From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-qk1-x744.google.com (mail-qk1-x744.google.com [IPv6:2607:f8b0:4864:20::744]) by sourceware.org (Postfix) with ESMTPS id 86535385DC1C for ; Sat, 25 Apr 2020 04:09:41 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.3.2 sourceware.org 86535385DC1C Received: by mail-qk1-x744.google.com with SMTP id n143so12488315qkn.8 for ; Fri, 24 Apr 2020 21:09:41 -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; bh=O2Qpvij6ih3WJ2B2PZi7UVKW1gSYnRHtA7ntgy0WsVk=; b=mBSDRx8v3Je4r1rf0PIjPEGQXwflzOxDuhk6bS480Et+jjUsR8LkXK48YhopdKQasd QC8yGu26WF5txbCBVNAHvQBuPqlOWudENZ2K3HdAGV6JNtYge+aUsy5PSqbEEptCScM7 fR2JfctKYPbl94BhoppqS/+QHjbEo1PnQpVLImLVmWNu0KyM6Qoqy6t0Uh6ErkN4KAon GxyceZbhhydAlRSmdxAIQoyjQhnjlUS0EQu8iPLlAP5A4CnukWXpI78H3JhL/Fuh+aRF KZ3bWiFWrEc/gS+2CuRp4mulOu3gwS5J4Y7qhyOw2wkVqByF/5iVkjBRRhYCOPqwSBzQ NC6w== X-Gm-Message-State: AGi0PuaWM2lytHCEIz1om3Rwrj0B1Ao7f33ERKe4vjmlknROfHf4Io1O 3608kSPb6eJQ0ZGOk2/G+t9VuQQqjcY= X-Google-Smtp-Source: APiQypJdiLNwtDCClVyeZX+m1bDjgwXrYCsL0p50W37FYfTDpbOqfGmGz3Fgf+bCpZkcWmgYY+vKeg== X-Received: by 2002:a37:9a57:: with SMTP id c84mr11984555qke.380.1587787780550; Fri, 24 Apr 2020 21:09:40 -0700 (PDT) Received: from localhost.localdomain ([191.251.166.36]) by smtp.gmail.com with ESMTPSA id k33sm5181486qtd.22.2020.04.24.21.09.38 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Fri, 24 Apr 2020 21:09:39 -0700 (PDT) From: Luis Machado To: gdb-patches@sourceware.org Cc: tromey@adacore.com, tdevries@suse.de, andrew.burgess@embecosm.com Subject: [PATCH] Fix remaining inline/tailcall unwinding breakage for x86_64 Date: Sat, 25 Apr 2020 01:09:34 -0300 Message-Id: <20200425040934.17011-1-luis.machado@linaro.org> X-Mailer: git-send-email 2.17.1 X-Spam-Status: No, score=-19.5 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, GIT_PATCH_0, GIT_PATCH_1, GIT_PATCH_2, GIT_PATCH_3, KAM_NUMSUBJECT, 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: gdb-patches@sourceware.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Gdb-patches mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Sat, 25 Apr 2020 04:09:43 -0000 Commit 5939967b355ba6a940887d19847b7893a4506067 fixed inline frame unwinding breakage for some targets (aarch64, riscv, s390...) but regressed a few amd64 testcases related to tailcalls. Given the following example situation... Frame #-1 - sentinel frame Frame # 0 - inline frame Frame # 1 - normal frame ... suppose we're at level #1 and call into dwarf2_tailcall_sniffer_first. We'll attempt to fetch PC, which used to be done via the gdbarch_unwind_pc call (before 5939967b355ba6a940887d19847b7893a4506067), but now it is being handled by the get_frame_register function. gdbarch_unwind_pc will attempt to use frame #1's cache to retrieve information about the PC. Here's where different architectures behave differently. x86_64 will find a dwarf rule to retrieve PC from memory, at a CFA + offset location. So the PC value is readily available and there is no need to create a lazy value. For aarch64 (and others), GCC doesn't emit an explicit location for PC, so we eventually will find that PC is DWARF2_FRAME_REG_UNSPECIFIED. This is known and is handled by GDB by assuming GCC really meant DWARF2_FRAME_REG_SAME_VALUE. This means we'll attempt to fetch the register value from frame #0, via a call to frame_unwind_got_register, which will trigger the creation of a lazy value that requires a valid frame id for frame #0. We don't have a valid id for frame #0 yet, so we assert. Given the above, the following patch attempts to handle the situation without being too hacky. We verify if the next frame is an inline frame and if its frame id has been computed already. If it hasn't been computed yet, then we use the safer get_frame_register function, otherwise we use the regular