Message ID | 20221118155252.113476-4-torbjorn.svensson@foss.st.com |
---|---|
State | Superseded |
Headers |
Return-Path: <gdb-patches-bounces+patchwork=sourceware.org@sourceware.org> X-Original-To: patchwork@sourceware.org Delivered-To: patchwork@sourceware.org Received: from server2.sourceware.org (localhost [IPv6:::1]) by sourceware.org (Postfix) with ESMTP id 600D5384D98B for <patchwork@sourceware.org>; Fri, 18 Nov 2022 15:54:26 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 600D5384D98B DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=sourceware.org; s=default; t=1668786866; bh=qJdTyUnFBrFGQyYpt/6GddPn19rNxpurYxcOUEnsCG8=; h=To:CC:Subject:Date:In-Reply-To:References:List-Id: List-Unsubscribe:List-Archive:List-Post:List-Help:List-Subscribe: From:Reply-To:From; b=jhIAKP/jJcT4mG/yCBly3zIVfn896M2ictPgjzvGTkZVfsVJMHJudV5PYAQlfbOW7 Y2Vh5jWh5Jn/oJ06KASmZaKOz6JPKKqcjZKuGPhUVJ3mC+8CdcPwDS8bmGk5iR+cKT WVBEWsN4uVq2A/CN1IBhIlwM69lxohQJs1sqPHCI= X-Original-To: gdb-patches@sourceware.org Delivered-To: gdb-patches@sourceware.org Received: from mx07-00178001.pphosted.com (mx07-00178001.pphosted.com [185.132.182.106]) by sourceware.org (Postfix) with ESMTPS id CF0AA385457F for <gdb-patches@sourceware.org>; Fri, 18 Nov 2022 15:53:48 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org CF0AA385457F Received: from pps.filterd (m0288072.ppops.net [127.0.0.1]) by mx07-00178001.pphosted.com (8.17.1.19/8.17.1.19) with ESMTP id 2AICS16A014845; Fri, 18 Nov 2022 16:53:46 +0100 Received: from beta.dmz-eu.st.com (beta.dmz-eu.st.com [164.129.1.35]) by mx07-00178001.pphosted.com (PPS) with ESMTPS id 3kx0ph4t0c-1 (version=TLSv1.2 cipher=ECDHE-RSA-AES256-GCM-SHA384 bits=256 verify=NOT); Fri, 18 Nov 2022 16:53:46 +0100 Received: from euls16034.sgp.st.com (euls16034.sgp.st.com [10.75.44.20]) by beta.dmz-eu.st.com (STMicroelectronics) with ESMTP id 8388C10003B; Fri, 18 Nov 2022 16:53:41 +0100 (CET) Received: from Webmail-eu.st.com (shfdag1node3.st.com [10.75.129.71]) by euls16034.sgp.st.com (STMicroelectronics) with ESMTP id 7C63523C68D; Fri, 18 Nov 2022 16:53:41 +0100 (CET) Received: from jkgcxl0002.jkg.st.com (10.210.54.218) by SHFDAG1NODE3.st.com (10.75.129.71) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256) id 15.1.2507.6; Fri, 18 Nov 2022 16:53:40 +0100 To: <gdb-patches@sourceware.org> CC: <luis.machado@arm.com>, <vanekt@volny.cz>, =?utf-8?q?Torbj=C3=B6rn_SVENS?= =?utf-8?q?SON?= <torbjorn.svensson@foss.st.com>, Yvan Roux <yvan.roux@foss.st.com> Subject: [PATCH v2 3/4] gdb: dwarf2 generic implementation for caching function data Date: Fri, 18 Nov 2022 16:52:52 +0100 Message-ID: <20221118155252.113476-4-torbjorn.svensson@foss.st.com> X-Mailer: git-send-email 2.25.1 In-Reply-To: <20221118155252.113476-1-torbjorn.svensson@foss.st.com> References: <20221118155252.113476-1-torbjorn.svensson@foss.st.com> MIME-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 8bit X-Originating-IP: [10.210.54.218] X-ClientProxiedBy: EQNCAS1NODE3.st.com (10.75.129.80) To SHFDAG1NODE3.st.com (10.75.129.71) X-Proofpoint-Virus-Version: vendor=baseguard engine=ICAP:2.0.219,Aquarius:18.0.895,Hydra:6.0.545,FMLib:17.11.122.1 definitions=2022-11-18_04,2022-11-18_01,2022-06-22_01 X-Spam-Status: No, score=-11.8 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, GIT_PATCH_0, RCVD_IN_DNSWL_LOW, SPF_HELO_NONE, SPF_PASS, TXREP autolearn=ham autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org X-BeenThere: gdb-patches@sourceware.