From patchwork Sat Nov 7 13:33:45 2015 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Richard Sandiford X-Patchwork-Id: 541302 Return-Path: X-Original-To: incoming@patchwork.ozlabs.org Delivered-To: patchwork-incoming@bilbo.ozlabs.org Received: from sourceware.org (server1.sourceware.org [209.132.180.131]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by ozlabs.org (Postfix) with ESMTPS id 94855140322 for ; Sun, 8 Nov 2015 00:34:04 +1100 (AEDT) Authentication-Results: ozlabs.org; dkim=pass (1024-bit key; unprotected) header.d=gcc.gnu.org header.i=@gcc.gnu.org header.b=cRtTXpiQ; dkim-atps=neutral DomainKey-Signature: a=rsa-sha1; c=nofws; d=gcc.gnu.org; h=list-id :list-unsubscribe:list-archive:list-post:list-help:sender:from :to:subject:date:message-id:mime-version:content-type :content-transfer-encoding; q=dns; s=default; b=rbGg2RlmacRaUvh9 o8zkp3fYD0hf2RG/rIqvhTZ0IfsHb3tvpOlnWJx80NjmVQWPw1egvkwwXdDwYL8+ B3w88IMBIRaHrdlMs4KhKW1yFNw+weshNrZhVq75dm2YbnPar32Dk2ucGLlQkNu/ kWNqrSsrxTfDYKlnKUdhaq5Y+7I= DKIM-Signature: v=1; a=rsa-sha1; c=relaxed; d=gcc.gnu.org; h=list-id :list-unsubscribe:list-archive:list-post:list-help:sender:from :to:subject:date:message-id:mime-version:content-type :content-transfer-encoding; s=default; bh=mntYpvPSapsfxUtT2Ak1ug FbH9A=; b=cRtTXpiQQfORC7tv+BJ69XTOaZl9IlDB6bX0o0llUsZZPRQb9HouKj BkVhsutj0alrgpj22unm9wA1Asy7k0MWWVgIj/Ccn+V7lO88QipI2vHdMIsi6OQ1 DTZWxy2t4tnOsk8nKurK6Jy39X1CzSq0Ltckhq694RAuhdw0H3IRc= Received: (qmail 130970 invoked by alias); 7 Nov 2015 13:33:57 -0000 Mailing-List: contact gcc-patches-help@gcc.gnu.org; run by ezmlm Precedence: bulk List-Id: List-Unsubscribe: List-Archive: List-Post: List-Help: Sender: gcc-patches-owner@gcc.gnu.org Delivered-To: mailing list gcc-patches@gcc.gnu.org Received: (qmail 130953 invoked by uid 89); 7 Nov 2015 13:33:55 -0000 Authentication-Results: sourceware.org; auth=none X-Virus-Found: No X-Spam-SWARE-Status: No, score=-1.5 required=5.0 tests=AWL, BAYES_00, SPF_PASS autolearn=ham version=3.3.2 X-HELO: eu-smtp-delivery-143.mimecast.com Received: from eu-smtp-delivery-143.mimecast.com (HELO eu-smtp-delivery-143.mimecast.com) (207.82.80.143) by sourceware.org (qpsmtpd/0.93/v0.84-503-g423c35a) with ESMTP; Sat, 07 Nov 2015 13:33:51 +0000 Received: from cam-owa1.Emea.Arm.com (fw-tnat.cambridge.arm.com [217.140.96.140]) by eu-smtp-1.mimecast.com with ESMTP id uk-mta-22-Jo7qE2fTSw2QKcxfXABufQ-1; Sat, 07 Nov 2015 13:33:46 +0000 Received: from localhost ([10.1.2.79]) by cam-owa1.Emea.Arm.com with Microsoft SMTPSVC(6.0.3790.3959); Sat, 7 Nov 2015 13:33:45 +0000 From: Richard Sandiford To: gcc-patches@gcc.gnu.org Mail-Followup-To: gcc-patches@gcc.gnu.org, richard.sandiford@arm.com Subject: Extend tree-call-cdce to calls whose result is used Date: Sat, 07 Nov 2015 13:33:45 +0000 Message-ID: <874mgyorwm.fsf@e105548-lin.cambridge.arm.com> User-Agent: Gnus/5.130012 (Ma Gnus v0.12) Emacs/24.3 (gnu/linux) MIME-Version: 1.0 X-MC-Unique: Jo7qE2fTSw2QKcxfXABufQ-1 For -fmath-errno, builtins.c currently expands calls to sqrt to: y = sqrt_optab (x); if (y != y) [ sqrt (x); or errno = EDOM; ] The drawbacks of this are: - the call to sqrt is protected by the result of the optab rather than the input. It would be better to check !(x >= 0), like tree-call-cdce.c does. - the branch isn't exposed at the gimple level and so gets little high-level optimisation. - we do this for log too, but for log a zero input produces -inf rather than a NaN, and sets errno to ERANGE rather than EDOM. This patch moves the code to tree-call-cdce.c instead, with the optab operation being represented as an internal function. This means that we can use the existing argument-based range checks rather than the result-based checks and that we get more gimple optimisation of the branch. Previously the pass was only enabled by default at -O2 or above, but the old builtins.c code was enabled at -O. The patch therefore enables the pass at -O as well. The previous patch to cfgexpand.c handled cases where functions don't (or are assumed not to) set errno, so this patch makes the builtins.c code dead. Tested on