From patchwork Tue Jun 22 12:50:49 2010 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Dmitry Plotnikov X-Patchwork-Id: 56474 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]) by ozlabs.org (Postfix) with SMTP id DC653B6F10 for ; Tue, 22 Jun 2010 22:51:01 +1000 (EST) Received: (qmail 7144 invoked by alias); 22 Jun 2010 12:50:59 -0000 Received: (qmail 7129 invoked by uid 22791); 22 Jun 2010 12:50:58 -0000 X-SWARE-Spam-Status: No, hits=-2.1 required=5.0 tests=AWL, BAYES_00, T_RP_MATCHES_RCVD X-Spam-Check-By: sourceware.org Received: from smtp.ispras.ru (HELO smtp.ispras.ru) (83.149.198.201) by sourceware.org (qpsmtpd/0.43rc1) with ESMTP; Tue, 22 Jun 2010 12:50:51 +0000 Received: from ispserv.ispras.ru (ispserv.ispras.ru [83.149.198.72]) by smtp.ispras.ru (Postfix) with ESMTP id 762675D40AF; Tue, 22 Jun 2010 16:44:02 +0400 (MSD) Received: from [10.10.3.60] (winnie.ispras.ru [83.149.198.236]) by ispserv.ispras.ru (Postfix) with ESMTP id 134213FC48; Tue, 22 Jun 2010 16:50:49 +0400 (MSD) Message-ID: <4C20B1A9.7060506@ispras.ru> Date: Tue, 22 Jun 2010 16:50:49 +0400 From: Dmitry Plotnikov User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.9.1.9) Gecko/20100423 Thunderbird/3.0.4 MIME-Version: 1.0 To: gcc-patches@gcc.gnu.org CC: rearnsha@arm.com, dm@ispras.ru Subject: [PATCH, ARM] VSHL, VSHR, VLSHR immediate values support X-IsSubscribed: yes 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 This is repost of our patch ( http://gcc.gnu.org/ml/gcc-patches/2010-02/msg00462.html), which also includes a bugfix to previous patch. This patch adds support for immediate shift values for NEON in ARM backend. We added new predicates for immediate shift value or register operands and used them in new patterns. The only tiny modification in target-independent part is in optabs.c and it has been tested on x86_64 without any regressions. Example: a[i] = a[i]>>4; Before: vmov.i32 q9, #4294967292 @ v4si ... vshl.u32 q8, q8, q9 With patch: vshr.u32 q8, q8, #4 The patch was tested on 2010-06-03 snapshot. OK for trunk? 2010-06-22 Dmitry Plotnikov Dmitry Melnik * config/arm/arm.c (neon_immediate_valid_for_shift): New function. (neon_output_shift_immediate): New function. * config/arm/neon.md (vashl3): Modified constraint. (vashr3_imm): New insn pattern. (vlshr3_imm): New insn pattern. (vashr3): Modified constraint. (vlshr3): Modified constraint. * config/arm/predicates.md (imm_for_neon_shift_operand): New predicate. (imm_shift_or_reg_neon): New predicate. * optabs.c (init_optabs): Init optab codes for vashl, vashr, vlshr. * testsuite/gcc.target/arm/neon-vshr-imm-1.c: New testcase. 2010-06-22 Dmitry Plotnikov Dmitry Melnik * config/arm/arm.c (neon_immediate_valid_for_shift): New function. (neon_output_shift_immediate): New function. * config/arm/neon.md (vashl3): Modified constraint. (vashr3_imm): New insn pattern. (vlshr3_imm): New insn pattern. (vashr3): Modified constraint. (vlshr3): Modified constraint. * config/arm/predicates.md (imm_for_neon_shift_operand): New predicate. (imm_shift_or_reg_neon): New predicate. * optabs.c (init_optabs): Init optab codes for vashl, vashr, vlshr. * testsuite/gcc.target/arm/neon-vshr-imm-1.c: New testcase. diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c index 84dd8fa..3fa19e8 100644 --- a/gcc/config/arm/arm.c +++ b/gcc/config/arm/arm.c @@ -8026,6 +8026,51 @@ neon_immediate_valid_for_logic (rtx op, enum machine_mode mode, int inverse, return 1; } +/* Return TRUE if rtx OP is legal for use in a VSHR or VSHL instruction. If + the immediate is valid, write a constant suitable for using as an operand + to VSHR/VSHL to *MODCONST and the corresponding element width to + *ELEMENTWIDTH. */ + +int +neon_immediate_valid_for_shift (rtx op, enum machine_mode mode, + rtx *modconst, int *elementwidth) +{ + unsigned int innersize = GET_MODE_SIZE (GET_MODE_INNER (mode)); + unsigned int n_elts = CONST_VECTOR_NUNITS (op), i; + unsigned HOST_WIDE_INT last_elt = 0; + + /* Split vector constant out into a byte vector. */ + for (i = 0; i < n_elts; i++) + { + rtx el = CONST_VECTOR_ELT (op, i); + unsigned HOST_WIDE_INT elpart; + + if (GET_CODE (el) == CONST_INT) + elpart = INTVAL (el); + else if (GET_CODE (el) == CONST_DOUBLE) + return 0; + else + gcc_unreachable (); + + if (i != 0 && elpart != last_elt) + return 0; + + last_elt = elpart; + } + + /* shift less than element size */ + if (last_elt > innersize * 8) + return 0; + + if (elementwidth) + *elementwidth = innersize * 8; + + if (modconst) + *modconst = CONST_VECTOR_ELT (op, 0); + + return 1; +} + /* Return a string suitable for output of Neon immediate logic operation MNEM. */ @@ -8048,6 +8093,28 @@ neon_output_logic_immediate (const char *mnem, rtx *op2, enum machine_mode mode, return templ; } +/* Return a string suitable for output of Neon immediate shift operation + (VSHR or VSHL) MNEM. */ + +char * +neon_output_shift_immediate (const char *mnem, char sign, rtx *op2, enum machine_mode mode, + int quad) +{ + int width, is_valid; + static char templ[40]; + + is_valid = neon_immediate_valid_for_shift (*op2, mode, op2, &width); + + gcc_assert (is_valid != 0); + + if (quad) + sprintf (templ, "%s.%c%d\t%%q0, %%q1, %%2", mnem, sign, width); + else + sprintf (templ, "%s.%c%d\t%%P0, %%P1, %%2", mnem, sign, width); + + return templ; +} + /* Output a sequence of pairwise operations to implement a reduction. NOTE: We do "too much work" here, because pairwise operations work on two registers-worth of operands in one go. Unfortunately we can't exploit those diff --git a/gcc/config/arm/neon.md b/gcc/config/arm/neon.md index 43b3805..43c282d 100644 --- a/gcc/config/arm/neon.md +++ b/gcc/config/arm/neon.md @@ -1135,11 +1135,53 @@ ; SImode elements. (define_insn "vashl3" + [(set (match_operand:VDQIW 0 "s_register_operand" "=w,w") + (ashift:VDQIW (match_operand:VDQIW 1 "s_register_operand" "w,w") + (match_operand:VDQIW 2 "imm_shift_or_reg_neon" "w,Dn")))] + "TARGET_NEON" + { + switch (which_alternative) + { + case 0: return "vshl.\t%0, %1, %2"; + case 1: return neon_output_shift_immediate ("vshl", 's', &operands[2], + mode, VALID_NEON_QREG_MODE (mode)); + default: gcc_unreachable (); + } + } + + [(set (attr "neon_type") + (if_then_else (ne (symbol_ref "") (const_int 0)) + (const_string "neon_vshl_ddd") + (const_string "neon_shift_3")))] +) + + +(define_insn "vashr3_imm" [(set (match_operand:VDQIW 0 "s_register_operand" "=w") - (ashift:VDQIW (match_operand:VDQIW 1 "s_register_operand" "w") - (match_operand:VDQIW 2 "s_register_operand" "w")))] + (ashiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "w") + (match_operand:VDQIW 2 "imm_for_neon_shift_operand" "Dn")))] "TARGET_NEON" - "vshl.