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Gdb-patches mailing list <gdb-patches.sourceware.org> List-Unsubscribe: <https://sourceware.org/mailman/options/gdb-patches>, <mailto:gdb-patches-request@sourceware.org?subject=unsubscribe> List-Archive: <https://sourceware.org/pipermail/gdb-patches/> List-Post: <mailto:gdb-patches@sourceware.org> List-Help: <mailto:gdb-patches-request@sourceware.org?subject=help> List-Subscribe: <https://sourceware.org/mailman/listinfo/gdb-patches>, <mailto:gdb-patches-request@sourceware.org?subject=subscribe> From: =?utf-8?q?Torbj=C3=B6rn_SVENSSON_via_Gdb-patches?= <gdb-patches@sourceware.org> Reply-To: =?utf-8?q?Torbj=C3=B6rn_SVENSSON?= <torbjorn.svensson@foss.st.com> Errors-To: gdb-patches-bounces+patchwork=sourceware.org@sourceware.org Sender: "Gdb-patches" <gdb-patches-bounces+patchwork=sourceware.org@sourceware.org> |
Series |
v2 gdb/arm: Fixes for Cortex-M stack unwinding
|
|
Commit Message
Torbjörn SVENSSON
Nov. 18, 2022, 3:52 p.m. UTC
When there is no dwarf2 data for a register, a function can be called to provide the value of this register. In some situations, it might not be trivial to determine the value to return and it would cause a performance bottleneck to do the computation each time. This patch allows the called function to have a "cache" object that it can use to store some metadata between calls to reduce the performance impact of the complex logic. The cache object is unique for each function and frame, so if there are more than one function pointer stored in the dwarf2_frame_cache->reg array, then the appropriate pointer will be supplied (the type is not known by the dwarf2 implementation). dwarf2_frame_get_fn_data can be used to retrieve the function unique cache object. dwarf2_frame_allocate_fn_data can be used to allocate and retrieve the function unqiue cache object. Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com> Signed-off-by: Yvan Roux <yvan.roux@foss.st.com> --- gdb/dwarf2/frame.c | 48 ++++++++++++++++++++++++++++++++++++++++++++++ gdb/dwarf2/frame.h | 20 +++++++++++++++++-- 2 files changed, 66 insertions(+), 2 deletions(-)
Comments
Hi, On 2022-11-18 16:52, Torbjörn SVENSSON wrote: > When there is no dwarf2 data for a register, a function can be called > to provide the value of this register. In some situations, it might > not be trivial to determine the value to return and it would cause a > performance bottleneck to do the computation each time. > > This patch allows the called function to have a "cache" object that it > can use to store some metadata between calls to reduce the performance > impact of the complex logic. > > The cache object is unique for each function and frame, so if there are > more than one function pointer stored in the dwarf2_frame_cache->reg > array, then the appropriate pointer will be supplied (the type is not > known by the dwarf2 implementation). > > dwarf2_frame_get_fn_data can be used to retrieve the function unique > cache object. > dwarf2_frame_allocate_fn_data can be used to allocate and retrieve the > function unqiue cache object. > > Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com> > Signed-off-by: Yvan Roux <yvan.roux@foss.st.com> > --- > gdb/dwarf2/frame.c | 48 ++++++++++++++++++++++++++++++++++++++++++++++ > gdb/dwarf2/frame.h | 20 +++++++++++++++++-- > 2 files changed, 66 insertions(+), 2 deletions(-) > > diff --git a/gdb/dwarf2/frame.c b/gdb/dwarf2/frame.c > index 3f884abe1d5..bff3b706e7e 100644 > --- a/gdb/dwarf2/frame.c > +++ b/gdb/dwarf2/frame.c > @@ -831,6 +831,14 @@ dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, CORE_ADDR pc, > } > > > +struct dwarf2_frame_fn_data > +{ > + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, > + int regnum); > + void *data; > + struct dwarf2_frame_fn_data* next; > +}; > + > struct dwarf2_frame_cache > { > /* DWARF Call Frame Address. */ > @@ -862,6 +870,8 @@ struct dwarf2_frame_cache > dwarf2_tailcall_frame_unwind unwinder so this field does not apply