x86_64-linux-gnu, aarch64-linux-gnu, arm-linux-gnueabi and visium-elf (for the EDOM stuff). OK to install? Thanks, Richard gcc/ * builtins.c (expand_errno_check, expand_builtin_mathfn) (expand_builtin_mathfn_2): Delete. (expand_builtin): Remove handling of functions with internal function equivalents. * internal-fn.def (SET_EDOM): New internal function. * internal-fn.h (set_edom_supported_p): Declare. * internal-fn.c (expand_SET_EDOM): New function. (set_edom_supported_p): Likewise. * tree-call-cdce.c: Include builtins.h and internal-fn.h. Rewrite comment at head of file. (is_call_dce_candidate): Rename to... (can_test_argument_range): ...this. Don't check gimple_call_lhs or gimple_call_builtin_p here. (edom_only_function): New function. (shrink_wrap_one_built_in_call_with_conds): New function, split out from... (shrink_wrap_one_built_in_call): ...here. (can_use_internal_fn, use_internal_fn): New functions. (shrink_wrap_conditional_dead_built_in_calls): Call use_internal_fn for calls that have an lhs. (pass_call_cdce::gate): Remove optimize_function_for_speed_p check. (pass_call_cdce::execute): Skip blocks that are optimized for size. Check gimple_call_builtin_p here. Use can_use_internal_fn for calls with an lhs. * opts.c (default_options_table): Enable -ftree-builtin-call-cdce at -O and above. diff --git a/gcc/builtins.c b/gcc/builtins.c index bbcc7dc3..1c13a51 100644 --- a/gcc/builtins.c +++ b/gcc/builtins.c @@ -101,9 +101,6 @@ static rtx expand_builtin_apply (rtx, rtx, rtx); static void expand_builtin_return (rtx); static enum type_class type_to_class (tree); static rtx expand_builtin_classify_type (tree); -static void expand_errno_check (tree, rtx); -static rtx expand_builtin_mathfn (tree, rtx, rtx); -static rtx expand_builtin_mathfn_2 (tree, rtx, rtx); static rtx expand_builtin_mathfn_3 (tree, rtx, rtx); static rtx expand_builtin_mathfn_ternary (tree, rtx, rtx); static rtx expand_builtin_interclass_mathfn (tree, rtx); @@ -1972,286 +1969,6 @@ replacement_internal_fn (gcall *call) return IFN_LAST; } -/* If errno must be maintained, expand the RTL to check if the result, - TARGET, of a built-in function call, EXP, is NaN, and if so set - errno to EDOM. */ - -static void -expand_errno_check (tree exp, rtx target) -{ - rtx_code_label *lab = gen_label_rtx (); - - /* Test the result; if it is NaN, set errno=EDOM because - the argument was not in the domain. */ - do_compare_rtx_and_jump (target, target, EQ, 0, GET_MODE (target), - NULL_RTX, NULL, lab, - /* The jump is very likely. */ - REG_BR_PROB_BASE - (REG_BR_PROB_BASE / 2000 - 1)); - -#ifdef TARGET_EDOM - /* If this built-in doesn't throw an exception, set errno directly. */ - if (TREE_NOTHROW (TREE_OPERAND (CALL_EXPR_FN (exp), 0))) - { -#ifdef GEN_ERRNO_RTX - rtx errno_rtx = GEN_ERRNO_RTX; -#else - rtx errno_rtx - = gen_rtx_MEM (word_mode, gen_rtx_SYMBOL_REF (Pmode, "errno")); -#endif - emit_move_insn (errno_rtx, - gen_int_mode (TARGET_EDOM, GET_MODE (errno_rtx))); - emit_label (lab); - return; - } -#endif - - /* Make sure the library call isn't expanded as a tail call. */ - CALL_EXPR_TAILCALL (exp) = 0; - - /* We can't set errno=EDOM directly; let the library call do it. - Pop the arguments right away in case the call gets deleted. */ - NO_DEFER_POP; - expand_call (exp, target, 0); - OK_DEFER_POP; - emit_label (lab); -} - -/* Expand a call to one of the builtin math functions (sqrt, exp, or log). - Return NULL_RTX if a normal call should be emitted rather than expanding - the function in-line. EXP is the expression that is a call to the builtin - function; if convenient, the result should be placed in TARGET. - SUBTARGET may be used as the target for computing one of EXP's operands. */ - -static rtx -expand_builtin_mathfn (tree exp, rtx target, rtx subtarget) -{ - optab builtin_optab; - rtx op0; - rtx_insn *insns; - tree fndecl = get_callee_fndecl (exp); - machine_mode mode; - bool errno_set = false; - bool try_widening = false; - tree arg; - - if (!validate_arglist (exp, REAL_TYPE, VOID_TYPE)) - return NULL_RTX; - - arg = CALL_EXPR_ARG (exp, 0); - - switch (DECL_FUNCTION_CODE (fndecl)) - { - CASE_FLT_FN (BUILT_IN_SQRT): - errno_set = ! tree_expr_nonnegative_p (arg); - try_widening = true; - builtin_optab = sqrt_optab; - break; - CASE_FLT_FN (BUILT_IN_EXP): - errno_set = true; builtin_optab = exp_optab; break; - CASE_FLT_FN (BUILT_IN_EXP10): - CASE_FLT_FN (BUILT_IN_POW10): - errno_set = true; builtin_optab = exp10_optab; break; - CASE_FLT_FN (BUILT_IN_EXP2): - errno_set = true; builtin_optab = exp2_optab; break; - CASE_FLT_FN (BUILT_IN_EXPM1): - errno_set = true; builtin_optab = expm1_optab; break; - CASE_FLT_FN (BUILT_IN_LOGB): - errno_set = true; builtin_optab = logb_optab; break; - CASE_FLT_FN (BUILT_IN_LOG): - errno_set = true; builtin_optab = log_optab; break; - CASE_FLT_FN (BUILT_IN_LOG10): - errno_set = true; builtin_optab = log10_optab; break; - CASE_FLT_FN (BUILT_IN_LOG2): - errno_set = true; builtin_optab = log2_optab; break; - CASE_FLT_FN (BUILT_IN_LOG1P): - errno_set = true; builtin_optab = log1p_optab; break; - CASE_FLT_FN (BUILT_IN_ASIN): - builtin_optab = asin_optab; break; - CASE_FLT_FN (BUILT_IN_ACOS): - builtin_optab = acos_optab; break; - CASE_FLT_FN (BUILT_IN_TAN): - builtin_optab = tan_optab; break; - CASE_FLT_FN (BUILT_IN_ATAN): - builtin_optab = atan_optab; break; - CASE_FLT_FN (BUILT_IN_FLOOR): - builtin_optab = floor_optab; break; - CASE_FLT_FN (BUILT_IN_CEIL): - builtin_optab = ceil_optab; break; - CASE_FLT_FN (BUILT_IN_TRUNC): - builtin_optab = btrunc_optab; break; - CASE_FLT_FN (BUILT_IN_ROUND): - builtin_optab = round_optab; break; - CASE_FLT_FN (BUILT_IN_NEARBYINT): - builtin_optab = nearbyint_optab; - if (flag_trapping_math) - break; - /* Else fallthrough and expand as rint. */ - CASE_FLT_FN (BUILT_IN_RINT): - builtin_optab = rint_optab; break; - CASE_FLT_FN (BUILT_IN_SIGNIFICAND): - builtin_optab = significand_optab; break; - default: - gcc_unreachable (); - } - - /* Make a suitable register to place result in. */ - mode = TYPE_MODE (TREE_TYPE (exp)); - - if (! flag_errno_math || ! HONOR_NANS (mode)) - errno_set = false; - - /* Before working hard, check whether the instruction is available, but try - to widen the mode for specific operations. */ - if ((optab_handler (builtin_optab, mode) != CODE_FOR_nothing - || (try_widening && !excess_precision_type (TREE_TYPE (exp)))) - && (!errno_set || !optimize_insn_for_size_p ())) - { - rtx result = gen_reg_rtx (mode); - - /* Wrap the computation of the argument in a SAVE_EXPR, as we may - need to expand the argument again. This way, we will not perform - side-effects more the once. */ - CALL_EXPR_ARG (exp, 0) = arg = builtin_save_expr (arg); - - op0 = expand_expr (arg, subtarget, VOIDmode, EXPAND_NORMAL); - - start_sequence (); - - /* Compute into RESULT. - Set RESULT to wherever the result comes back. */ - result = expand_unop (mode, builtin_optab, op0, result, 0); - - if (result != 0) - { - if (errno_set) - expand_errno_check (exp, result); - - /* Output the entire sequence. */ - insns = get_insns (); - end_sequence (); - emit_insn (insns); - return result; - } - - /* If we were unable to expand via the builtin, stop the sequence - (without outputting the insns) and call to the library function - with the stabilized argument list. */ - end_sequence (); - } - - return expand_call (exp, target, target == const0_rtx); -} - -/* Expand a call to the builtin binary math functions (pow and atan2). - Return NULL_RTX if a normal call should be emitted rather than expanding the - function in-line. EXP is the expression that is a call to the builtin - function; if convenient, the result should be placed in TARGET. - SUBTARGET may be used as the target for computing one of EXP's - operands. */ - -static rtx -expand_builtin_mathfn_2 (tree exp, rtx target, rtx subtarget) -{ - optab builtin_optab; - rtx op0, op1, result; - rtx_insn *insns; - int op1_type = REAL_TYPE; - tree fndecl = get_callee_fndecl (exp); - tree arg0, arg1; - machine_mode mode; - bool errno_set = true; - - switch (DECL_FUNCTION_CODE (fndecl)) - { - CASE_FLT_FN (BUILT_IN_SCALBN): - CASE_FLT_FN (BUILT_IN_SCALBLN): - CASE_FLT_FN (BUILT_IN_LDEXP): - op1_type = INTEGER_TYPE; - default: - break; - } - - if (!validate_arglist (exp, REAL_TYPE, op1_type, VOID_TYPE)) - return NULL_RTX; - - arg0 = CALL_EXPR_ARG (exp, 0); - arg1 = CALL_EXPR_ARG (exp, 1); - - switch (DECL_FUNCTION_CODE (fndecl)) - { - CASE_FLT_FN (BUILT_IN_POW): - builtin_optab = pow_optab; break; - CASE_FLT_FN (BUILT_IN_ATAN2): - builtin_optab = atan2_optab; break; - CASE_FLT_FN (BUILT_IN_SCALB): - if (REAL_MODE_FORMAT (TYPE_MODE (TREE_TYPE (exp)))->b != 2) - return 0; - builtin_optab = scalb_optab; break; - CASE_FLT_FN (BUILT_IN_SCALBN): - CASE_FLT_FN (BUILT_IN_SCALBLN): - if (REAL_MODE_FORMAT (TYPE_MODE (TREE_TYPE (exp)))->b != 2) - return 0; - /* Fall through... */ - CASE_FLT_FN (BUILT_IN_LDEXP): - builtin_optab = ldexp_optab; break; - CASE_FLT_FN (BUILT_IN_FMOD): - builtin_optab = fmod_optab; break; - CASE_FLT_FN (BUILT_IN_REMAINDER): - CASE_FLT_FN (BUILT_IN_DREM): - builtin_optab = remainder_optab; break; - default: - gcc_unreachable (); - } - - /* Make a suitable register to place result in. */ - mode = TYPE_MODE (TREE_TYPE (exp)); - - /* Before working hard, check whether the instruction is available. */ - if (optab_handler (builtin_optab, mode) == CODE_FOR_nothing) - return NULL_RTX; - - result = gen_reg_rtx (mode); - - if (! flag_errno_math || ! HONOR_NANS (mode)) - errno_set = false; - - if (errno_set && optimize_insn_for_size_p ()) - return 0; - - /* Always stabilize the argument list. */ - CALL_EXPR_ARG (exp, 0) = arg0 = builtin_save_expr (arg0); - CALL_EXPR_ARG (exp, 1) = arg1 = builtin_save_expr (arg1); - - op0 = expand_expr (arg0, subtarget, VOIDmode, EXPAND_NORMAL); - op1 = expand_normal (arg1); - - start_sequence (); - - /* Compute into RESULT. - Set RESULT to wherever the result comes back. */ - result = expand_binop (mode, builtin_optab, op0, op1, - result, 0, OPTAB_DIRECT); - - /* If we were unable to expand via the builtin, stop the sequence - (without outputting the insns) and call to the library function - with the stabilized argument list. */ - if (result == 0) - { - end_sequence (); - return expand_call (exp, target, target == const0_rtx); - } - - if (errno_set) - expand_errno_check (exp, result); - - /* Output the entire sequence. */ - insns = get_insns (); - end_sequence (); - emit_insn (insns); - - return result; -} - /* Expand a call to the builtin trinary math functions (fma). Return NULL_RTX if a normal call should be emitted rather than expanding the function in-line. EXP is the expression that is a call to the builtin @@ -5984,37 +5701,6 @@ expand_builtin (tree exp, rtx target, rtx subtarget, machine_mode mode, CASE_FLT_FN (BUILT_IN_CABS): break; - CASE_FLT_FN (BUILT_IN_EXP): - CASE_FLT_FN (BUILT_IN_EXP10): - CASE_FLT_FN (BUILT_IN_POW10): - CASE_FLT_FN (BUILT_IN_EXP2): - CASE_FLT_FN (BUILT_IN_EXPM1): - CASE_FLT_FN (BUILT_IN_LOGB): - CASE_FLT_FN (BUILT_IN_LOG): - CASE_FLT_FN (BUILT_IN_LOG10): - CASE_FLT_FN (BUILT_IN_LOG2): - CASE_FLT_FN (BUILT_IN_LOG1P): - CASE_FLT_FN (BUILT_IN_TAN): - CASE_FLT_FN (BUILT_IN_ASIN): - CASE_FLT_FN (BUILT_IN_ACOS): - CASE_FLT_FN (BUILT_IN_ATAN): - CASE_FLT_FN (BUILT_IN_SIGNIFICAND): - /* Treat these like sqrt only if unsafe math optimizations are allowed, - because of possible accuracy problems. */ - if (! flag_unsafe_math_optimizations) - break; - CASE_FLT_FN (BUILT_IN_SQRT): - CASE_FLT_FN (BUILT_IN_FLOOR): - CASE_FLT_FN (BUILT_IN_CEIL): - CASE_FLT_FN (BUILT_IN_TRUNC): - CASE_FLT_FN (BUILT_IN_ROUND): - CASE_FLT_FN (BUILT_IN_NEARBYINT): - CASE_FLT_FN (BUILT_IN_RINT): - target = expand_builtin_mathfn (exp, target, subtarget); - if (target) - return target; - break; - CASE_FLT_FN (BUILT_IN_FMA): target = expand_builtin_mathfn_ternary (exp, target, subtarget); if (target) @@ -6061,23 +5747,6 @@ expand_builtin (tree exp, rtx target, rtx subtarget, machine_mode mode, return target; break; - CASE_FLT_FN (BUILT_IN_ATAN2): - CASE_FLT_FN (BUILT_IN_LDEXP): - CASE_FLT_FN (BUILT_IN_SCALB): - CASE_FLT_FN (BUILT_IN_SCALBN): - CASE_FLT_FN (BUILT_IN_SCALBLN): - if (! flag_unsafe_math_optimizations) - break; - - CASE_FLT_FN (BUILT_IN_FMOD): - CASE_FLT_FN (BUILT_IN_REMAINDER): - CASE_FLT_FN (BUILT_IN_DREM): - CASE_FLT_FN (BUILT_IN_POW): - target = expand_builtin_mathfn_2 (exp, target, subtarget); - if (target) - return target; - break; - CASE_FLT_FN (BUILT_IN_CEXPI): target = expand_builtin_cexpi (exp, target); gcc_assert (target); diff --git a/gcc/internal-fn.c b/gcc/internal-fn.c index c03c0fc..898c83d 100644 --- a/gcc/internal-fn.c +++ b/gcc/internal-fn.c @@ -2069,6 +2069,24 @@ expand_GOACC_REDUCTION (internal_fn, gcall *) gcc_unreachable (); } +/* Set errno to EDOM. */ + +static void +expand_SET_EDOM (internal_fn, gcall *) +{ +#ifdef TARGET_EDOM +#ifdef GEN_ERRNO_RTX + rtx errno_rtx = GEN_ERRNO_RTX; +#else + rtx errno_rtx = gen_rtx_MEM (word_mode, gen_rtx_SYMBOL_REF (Pmode, "errno")); +#endif + emit_move_insn (errno_rtx, + gen_int_mode (TARGET_EDOM, GET_MODE (errno_rtx))); +#else + gcc_unreachable (); +#endif +} + /* Expand a call to FN using the operands in STMT. FN has a single output operand and NARGS input operands. */ @@ -2213,6 +2231,18 @@ direct_internal_fn_supported_p (internal_fn fn, tree type) return direct_internal_fn_supported_p (fn, tree_pair (type, type)); } +/* Return true if IFN_SET_EDOM is supported. */ + +bool +set_edom_supported_p (void) +{ +#ifdef TARGET_EDOM + return true; +#else + return false; +#endif +} + #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ static void \ expand_##CODE (internal_fn fn, gcall *stmt) \ diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def index bf8047a..825dba1 100644 --- a/gcc/internal-fn.def +++ b/gcc/internal-fn.def @@ -181,6 +181,10 @@ DEF_INTERNAL_FN (GOACC_LOOP, ECF_PURE | ECF_NOTHROW, NULL) /* OpenACC reduction abstraction. See internal-fn.h for usage. */ DEF_INTERNAL_FN (GOACC_REDUCTION, ECF_NOTHROW | ECF_LEAF, NULL) +/* Set errno to EDOM, if GCC knows how to do that directly for the + current target. */ +DEF_INTERNAL_FN (SET_EDOM, ECF_LEAF | ECF_NOTHROW, NULL) + #undef DEF_INTERNAL_INT_FN #undef DEF_INTERNAL_FLT_FN #undef DEF_INTERNAL_OPTAB_FN diff --git a/gcc/internal-fn.h b/gcc/internal-fn.h index 5ee43b8..6cb123f 100644 --- a/gcc/internal-fn.h +++ b/gcc/internal-fn.h @@ -160,6 +160,7 @@ extern tree_pair direct_internal_fn_types (internal_fn, tree, tree *); extern tree_pair direct_internal_fn_types (internal_fn, gcall *); extern bool direct_internal_fn_supported_p (internal_fn, tree_pair); extern bool direct_internal_fn_supported_p (internal_fn, tree); +extern bool set_edom_supported_p (void); extern void expand_internal_call (gcall *); extern void expand_internal_call (internal_fn, gcall *); diff --git a/gcc/opts.c b/gcc/opts.c index 9a3fbb3..99d35ae 100644 --- a/gcc/opts.c +++ b/gcc/opts.c @@ -464,6 +464,7 @@ static const struct default_options default_options_table[] = { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_fmove_loop_invariants, NULL, 1 }, { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_ftree_pta, NULL, 1 }, { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_fssa_phiopt, NULL, 1 }, + { OPT_LEVELS_1_PLUS, OPT_ftree_builtin_call_dce, NULL, 1 }, /* -O2 optimizations. */ { OPT_LEVELS_2_PLUS, OPT_finline_small_functions, NULL, 1 }, @@ -489,7 +490,6 @@ static const struct default_options default_options_table[] = REORDER_BLOCKS_ALGORITHM_STC }, { OPT_LEVELS_2_PLUS, OPT_freorder_functions, NULL, 1 }, { OPT_LEVELS_2_PLUS, OPT_ftree_vrp, NULL, 1 }, - { OPT_LEVELS_2_PLUS, OPT_ftree_builtin_call_dce, NULL, 1 }, { OPT_LEVELS_2_PLUS, OPT_ftree_pre, NULL, 1 }, { OPT_LEVELS_2_PLUS, OPT_ftree_switch_conversion, NULL, 1 }, { OPT_LEVELS_2_PLUS, OPT_fipa_cp, NULL, 1 }, diff --git a/gcc/tree-call-cdce.c b/gcc/tree-call-cdce.c index a5f38ce..986c218 100644 --- a/gcc/tree-call-cdce.c +++ b/gcc/tree-call-cdce.c @@ -33,45 +33,77 @@ along with GCC; see the file COPYING3. If not see #include "gimple-iterator.h" #include "tree-cfg.h" #include "tree-into-ssa.h" +#include "builtins.h" +#include "internal-fn.h" -/* Conditional dead call elimination +/* This pass serves two closely-related purposes: - Some builtin functions can set errno on error conditions, but they - are otherwise pure. If the result of a call to such a function is - not used, the compiler can still not eliminate the call without - powerful