs1 (store)) > > > + return false; > > > + > > > + /* Load and store must be in the same loop nest. */ > > > + if (bb_st->loop_father != bb_ld->loop_father) > > > return false; > > > > > > edge e = single_exit (bb_st->loop_father); > > > @@ -1681,6 +1758,84 @@ alloc_builtin (data_reference_p dst_dr, data_reference_p src_dr, > > > return builtin; > > > } > > > > > > +/* Given data reference DR in loop nest LOOP, classify if it forms builtin > > > + rawmemchr{8,16,32} call. */ > > > + > > > +static bool > > > +classify_builtin_rawmemchr (loop_p loop, partition *partition, data_reference_p dr, tree loop_result) > > > +{ > > > + tree dr_ref = DR_REF (dr); > > > + tree dr_access_base = build_fold_addr_expr (dr_ref); > > > + tree dr_access_size = TYPE_SIZE_UNIT (TREE_TYPE (dr_ref)); > > > + gimple *dr_stmt = DR_STMT (dr); > > > + tree rhs1 = gimple_assign_rhs1 (dr_stmt); > > > + affine_iv iv; > > > + tree pattern; > > > + > > > + if (TREE_OPERAND (rhs1, 0) != loop_result) > > > + return false; > > > + > > > + /* A limitation of the current implementation is that we only support > > > + constant patterns. */ > > > + gcond *cond_stmt = as_a (last_stmt (loop->header)); > > > + pattern = gimple_cond_rhs (cond_stmt); > > > + if (gimple_cond_code (cond_stmt) != NE_EXPR > > > + || gimple_cond_lhs (cond_stmt) != gimple_assign_lhs (dr_stmt) > > > + || TREE_CODE (pattern) != INTEGER_CST) > > > + return false; > > > + > > > + /* Bail out if no affine induction variable with constant step can be > > > + determined. */ > > > + if (!simple_iv (loop, loop, dr_access_base, &iv, false)) > > > + return false; > > > + > > > + /* Bail out if memory accesses are not consecutive. */ > > > + if (!operand_equal_p (iv.step, dr_access_size, 0)) > > > + return false; > > > + > > > + /* Bail out if direction of memory accesses is not growing. */ > > > + if (get_range_pos_neg (iv.step) != 1) > > > + return false; > > > + > > > + internal_fn fn; > > > + switch (TREE_INT_CST_LOW (iv.step)) > > > + { > > > + case 1: > > > + if (!targetm.have_rawmemchr8 ()) > > > + return false; > > > + fn = IFN_RAWMEMCHR8; > > > + break; > > > + > > > + case 2: > > > + if (!targetm.have_rawmemchr16 ()) > > > + return false; > > > + fn = IFN_RAWMEMCHR16; > > > + break; > > > + > > > + case 4: > > > + if (!targetm.have_rawmemchr32 ()) > > > + return false; > > > + fn = IFN_RAWMEMCHR32; > > > + break; > > > + > > > + default: > > > + return false; > > > + } > > > + > > > + struct builtin_info *builtin; > > > + builtin = alloc_builtin (NULL, NULL, NULL_TREE, NULL_TREE, NULL_TREE); > > > + builtin->src_dr = dr; > > > + builtin->src_base = iv.base; > > > + builtin->pattern = pattern; > > > + builtin->fn = fn; > > > + > > > + partition->loc = gimple_location (dr_stmt); > > > + partition->builtin = builtin; > > > + partition->kind = PKIND_RAWMEMCHR; > > > + > > > + return true; > > > +} > > > + > > > /* Given data reference DR in loop nest LOOP, classify if it forms builtin > > > memset call. */ > > > > > > @@ -1792,12 +1947,16 @@ loop_distribution::classify_builtin_ldst (loop_p loop, struct graph *rdg, > > > bool > > > loop_distribution::classify_partition (loop_p loop, > > > struct graph *rdg, partition *partition, > > > - bitmap stmt_in_all_partitions) > > > + bitmap stmt_in_all_partitions, > > > + vec *partitions) > > > { > > > bitmap_iterator bi; > > > unsigned i; > > > data_reference_p single_ld = NULL, single_st = NULL; > > > bool volatiles_p = false, has_reduction = false; > > > + unsigned nreductions = 0; > > > + gimple *reduction_stmt = NULL; > > > + bool has_interpar_reduction = false; > > > > > > EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) > > > { > > > @@ -1821,6 +1980,19 @@ loop_distribution::classify_partition (loop_p loop, > > > partition->reduction_p = true; > > > else > > > has_reduction = true; > > > + > > > + /* Determine whether the reduction statement occurs in other > > > + partitions than the current one. */ > > > + struct partition *piter; > > > + for (unsigned j = 0; partitions->iterate (j, &piter); ++j) > > > + { > > > + if (piter == partition) > > > + continue; > > > + if (bitmap_bit_p (piter->stmts, i)) > > > + has_interpar_reduction = true; > > > + } > > > + reduction_stmt = stmt; > > > + ++nreductions; > > > } > > > } > > > > > > @@ -1840,6 +2012,30 @@ loop_distribution::classify_partition (loop_p loop, > > > if (!find_single_drs (loop, rdg, partition, &single_st, &single_ld)) > > > return has_reduction; > > > > > > + /* If we determined a single load and a single reduction statement which does > > > + not occur in any other partition, then try to classify this partition as a > > > + rawmemchr builtin. */ > > > + if (single_ld != NULL > > > + && single_st == NULL > > > + && nreductions == 1 > > > + && !has_interpar_reduction > > > + && is_gimple_assign (reduction_stmt)) > > > + { > > > + /* If we classified the partition as a builtin, then ignoring the single > > > + reduction is safe, since the reduction variable is not used in other > > > + partitions. */ > > > + tree reduction_var = gimple_assign_lhs (reduction_stmt); > > > + return !classify_builtin_rawmemchr (loop, partition, single_ld, reduction_var); > > > + } > > > + > > > + if (single_st == NULL) > > > + return has_reduction; > > > + > > > + /* Don't distribute loop if niters is unknown. */ > > > + tree niters = number_of_latch_executions (loop); > > > + if (niters == NULL_TREE || niters == chrec_dont_know) > > > + return has_reduction; > > > + > > > partition->loc = gimple_location (DR_STMT (single_st)); > > > > > > /* Classify the builtin kind. */ > > > @@ -2979,7 +3175,7 @@ loop_distribution::distribute_loop (class loop *loop, vec stmts, > > > FOR_EACH_VEC_ELT (partitions, i, partition) > > > { > > > reduction_in_all > > > - |= classify_partition (loop, rdg, partition, stmt_in_all_partitions); > > > + |= classify_partition (loop, rdg, partition, stmt_in_all_partitions, &partitions); > > > any_builtin |= partition_builtin_p (partition); > > > } > > > > > > @@ -3290,11 +3486,6 @@ loop_distribution::execute (function *fun) > > > && !optimize_loop_for_speed_p (loop))) > > > continue; > > > > > > - /* Don't distribute loop if niters is unknown. */ > > > - tree niters = number_of_latch_executions (loop); > > > - if (niters == NULL_TREE || niters == chrec_dont_know) > > > - continue; > > > - > > > /* Get the perfect loop nest for distribution. */ > > > loop = prepare_perfect_loop_nest (loop); > > > for (; loop; loop = loop->inner) > > > -- > > > 2.23.0 > > >