for > them. */ > void *tailcall_cache; > + > + struct dwarf2_frame_fn_data *fn_data; > }; > > static struct dwarf2_frame_cache * > @@ -1221,6 +1231,44 @@ dwarf2_frame_prev_register (frame_info_ptr this_frame, void **this_cache, > } > } > > +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void **this_cache, > + fn_prev_register fn) > +{ > + struct dwarf2_frame_fn_data *fn_data = nullptr; > + struct dwarf2_frame_cache *cache > + = dwarf2_frame_cache (this_frame, this_cache); > + > + /* Find the object for the function. */ > + for (fn_data = cache->fn_data; fn_data; fn_data = fn_data->next) > + if (fn_data->fn == fn) > + return fn_data->data; > + > + return nullptr; > +} > + > +void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, > + void **this_cache, > + fn_prev_register fn, unsigned long size) > +{ > + struct dwarf2_frame_fn_data *fn_data = nullptr; > + struct dwarf2_frame_cache *cache > + = dwarf2_frame_cache (this_frame, this_cache); > + > + /* First try to find an existing object. */ > + void *data = dwarf2_frame_get_fn_data (this_frame, this_cache, fn); > + if (data) > + return data; > + > + /* No object found, lets create a new instance. */ > + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); > + fn_data->fn = fn; > + fn_data->data = frame_obstack_zalloc (size); Since these 2 blocks (fn_data and fn_data->fn) are allocated on the obstack, do we need to release them in dwarf2_frame_dealloc_cache() or just leave them to the generic garbage collection? Kind regards, Torbjörn > + fn_data->next = cache->fn_data; > + cache->fn_data = fn_data; > + > + return fn_data->data; > +} > + > /* Proxy for tailcall_frame_dealloc_cache for bottom frame of a virtual tail > call frames chain. */ > > diff --git a/gdb/dwarf2/frame.h b/gdb/dwarf2/frame.h > index 06c8a10c178..444afd9f8eb 100644 > --- a/gdb/dwarf2/frame.h > +++ b/gdb/dwarf2/frame.h > @@ -66,6 +66,9 @@ enum dwarf2_frame_reg_rule > > /* Register state. */ > > +typedef struct value *(*fn_prev_register) (frame_info_ptr this_frame, > + void **this_cache, int regnum); > + > struct dwarf2_frame_state_reg > { > /* Each register save state can be described in terms of a CFA slot, > @@ -78,8 +81,7 @@ struct dwarf2_frame_state_reg > const gdb_byte *start; > ULONGEST len; > } exp; > - struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, > - int regnum); > + fn_prev_register fn; > } loc; > enum dwarf2_frame_reg_rule how; > }; > @@ -262,4 +264,18 @@ extern int dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, CORE_ADDR pc, > const gdb_byte **cfa_start_out, > const gdb_byte **cfa_end_out); > > + > +/* Allocate a new instance of the function unique data. */ > + > +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, > + void **this_cache, > + fn_prev_register fn, > + unsigned long size); > + > +/* Retrieve the function unique data for this frame. */ > + > +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, > + void **this_cache, > + fn_prev_register fn); > + > #endif /* dwarf2-frame.h */
Hi, On 11/18/22 15:52, Torbjörn SVENSSON wrote: > When there is no dwarf2 data for a register, a function can be called > to provide the value of this register. In some situations, it might > not be trivial to determine the value to return and it would cause a > performance bottleneck to do the computation each time. > > This patch allows the called function to have a "cache" object that it > can use to store some metadata between calls to reduce the performance > impact of the complex logic. > > The cache object is unique for each function and frame, so if there are > more than one function pointer stored in the dwarf2_frame_cache->reg > array, then the appropriate pointer will be supplied (the type is not > known