From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from xry111.site (xry111.site [89.208.246.23]) by sourceware.org (Postfix) with ESMTPS id 5B3AF3858291 for ; Fri, 7 Jul 2023 13:22:15 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.2 sourceware.org 5B3AF3858291 Authentication-Results: sourceware.org; dmarc=pass (p=reject dis=none) header.from=xry111.site Authentication-Results: sourceware.org; spf=pass smtp.mailfrom=xry111.site DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=xry111.site; s=default; t=1688736134; bh=ZqVoHGeeeIVKSCdoq9LoHxTRH+ne/lmKoBF3RulRHKs=; h=From:To:Cc:Subject:Date:From; b=R8N5TM9bOfKqdsU4Os6jjTVfsEEduynxS0Lq26879i44ID6uoxC/ei8iDMwKLFf6e k49HA00UpMLP3giGJIkSjjl74/GiQjeTmvzhJNFk6BONcU/8wRFmhKqLN+yGU/URQy MhyXumA0S6ijDyEhYPOtWfLRGtXHSBphiyO0z4Sc= Received: from stargazer.. (unknown [IPv6:240e:358:110b:4200:dc73:854d:832e:5]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature ECDSA (P-384) server-digest SHA384) (Client did not present a certificate) (Authenticated sender: xry111@xry111.site) by xry111.site (Postfix) with ESMTPSA id 1A2F966311; Fri, 7 Jul 2023 09:22:06 -0400 (EDT) From: Xi Ruoyao To: gcc-patches@gcc.gnu.org Cc: Andre Vieira , Richard Biener , Jakub Jelinek , Hongtao Liu , Xi Ruoyao Subject: [PATCH v2] vect: Fix vectorized BIT_FIELD_REF for signed bit-fields [PR110557] Date: Fri, 7 Jul 2023 21:18:58 +0800 Message-ID: <20230707131857.2386125-2-xry111@xry111.site> X-Mailer: git-send-email 2.41.0 MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-Spam-Status: No, score=-8.9 required=5.0 tests=BAYES_00,DKIM_SIGNED,DKIM_VALID,DKIM_VALID_AU,DKIM_VALID_EF,GIT_PATCH_0,LIKELY_SPAM_FROM,SPF_HELO_PASS,SPF_PASS,TXREP,T_SCC_BODY_TEXT_LINE autolearn=ham autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org List-Id: If a bit-field is signed and it's wider than the output type, we must ensure the extracted result sign-extended. But this was not handled correctly. For example: int x : 8; long y : 55; bool z : 1; The vectorized extraction of y was: vect__ifc__49.29_110 = MEM [(struct Item *)vectp_a.27_108]; vect_patt_38.30_112 = vect__ifc__49.29_110 & { 9223372036854775552, 9223372036854775552 }; vect_patt_39.31_113 = vect_patt_38.30_112 >> 8; vect_patt_40.32_114 = VIEW_CONVERT_EXPR(vect_patt_39.31_113); This is obviously incorrect. This pach has implemented it as: vect__ifc__25.16_62 = MEM [(struct Item *)vectp_a.14_60]; vect_patt_31.17_63 = VIEW_CONVERT_EXPR(vect__ifc__25.16_62); vect_patt_32.18_64 = vect_patt_31.17_63 << 1; vect_patt_33.19_65 = vect_patt_32.18_64 >> 9; gcc/ChangeLog: PR tree-optimization/110557 * tree-vect-patterns.cc (vect_recog_bitfield_ref_pattern): Ensure the output sign-extended if necessary. gcc/testsuite/ChangeLog: PR tree-optimization/110557 * g++.dg/vect/pr110557.cc: New test. --- Change v1 -> v2: - Rename two variables for readability. - Remove a redundant useless_type_conversion_p check. - Edit the comment for early conversion to show the rationale of "|| ref_sext". Bootstrapped (with BOOT_CFLAGS="-O3 -mavx2") and regtested on x86_64-linux-gnu. Ok for trunk and gcc-13? gcc/testsuite/g++.dg/vect/pr110557.cc | 37 ++++++++++++++++ gcc/tree-vect-patterns.cc | 62 ++++++++++++++++++++------- 2 files changed, 83 insertions(+), 16 deletions(-) create mode 100644 gcc/testsuite/g++.dg/vect/pr110557.cc diff --git a/gcc/testsuite/g++.dg/vect/pr110557.cc b/gcc/testsuite/g++.dg/vect/pr110557.cc new file mode 100644 index 00000000000..e1fbe1caac4 --- /dev/null +++ b/gcc/testsuite/g++.dg/vect/pr110557.cc @@ -0,0 +1,37 @@ +// { dg-additional-options "-mavx" { target { avx_runtime } } } + +static inline long +min (long a, long b) +{ + return a < b ? a : b; +} + +struct Item +{ + int x : 8; + long y : 55; + bool z : 1; +}; + +__attribute__ ((noipa)) long +test (Item *a, int cnt) +{ + long size = 0; + for (int i = 0; i < cnt; i++) + size = min ((long)a[i].y, size); + return size; +} + +int +main () +{ + struct Item items[] = { + { 1, -1 }, + { 2, -2 }, + { 3, -3 }, + { 4, -4 }, + }; + + if (test (items, 4) != -4) + __builtin_trap (); +} diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc index 1bc36b043a0..c0832e8679f 100644 --- a/gcc/tree-vect-patterns.cc +++ b/gcc/tree-vect-patterns.cc @@ -2566,7 +2566,7 @@ vect_recog_widen_sum_pattern (vec_info *vinfo, Widening with mask first, shift later: container = (type_out) container; masked = container & (((1 << bitsize) - 1) << bitpos); - result = patt2 >> masked; + result = masked >> bitpos; Widening with shift first, mask last: container = (type_out) container; @@ -2578,6 +2578,15 @@ vect_recog_widen_sum_pattern (vec_info *vinfo, result = masked >> bitpos; result = (type_out) result; + If the bitfield is signed and it's wider than type_out, we need to + keep the result sign-extended: + container = (type) container; + masked = container << (prec - bitsize - bitpos); + result = (type_out) (masked >> (prec - bitsize)); + + Here type is the signed variant of the wider of type_out and the type + of container. + The shifting is always optional depending on whether bitpos != 0. */ @@ -2636,14 +2645,22 @@ vect_recog_bitfield_ref_pattern (vec_info *vinfo, stmt_vec_info stmt_info, if (BYTES_BIG_ENDIAN) shift_n = prec - shift_n - mask_width; + bool ref_sext = (!TYPE_UNSIGNED (TREE_TYPE (bf_ref)) && + TYPE_PRECISION (ret_type) > mask_width); + bool load_widen = (TYPE_PRECISION (TREE_TYPE (container)) < + TYPE_PRECISION (ret_type)); + /* We move the conversion earlier if the loaded type is smaller than the - return type to enable the use of widening loads. */ - if (TYPE_PRECISION (TREE_TYPE (container)) < TYPE_PRECISION (ret_type) - && !useless_type_conversion_p (TREE_TYPE (container), ret_type)) - { - pattern_stmt - = gimple_build_assign (vect_recog_temp_ssa_var (ret_type), - NOP_EXPR, container); + return type to enable the use of widening loads. And if we need a + sign extension, we need to convert the loaded value early to a signed + type as well. */ + if (ref_sext || load_widen) + { + tree type = load_widen ? ret_type : container_type; + if (ref_sext) + type = gimple_signed_type (type); + pattern_stmt = gimple_build_assign (vect_recog_temp_ssa_var (type), + NOP_EXPR, container); container = gimple_get_lhs (pattern_stmt); container_type = TREE_TYPE (container); prec = tree_to_uhwi (TYPE_SIZE (container_type)); @@ -2671,7 +2688,7 @@ vect_recog_bitfield_ref_pattern (vec_info *vinfo, stmt_vec_info stmt_info, shift_first = true; tree result; - if (shift_first) + if (shift_first && !ref_sext) { tree shifted = container; if (shift_n) @@ -2694,14 +2711,27 @@ vect_recog_bitfield_ref_pattern (vec_info *vinfo, stmt_vec_info stmt_info, } else { - tree mask = wide_int_to_tree (container_type, - wi::shifted_mask (shift_n, mask_width, - false, prec)); - pattern_stmt - = gimple_build_assign (vect_recog_temp_ssa_var (container_type), - BIT_AND_EXPR, container, mask); - tree masked = gimple_assign_lhs (pattern_stmt); + tree temp = vect_recog_temp_ssa_var (container_type); + if (!ref_sext) + { + tree mask = wide_int_to_tree (container_type, + wi::shifted_mask (shift_n, + mask_width, + false, prec)); + pattern_stmt = gimple_build_assign (temp, BIT_AND_EXPR, + container, mask); + } + else + { + HOST_WIDE_INT shl = prec - shift_n - mask_width; + shift_n += shl; + pattern_stmt = gimple_build_assign (temp, LSHIFT_EXPR, + container, + build_int_cst (sizetype, + shl)); + } + tree masked = gimple_assign_lhs (pattern_stmt); append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype); pattern_stmt = gimple_build_assign (vect_recog_temp_ssa_var (container_type), -- 2.41.0