gdbarch_unwind_pc hook. I've verified this makes both aarch64-linux and x86_64 happy testsuite-wise. gdb/ChangeLog: 2020-04-25 Luis Machado * dwarf2/frame-tailcall.c (dwarf2_tailcall_sniffer_first): Handle problematic inline frame unwinding situation. * frame.c (frame_id_computed_p): New function. * frame.h (frame_id_computed_p): New prototype. --- gdb/dwarf2/frame-tailcall.c | 39 ++++++++++++++++++++++++++++++++++--- gdb/frame.c | 8 ++++++++ gdb/frame.h | 4 ++++ 3 files changed, 48 insertions(+), 3 deletions(-) diff --git a/gdb/dwarf2/frame-tailcall.c b/gdb/dwarf2/frame-tailcall.c index 01bb134a5c..16dba2b201 100644 --- a/gdb/dwarf2/frame-tailcall.c +++ b/gdb/dwarf2/frame-tailcall.c @@ -384,10 +384,43 @@ dwarf2_tailcall_sniffer_first (struct frame_info *this_frame, prev_gdbarch = frame_unwind_arch (this_frame); + /* The dwarf2 tailcall sniffer runs early, at the end of populating the + dwarf2 frame cache for the current frame. If there exists inline + frames inner (next) to the current frame, there is a good possibility + of that inline frame not having a computed frame id yet. + + This is because computing such a frame id requires us to walk through + the frame chain until we find the first normal frame after the inline + frame and then compute the normal frame's id first. + + Some architectures' compilers generate enough register location + information for a dwarf unwinder to fetch PC without relying on inner + frames (x86_64 for example). In this case the PC is retrieved + according to dwarf rules. + + But others generate less strict dwarf data for which assumptions are + made (like interpreting DWARF2_FRAME_REG_UNSPECIFIED as + DWARF2_FRAME_REG_SAME_VALUE). For such cases, GDB may attempt to + create lazy values for registers, and those lazy values must be + created with a valid frame id, but we potentially have no valid id. + + So, to avoid breakage, if we see a dangerous situation with inline + frames without a computed id, use safer functions to retrieve the + current frame's PC. Otherwise use the provided dwarf rules. */ + frame_info *next_frame = get_next_frame (this_frame); + /* Simulate frame_unwind_pc without setting this_frame->prev_pc.p. */ - get_frame_register (this_frame, gdbarch_pc_regnum (prev_gdbarch), - (gdb_byte *) &prev_pc); - prev_pc = gdbarch_addr_bits_remove (prev_gdbarch, prev_pc); + if (next_frame != nullptr && get_frame_type (next_frame) == INLINE_FRAME + && !frame_id_computed_p (next_frame)) + { + /* The next frame is an inline frame and its frame id has not been + computed yet. */ + get_frame_register (this_frame, gdbarch_pc_regnum (prev_gdbarch), + (gdb_byte *) &prev_pc); + prev_pc = gdbarch_addr_bits_remove (prev_gdbarch, prev_pc); + } + else + prev_pc = gdbarch_unwind_pc (prev_gdbarch, this_frame); /* call_site_find_chain can throw an exception. */ chain = call_site_find_chain (prev_gdbarch, prev_pc, this_pc); diff --git a/gdb/frame.c b/gdb/frame.c index ac1016b083..ff27b9f00e 100644 --- a/gdb/frame.c +++ b/gdb/frame.c @@ -687,6 +687,14 @@ frame_id_build_wild (CORE_ADDR stack_addr) return id; } +bool +frame_id_computed_p (struct frame_info *frame) +{ + gdb_assert (frame != nullptr); + + return frame->this_id.p != 0; +} + int frame_id_p (struct frame_id l) { diff --git a/gdb/frame.h b/gdb/frame.h index cfc15022ed..e835d49f9c 100644 --- a/gdb/frame.h +++ b/gdb/frame.h @@ -236,6 +236,10 @@ extern struct frame_id as the special identifier address are set to indicate wild cards. */ extern struct frame_id frame_id_build_wild (CORE_ADDR stack_addr); +/* Returns true if FRAME's id has been computed. + Returns false otherwise. */ +extern bool frame_id_computed_p (struct frame_info *frame); + /* Returns non-zero when L is a valid frame (a valid frame has a non-zero .base). The outermost frame is valid even without an ID. */ -- 2.17.1