interprocedural analysis to prove that the errno is not - checked. However, if the conditions under which the error occurs - are known, the compiler can conditionally dead code eliminate the - calls by shrink-wrapping the semi-dead calls into the error condition: + 1. It conditionally executes calls that set errno if (a) the result of + the call is unused and (b) a simple range check on the arguments can + detect most cases where errno does not need to be set. - built_in_call (args) - ==> - if (error_cond (args)) - built_in_call (args) + This is the "conditional dead-code elimination" that gave the pass + its original name, since the call is dead for most argument values. + The calls for which it helps are usually part of the C++ abstraction + penalty exposed after inlining. - An actual simple example is : - log (x); // Mostly dead call + 2. It looks for calls to built-in functions that set errno and whose + result is used. It checks whether there is an associated internal + function that doesn't set errno and whether the target supports + that internal function. If so, the pass uses the internal function + to compute the result of the built-in function but still arranges + for errno to be set when necessary. There are two ways of setting + errno: + + a. by protecting the original call with the same argument checks as (1) + + b. by protecting the original call with a check that the result + of the internal function is not equal to itself (i.e. is NaN). + + (b) requires that NaNs are the only erroneous results. It is not + appropriate for functions like log, which returns ERANGE for zero + arguments. (b) is also likely to perform worse than (a) because it + requires the result to be calculated first. The pass therefore uses + (a) when it can and uses (b) as a fallback. + + For (b) the pass can replace the original call with a call to + IFN_SET_EDOM, if the target supports direct assignments to errno. + + In both cases, arguments that require errno to be set should occur + rarely in practice. Checks of the errno result should also be rare, + but the compiler would need powerful interprocedural analysis to + prove that errno is not checked. It's much easier to add argument + checks or result checks instead. + + An example of (1) is: + + log (x); // Mostly dead call ==> - if (x <= 0) - log (x); + if (x <= 0) + log (x); + With this change, call to log (x) is effectively eliminated, as - in majority of the cases, log won't be called with x out of + in the majority of the cases, log won't be called with x out of range. The branch is totally predictable, so the branch cost is low. + An example of (2) is: + + y = sqrt (x); + ==> + y = IFN_SQRT (x); + if (x < 0) + sqrt (x); + + In the vast majority of cases we should then never need to call sqrt. + Note that library functions are not supposed to clear errno to zero without error. See IEEE Std 1003.1, section 2.3 Error Numbers, and section 7.5:3 of ISO/IEC 9899 (C99). The condition wrapping the builtin call is conservatively set to avoid too - aggressive (wrong) shrink wrapping. The optimization is called conditional - dead call elimination because the call is eliminated under the condition - that the input arguments would not lead to domain or range error (for - instance when x <= 0 for a log (x) call), however the chances that the error - condition is hit is very low (those builtin calls which are conditionally - dead are usually part of the C++ abstraction penalty exposed after - inlining). */ + aggressive (wrong) shrink wrapping. */ /* A structure for representing input domain of @@ -250,28 +282,15 @@ check_builtin_call (gcall *bcall) return check_target_format (arg); } -/* A helper function to determine if a builtin function call is a - candidate for conditional DCE. Returns true if the builtin call - is a candidate. */ +/* Return true if built-in function call CALL calls a math function + and if we know how to test the range of its arguments to detect _most_ + situations in which errno is not set. The test must err on the side + of treating non-erroneous values as potentially erroneous. */ static bool -is_call_dce_candidate (gcall *call) +can_test_argument_range (gcall *call) { - tree fn; - enum built_in_function fnc; - - /* Only potentially dead calls are considered. */ - if (gimple_call_lhs (call)) - return false; - - fn = gimple_call_fndecl (call); - if (!fn - || !DECL_BUILT_IN (fn) - || (DECL_BUILT_IN_CLASS (fn) != BUILT_IN_NORMAL)) - return false; - - fnc = DECL_FUNCTION_CODE (fn); - switch (fnc) + switch (DECL_FUNCTION_CODE (gimple_call_fndecl (call))) { /* Trig functions. */ CASE_FLT_FN (BUILT_IN_ACOS): @@ -305,6 +324,39 @@ is_call_dce_candidate (gcall *call) return false; } +/* Return true if CALL can produce a domain error (EDOM) but can never + produce a pole or range overflow error (ERANGE). This means that we + can tell whether a function would have set errno by testing whether + the result is a NaN. + + Note that we do not consider range underflow errors, which are + implementation-defined in C99. */ + +static bool +edom_only_function (gcall *call) +{ + switch (DECL_FUNCTION_CODE (gimple_call_fndecl (call))) + { + CASE_FLT_FN (BUILT_IN_ACOS): + CASE_FLT_FN (BUILT_IN_ASIN): + CASE_FLT_FN (BUILT_IN_ATAN): + CASE_FLT_FN (BUILT_IN_COS): + CASE_FLT_FN (BUILT_IN_EXP): + CASE_FLT_FN (BUILT_IN_EXP10): + CASE_FLT_FN (BUILT_IN_EXP2): + CASE_FLT_FN (BUILT_IN_EXPM1): + CASE_FLT_FN (BUILT_IN_SIGNIFICAND): + CASE_FLT_FN (BUILT_IN_SIN): + CASE_FLT_FN (BUILT_IN_SQRT): + CASE_FLT_FN (BUILT_IN_ATAN2): + CASE_FLT_FN (BUILT_IN_FMOD): + CASE_FLT_FN (BUILT_IN_REMAINDER): + return true; + + default: + return false; + } +} /* A helper function to generate gimple statements for one bound comparison. ARG is the call argument to @@ -704,33 +756,24 @@ gen_shrink_wrap_conditions (gcall *bi_call, vec conds, /* Probability of the branch (to the call) is taken. */ #define ERR_PROB 0.01 -/* The function to shrink wrap a partially dead builtin call - whose return value is not used anywhere, but has to be kept - live due to potential error condition. Returns true if the - transformation actually happens. */ +/* Shrink-wrap BI_CALL so that it is only called when the NCONDS conditions + in CONDS are true. + + Return true on success, in which case the cfg will have been updated. */ static bool -shrink_wrap_one_built_in_call (gcall *bi_call) +shrink_wrap_one_built_in_call_with_conds (gcall *bi_call, vec conds, + unsigned int nconds) { gimple_stmt_iterator bi_call_bsi; basic_block bi_call_bb, join_tgt_bb, guard_bb; edge join_tgt_in_edge_from_call, join_tgt_in_edge_fall_thru; edge bi_call_in_edge0, guard_bb_in_edge; - unsigned tn_cond_stmts, nconds; + unsigned tn_cond_stmts; unsigned ci; gimple *cond_expr = NULL; gimple *cond_expr_start; - auto_vec conds; - gen_shrink_wrap_conditions (bi_call, conds, &nconds); - - /* This can happen if the condition generator decides - it is not beneficial to do the transformation. Just - return false and do not do any transformation for - the call. */ - if (nconds == 0) - return false; - /* The cfg we want to create looks like this: [guard n-1] <- guard_bb (old block) @@ -869,6 +912,117 @@ shrink_wrap_one_built_in_call (gcall *bi_call) return true; } +/* Shrink-wrap BI_CALL so that it is only called when it might set errno + (but is always called if it would set errno). + + Return true on success, in which case the cfg will have been updated. */ + +static bool +shrink_wrap_one_built_in_call (gcall *bi_call) +{ + unsigned nconds = 0; + auto_vec conds; + gen_shrink_wrap_conditions (bi_call, conds, &nconds); + /* This can happen if the condition generator decides + it is not beneficial to do the transformation. Just + return false and do not do any transformation for + the call. */ + if (nconds == 0) + return false; + return shrink_wrap_one_built_in_call_with_conds (bi_call, conds, nconds); +} + +/* Return true if built-in function call CALL could be implemented using + a combination of an internal function to compute the result and a + separate call to set errno. */ + +static bool +can_use_internal_fn (gcall *call) +{ + /* Only replace calls that set errno. */ + if (!gimple_vdef (call)) + return false; + + /* Punt if we can't