by the dwarf2 implementation). > > dwarf2_frame_get_fn_data can be used to retrieve the function unique > cache object. > dwarf2_frame_allocate_fn_data can be used to allocate and retrieve the > function unqiue cache object. unqiue -> unique > > Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com> > Signed-off-by: Yvan Roux <yvan.roux@foss.st.com> > --- > gdb/dwarf2/frame.c | 48 ++++++++++++++++++++++++++++++++++++++++++++++ > gdb/dwarf2/frame.h | 20 +++++++++++++++++-- > 2 files changed, 66 insertions(+), 2 deletions(-) > > diff --git a/gdb/dwarf2/frame.c b/gdb/dwarf2/frame.c > index 3f884abe1d5..bff3b706e7e 100644 > --- a/gdb/dwarf2/frame.c > +++ b/gdb/dwarf2/frame.c > @@ -831,6 +831,14 @@ dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, CORE_ADDR pc, > } > > > +struct dwarf2_frame_fn_data > +{ > + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, > + int regnum); > + void *data; > + struct dwarf2_frame_fn_data* next; > +}; > + I'm wondering if we really need to have a function pointer here. Isn't the cache supposed to be frame-wide and not function-specific? If we don't need it, the cache just becomes an opaque data pointer. > struct dwarf2_frame_cache > { > /* DWARF Call Frame Address. */ > @@ -862,6 +870,8 @@ struct dwarf2_frame_cache > dwarf2_tailcall_frame_unwind unwinder so this field does not apply for > them. */ > void *tailcall_cache; > + > + struct dwarf2_frame_fn_data *fn_data; > }; > > static struct dwarf2_frame_cache * > @@ -1221,6 +1231,44 @@ dwarf2_frame_prev_register (frame_info_ptr this_frame, void **this_cache, > } > } > > +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void **this_cache, > + fn_prev_register fn) > +{ > + struct dwarf2_frame_fn_data *fn_data = nullptr; > + struct dwarf2_frame_cache *cache > + = dwarf2_frame_cache (this_frame, this_cache); > + > + /* Find the object for the function. */ > + for (fn_data = cache->fn_data; fn_data; fn_data = fn_data->next) > + if (fn_data->fn == fn) > + return fn_data->data; > + > + return nullptr; > +} > + > +void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, > + void **this_cache, > + fn_prev_register fn, unsigned long size) > +{ > + struct dwarf2_frame_fn_data *fn_data = nullptr; > + struct dwarf2_frame_cache *cache > + = dwarf2_frame_cache (this_frame, this_cache); > + > + /* First try to find an existing object. */ > + void *data = dwarf2_frame_get_fn_data (this_frame, this_cache, fn); > + if (data) > + return data; > + > + /* No object found, lets create a new instance. */ > + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); > + fn_data->fn = fn; > + fn_data->data = frame_obstack_zalloc (size); > + fn_data->next = cache->fn_data; > + cache->fn_data = fn_data; > + > + return fn_data->data; > +} And if we only have a data pointer, we can return a reference to it through the argument, and then DWARF can cache it. We could even have a destructor/cleanup that can get called once the frames are destroyed. > + > /* Proxy for tailcall_frame_dealloc_cache for bottom frame of a virtual tail > call frames chain. */ > > diff --git a/gdb/dwarf2/frame.h b/gdb/dwarf2/frame.h > index 06c8a10c178..444afd9f8eb 100644 > --- a/gdb/dwarf2/frame.h > +++ b/gdb/dwarf2/frame.h > @@ -66,6 +66,9 @@ enum dwarf2_frame_reg_rule > > /* Register state. */ > > +typedef struct value *(*fn_prev_register) (frame_info_ptr this_frame, > + void **this_cache, int regnum); > + > struct dwarf2_frame_state_reg > { > /* Each register save state can be described in terms of a CFA slot, > @@ -78,8 +81,7 @@ struct dwarf2_frame_state_reg > const gdb_byte *start; > ULONGEST len; > } exp; > - struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, > - int regnum); > + fn_prev_register fn; > } loc; > enum