\t%0, %1, %2" + { + return neon_output_shift_immediate ("vshr", 's', &operands[2], + mode, VALID_NEON_QREG_MODE (mode)); + } + + [(set (attr "neon_type") + (if_then_else (ne (symbol_ref "") (const_int 0)) + (const_string "neon_vshl_ddd") + (const_string "neon_shift_3")))] +) + +(define_insn "vlshr3_imm" + [(set (match_operand:VDQIW 0 "s_register_operand" "=w") + (lshiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "w") + (match_operand:VDQIW 2 "imm_for_neon_shift_operand" "Dn")))] + "TARGET_NEON" + { + return neon_output_shift_immediate ("vshr", 'u', &operands[2], + mode, VALID_NEON_QREG_MODE (mode)); + } + [(set (attr "neon_type") (if_then_else (ne (symbol_ref "") (const_int 0)) (const_string "neon_vshl_ddd") @@ -1182,29 +1224,35 @@ (define_expand "vashr3" [(set (match_operand:VDQIW 0 "s_register_operand" "") - (ashiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "") - (match_operand:VDQIW 2 "s_register_operand" "")))] + (ashiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "") + (match_operand:VDQIW 2 "imm_shift_or_reg_neon" "")))] "TARGET_NEON" { rtx neg = gen_reg_rtx (mode); - - emit_insn (gen_neg2 (neg, operands[2])); - emit_insn (gen_ashl3_signed (operands[0], operands[1], neg)); - + if (REG_P (operands[2])) + { + emit_insn (gen_neg2 (neg, operands[2])); + emit_insn (gen_ashl3_signed (operands[0], operands[1], neg)); + } + else + emit_insn(gen_vashr3_imm (operands[0], operands[1], operands[2])); DONE; }) (define_expand "vlshr3" [(set (match_operand:VDQIW 0 "s_register_operand" "") - (lshiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "") - (match_operand:VDQIW 2 "s_register_operand" "")))] + (lshiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "") + (match_operand:VDQIW 2 "imm_shift_or_reg_neon" "")))] "TARGET_NEON" { rtx neg = gen_reg_rtx (mode); - - emit_insn (gen_neg2 (neg, operands[2])); - emit_insn (gen_ashl3_unsigned (operands[0], operands[1], neg)); - + if (REG_P (operands[2])) + { + emit_insn (gen_neg2 (neg, operands[2])); + emit_insn (gen_ashl3_unsigned (operands[0], operands[1], neg)); + } + else + emit_insn(gen_vlshr3_imm (operands[0], operands[1], operands[2])); DONE; }) diff --git a/gcc/config/arm/predicates.md b/gcc/config/arm/predicates.md index d351f44..02fced4 100644 --- a/gcc/config/arm/predicates.md +++ b/gcc/config/arm/predicates.md @@ -496,6 +496,17 @@ return neon_immediate_valid_for_move (op, mode, NULL, NULL); }) +(define_predicate "imm_for_neon_shift_operand" + (match_code "const_vector") +{ + return neon_immediate_valid_for_shift (op, mode, NULL, NULL); +}) + + +(define_predicate "imm_shift_or_reg_neon" + (ior (match_operand 0 "s_register_operand") + (match_operand 0 "imm_for_neon_shift_operand"))) + (define_predicate "imm_for_neon_logic_operand" (match_code "const_vector") { diff --git a/gcc/optabs.c b/gcc/optabs.c index 555e256..dea1d9c 100644 --- a/gcc/optabs.c +++ b/gcc/optabs.c @@ -6260,6 +6260,9 @@ init_optabs (void) init_optab (usashl_optab, US_ASHIFT); init_optab (ashr_optab, ASHIFTRT); init_optab (lshr_optab, LSHIFTRT); + init_optabv (vashl_optab, ASHIFT); + init_optabv (vashr_optab, ASHIFTRT); + init_optabv (vlshr_optab, LSHIFTRT); init_optab (rotl_optab, ROTATE); init_optab (rotr_optab, ROTATERT); init_optab (smin_optab, SMIN);