conditionalize the call. */ + basic_block bb = gimple_bb (call); + if (stmt_ends_bb_p (call) && !find_fallthru_edge (bb->succs)) + return false; + + /* See whether there is an internal function for this built-in. */ + if (replacement_internal_fn (call) == IFN_LAST) + return false; + + /* See whether we can catch all cases where errno would be set, + while still avoiding the call in most cases. */ + if (!can_test_argument_range (call) + && !edom_only_function (call)) + return false; + + return true; +} + +/* Implement built-in function call CALL using an internal function. + Return true on success, in which case the cfg will have changed. */ + +static bool +use_internal_fn (gcall *call) +{ + unsigned nconds = 0; + auto_vec conds; + gen_shrink_wrap_conditions (call, conds, &nconds); + if (nconds == 0 && !edom_only_function (call)) + return false; + + internal_fn ifn = replacement_internal_fn (call); + gcc_assert (ifn != IFN_LAST); + + /* Construct the new call, with the same arguments as the original one. */ + auto_vec args; + unsigned int nargs = gimple_call_num_args (call); + for (unsigned int i = 0; i < nargs; ++i) + args.safe_push (gimple_call_arg (call, i)); + gcall *new_call = gimple_build_call_internal_vec (ifn, args); + gimple_set_location (new_call, gimple_location (call)); + + /* Transfer the LHS to the new call. */ + tree lhs = gimple_call_lhs (call); + gimple_call_set_lhs (new_call, lhs); + gimple_call_set_lhs (call, NULL_TREE); + SSA_NAME_DEF_STMT (lhs) = new_call; + + /* Insert the new call. */ + gimple_stmt_iterator gsi = gsi_for_stmt (call); + gsi_insert_before (&gsi, new_call, GSI_SAME_STMT); + + if (nconds == 0) + { + /* Check LHS != LHS. If we reach here, EDOM is the only + valid errno value and it is used iff the result is NaN. */ + conds.quick_push (gimple_build_cond (NE_EXPR, lhs, lhs, + NULL_TREE, NULL_TREE)); + nconds++; + + /* Try replacing the original call with a direct assignment to + errno, via an internal function. */ + if (set_edom_supported_p () && !stmt_ends_bb_p (call)) + { + gimple_stmt_iterator gsi = gsi_for_stmt (call); + gcall *new_call = gimple_build_call_internal (IFN_SET_EDOM, 0); + gimple_set_vuse (new_call, gimple_vuse (call)); + gimple_set_vdef (new_call, gimple_vdef (call)); + SSA_NAME_DEF_STMT (gimple_vdef (new_call)) = new_call; + gimple_set_location (new_call, gimple_location (call)); + gsi_replace (&gsi, new_call, false); + call = new_call; + } + } + + if (!shrink_wrap_one_built_in_call_with_conds (call, conds, nconds)) + /* It's too late to back out now. */ + gcc_unreachable (); + return true; +} + /* The top level function for conditional dead code shrink wrapping transformation. */ @@ -885,7 +1039,10 @@ shrink_wrap_conditional_dead_built_in_calls (vec calls) for (; i < n ; i++) { gcall *bi_call = calls[i]; - changed |= shrink_wrap_one_built_in_call (bi_call); + if (gimple_call_lhs (bi_call)) + changed |= use_internal_fn (bi_call); + else + changed |= shrink_wrap_one_built_in_call (bi_call); } return changed; @@ -914,13 +1071,12 @@ public: {} /* opt_pass methods: */ - virtual bool gate (function *fun) + virtual bool gate (function *) { /* The limit constants used in the implementation assume IEEE floating point format. Other formats can be supported in the future if needed. */ - return flag_tree_builtin_call_dce != 0 - && optimize_function_for_speed_p (fun); + return flag_tree_builtin_call_dce != 0; } virtual unsigned int execute (function *); @@ -936,11 +1092,20 @@ pass_call_cdce::execute (function *fun) auto_vec cond_dead_built_in_calls; FOR_EACH_BB_FN (bb, fun) { + /* Skip blocks that are being optimized for size, since our + transformation always increases code size. */ + if (optimize_bb_for_size_p (bb)) + continue; + /* Collect dead call candidates. */ for (i = gsi_start_bb (bb); !gsi_end_p (i); gsi_next (&i)) { gcall *stmt = dyn_cast (gsi_stmt (i)); - if (stmt && is_call_dce_candidate (stmt)) + if (stmt + && gimple_call_builtin_p (stmt, BUILT_IN_NORMAL) + && (gimple_call_lhs (stmt) + ? can_use_internal_fn (stmt) + : can_test_argument_range (stmt))) { if (dump_file && (dump_flags & TDF_DETAILS)) {