dwarf2_frame_reg_rule how; > }; > @@ -262,4 +264,18 @@ extern int dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, CORE_ADDR pc, > const gdb_byte **cfa_start_out, > const gdb_byte **cfa_end_out); > > + > +/* Allocate a new instance of the function unique data. */ > + > +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, > + void **this_cache, > + fn_prev_register fn, > + unsigned long size); > + > +/* Retrieve the function unique data for this frame. */ > + > +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, > + void **this_cache, > + fn_prev_register fn); > + > #endif /* dwarf2-frame.h */ As we've discussed before, I think the cache idea is nice if we have to deal with targets with multiple CFA's (in our case, we have either 4 SP's or 2 SP's, plus aliases). DWARF doesn't seem to support this at the moment, and the function HOW for DWARF is not smart enough to remember a previously-fetched value. So it seems we have room for some improvement, unless there is enough reason elsewhere about why we shouldn't have a cache. It would be nice to have some opinions from others, so we can potentially shape this in a way that makes it useful for the general case.
Hi, I've had a long discussion with Luis on IRC regarding the points mentioned here, but I'll reply to the list now in order to get more eyes on the topic. On 2022-11-21 22:16, Luis Machado wrote: > Hi, > > On 11/18/22 15:52, Torbjörn SVENSSON wrote: >> When there is no dwarf2 data for a register, a function can be called >> to provide the value of this register. In some situations, it might >> not be trivial to determine the value to return and it would cause a >> performance bottleneck to do the computation each time. >> >> This patch allows the called function to have a "cache" object that it >> can use to store some metadata between calls to reduce the performance >> impact of the complex logic. >> >> The cache object is unique for each function and frame, so if there are >> more than one function pointer stored in the dwarf2_frame_cache->reg >> array, then the appropriate pointer will be supplied (the type is not >> known by the dwarf2 implementation). >> >> dwarf2_frame_get_fn_data can be used to retrieve the function unique >> cache object. >> dwarf2_frame_allocate_fn_data can be used to allocate and retrieve the >> function unqiue cache object. > > unqiue -> unique > >> >> Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com> >> Signed-off-by: Yvan Roux <yvan.roux@foss.st.com> >> --- >> gdb/dwarf2/frame.c | 48 ++++++++++++++++++++++++++++++++++++++++++++++ >> gdb/dwarf2/frame.h | 20 +++++++++++++++++-- >> 2 files changed, 66 insertions(+), 2 deletions(-) >> >> diff --git a/gdb/dwarf2/frame.c b/gdb/dwarf2/frame.c >> index 3f884abe1d5..bff3b706e7e 100644 >> --- a/gdb/dwarf2/frame.c >> +++ b/gdb/dwarf2/frame.c >> @@ -831,6 +831,14 @@ dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, >> CORE_ADDR pc, >> } >> >> +struct dwarf2_frame_fn_data >> +{ >> + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >> + int regnum); >> + void *data; >> + struct dwarf2_frame_fn_data* next; >> +}; >> + > > I'm wondering if we really need to have a function pointer here. Isn't > the cache supposed to be frame-wide and not > function-specific? > > If we don't need it, the cache just becomes an opaque data pointer. > >> struct dwarf2_frame_cache >> { >> /* DWARF Call Frame Address. */ >> @@ -862,6 +870,8 @@ struct dwarf2_frame_cache >> dwarf2_tailcall_frame_unwind unwinder so this field does not >> apply for >> them. */ >> void *tailcall_cache; >> + >> + struct dwarf2_frame_fn_data *fn_data; >> }; >> static struct dwarf2_frame_cache * >> @@ -1221,6 +1231,44 @@ dwarf2_frame_prev_register (frame_info_ptr >> this_frame, void **this_cache, >> } >> } >> +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void >> **this_cache, >> + fn_prev_register fn) >> +{ >> + struct dwarf2_frame_fn_data *fn_data = nullptr; >> + struct dwarf2_frame_cache *cache >> + = dwarf2_frame_cache (this_frame, this_cache); >> + >> + /* Find the object for the function. */ >> + for (fn_data = cache->fn_data; fn_data; fn_data = fn_data->next) >> + if (fn_data->fn == fn) >> + return fn_data->data; >> + >> + return nullptr; >> +} >> + >> +void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >> + void **this_cache, >> + fn_prev_register fn, unsigned long size) >> +{ >> + struct dwarf2_frame_fn_data *fn_data = nullptr; >> + struct dwarf2_frame_cache *cache >> + = dwarf2_frame_cache (this_frame, this_cache); >> + >> + /* First try to find an existing object. */ >> + void *data = dwarf2_frame_get_fn_data (this_frame, this_cache, fn); >> + if (data) >> + return data; >> + >> + /* No object found, lets create a new instance. */ >> + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); >> + fn_data->fn = fn; >> + fn_data->data = frame_obstack_zalloc (size); >> + fn_data->next = cache->fn_data; >> + cache->fn_data = fn_data; >> + >> + return fn_data->data; >> +} > > And if we only have a data pointer, we can return a reference to it > through the argument, and then DWARF can cache it. > > We could even have a destructor/cleanup that can get called once the > frames are destroyed. I don't think we can do that without introducing a lot more changes to the common code. My changes are designed in a way that would only have an impact on arm (as they are the only users for the functionality right now) and not for every target out there that GDB supports. If going for a simpler solution, it would mean that every target needs to be re-tested in order to get the confirmation that the implementation would not break some other target. > >> + >> /* Proxy for tailcall_frame_dealloc_cache for bottom frame of a >> virtual tail >> call frames chain. */ >> diff --git a/gdb/dwarf2/frame.h b/gdb/dwarf2/frame.h >> index 06c8a10c178..444afd9f8eb 100644 >> --- a/gdb/dwarf2/frame.h >> +++ b/gdb/dwarf2/frame.h >> @@ -66,6 +66,9 @@ enum dwarf2_frame_reg_rule >> /* Register state. */ >> +typedef struct value *(*fn_prev_register) (frame_info_ptr this_frame, >> + void **this_cache, int regnum); >> + >> struct dwarf2_frame_state_reg >> { >> /* Each register save state can be described in terms of a CFA slot, >> @@ -78,8 +81,7 @@ struct dwarf2_frame_state_reg >> const gdb_byte *start; >> ULONGEST len; >> } exp; >> - struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >> - int regnum); >> + fn_prev_register fn; >> } loc; >> enum dwarf2_frame_reg_rule how; >> }; >> @@ -262,4 +264,18 @@ extern int dwarf2_fetch_cfa_info (struct gdbarch >> *gdbarch, CORE_ADDR pc, >> const gdb_byte **cfa_start_out, >> const gdb_byte **cfa_end_out); >> + >> +/* Allocate a new instance of the function unique data. */ >> + >> +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >> + void **this_cache, >> + fn_prev_register fn, >> + unsigned long size); >> + >> +/* Retrieve the function unique data for this frame. */ >> + >> +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, >> + void **this_cache, >> + fn_prev_register fn); >> + >> #endif /* dwarf2-frame.h */ > > As we've discussed before, I think the cache idea is nice if we have to > deal with targets with multiple CFA's (in our case, we have either 4 > SP's or 2 SP's, plus aliases). > > DWARF doesn't seem to support this at the moment, and the function HOW > for DWARF is not smart enough to remember a previously-fetched value. So > it seems we have room > for some improvement, unless there is enough reason elsewhere about why > we shouldn't have a cache. This patch does not provide a cache or anything, it just provides a way for the callback function to save some additional data between calls for the same frame. The code above is generic in the way that it has one data object per function and frame. The reason for this implementation is that it's rather easy to ensure that the data object is okay for the function that uses it without any inter-dependencies with some other function that might be called for some other register on the same frame. You could even consider having a shared function to be a callback function. In the case of a shared function, that would mean that the object would be large and public and then it would simply make more sense to make the dwarf2 object public and extend it instead. My approach ensures that each callback function has its own data and the data structure is "private" to the function. It's possible to share the struct for the data object between 2 functions, but it's not possible to share the instance of the struct between 2 functions. > It would be nice to have some opinions from others, so we can > potentially shape this in a way that makes it useful for the general case. Yes. Please give me some more feedback to work on! Kind regards, Torbjörn
Hi Torbjorn, On 29/11/2022 16:19, Torbjorn SVENSSON wrote: > Hi, > > I've had a long discussion with Luis on IRC regarding the points > mentioned here, but I'll reply to the list now in order to get more > eyes on the topic. > > > On 2022-11-21 22:16, Luis Machado wrote: >> Hi, >> >> On 11/18/22 15:52, Torbjörn SVENSSON wrote: >>> When there is no dwarf2 data for a register, a function can be called >>> to provide the value of this register. In some situations, it might >>> not be trivial to determine the value to return and it would cause a >>> performance bottleneck to do the computation each time. >>> >>> This patch allows the called function to have a "cache" object that it >>> can use to store some metadata between calls to reduce the performance >>> impact of the complex logic. >>> >>> The cache object is unique for each function and frame, so if there are >>> more than one function pointer stored in the dwarf2_frame_cache->reg >>> array, then the appropriate pointer will be supplied (the type is not >>> known by the dwarf2 implementation). >>> >>> dwarf2_frame_get_fn_data can be used to retrieve the function unique >>> cache object. >>> dwarf2_frame_allocate_fn_data can be used to allocate and retrieve the >>> function unqiue cache object. >> >> unqiue -> unique >> >>> >>> Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com> >>> Signed-off-by: Yvan Roux <yvan.roux@foss.st.com> >>> --- >>> gdb/dwarf2/frame.c | 48 >>> ++++++++++++++++++++++++++++++++++++++++++++++ >>> gdb/dwarf2/frame.h | 20 +++++++++++++++++-- >>> 2 files changed, 66 insertions(+), 2 deletions(-) >>> >>> diff --git a/gdb/dwarf2/frame.c b/gdb/dwarf2/frame.c >>> index 3f884abe1d5..bff3b706e7e 100644 >>> --- a/gdb/dwarf2/frame.c >>> +++ b/gdb/dwarf2/frame.c >>> @@ -831,6 +831,14 @@ dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, >>> CORE_ADDR pc, >>> } >>> >>> +struct dwarf2_frame_fn_data >>> +{ >>> + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >>> + int regnum); >>> + void *data; >>> + struct dwarf2_frame_fn_data* next; >>> +}; >>> + >> >> I'm wondering if we really need to have a function pointer here. >> Isn't the cache supposed to be frame-wide and not >> function-specific? >> >> If we don't need it, the cache just becomes an opaque data pointer. >> >>> struct dwarf2_frame_cache >>> { >>> /* DWARF Call Frame Address. */ >>> @@ -862,6 +870,8 @@ struct dwarf2_frame_cache >>> dwarf2_tailcall_frame_unwind unwinder so this field does not >>> apply for >>> them. */ >>> void *tailcall_cache; >>> + >>> + struct dwarf2_frame_fn_data *fn_data; >>> }; >>> static struct dwarf2_frame_cache * >>> @@ -1221,6 +1231,44 @@ dwarf2_frame_prev_register (frame_info_ptr >>> this_frame, void **this_cache, >>> } >>> } >>> +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void >>> **this_cache, >>> + fn_prev_register fn) >>> +{ >>> + struct dwarf2_frame_fn_data *fn_data = nullptr; >>> + struct dwarf2_frame_cache *cache >>> + = dwarf2_frame_cache (this_frame, this_cache); >>> + >>> + /* Find the object for the function. */ >>> + for (fn_data = cache->fn_data; fn_data; fn_data = fn_data->next) >>> + if (fn_data->fn == fn) >>> + return fn_data->data; >>> + >>> + return nullptr; >>> +} >>> + >>> +void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >>> + void **this_cache, >>> + fn_prev_register fn, unsigned long size) >>> +{ >>> + struct dwarf2_frame_fn_data *fn_data = nullptr; >>> + struct dwarf2_frame_cache *cache >>> + = dwarf2_frame_cache (this_frame, this_cache); >>> + >>> + /* First try to find an existing object. */ >>> + void *data = dwarf2_frame_get_fn_data (this_frame, this_cache, fn); >>> + if (data) >>> + return data; >>> + >>> + /* No object found, lets create a new instance. */ >>> + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); >>> + fn_data->fn = fn; >>> + fn_data->data = frame_obstack_zalloc (size); >>> + fn_data->next = cache->fn_data; >>> + cache->fn_data = fn_data; >>> + >>> + return fn_data->data; >>> +} >> >> And if we only have a data pointer, we can return a reference to it >> through the argument, and then DWARF can cache it. >> >> We could even have a destructor/cleanup that can get called once the >> frames are destroyed. > > I don't think we can do that without introducing a lot more changes to > the common code. My changes are designed in a way that would only have > an impact on arm (as they are the only users for the functionality > right now) and not for every target out there that GDB supports. If > going for a simpler solution, it would mean that every target needs to > be re-tested in order to get the confirmation that the implementation > would not break some other target. > > >> >>> + >>> /* Proxy for tailcall_frame_dealloc_cache for bottom frame of a >>> virtual tail >>> call frames chain. */ >>> diff --git a/gdb/dwarf2/frame.h b/gdb/dwarf2/frame.h >>> index 06c8a10c178..444afd9f8eb 100644 >>> --- a/gdb/dwarf2/frame.h >>> +++ b/gdb/dwarf2/frame.h >>> @@ -66,6 +66,9 @@ enum dwarf2_frame_reg_rule >>> /* Register state. */ >>> +typedef struct value *(*fn_prev_register) (frame_info_ptr this_frame, >>> + void **this_cache, int regnum); >>> + >>> struct dwarf2_frame_state_reg >>> { >>> /* Each register save state can be described in terms of a CFA >>> slot, >>> @@ -78,8 +81,7 @@ struct dwarf2_frame_state_reg >>> const gdb_byte *start; >>> ULONGEST len; >>> } exp; >>> - struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >>> - int regnum); >>> + fn_prev_register fn; >>> } loc; >>> enum dwarf2_frame_reg_rule how; >>> }; >>> @@ -262,4 +264,18 @@ extern int dwarf2_fetch_cfa_info (struct >>> gdbarch *gdbarch, CORE_ADDR pc, >>> const gdb_byte **cfa_start_out, >>> const gdb_byte **cfa_end_out); >>> + >>> +/* Allocate a new instance of the function unique data. */ >>> + >>> +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >>> + void **this_cache, >>> + fn_prev_register fn, >>> + unsigned long size); >>> + >>> +/* Retrieve the function unique data for this frame. */ >>> + >>> +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, >>> + void **this_cache, >>> + fn_prev_register fn); >>> + >>> #endif /* dwarf2-frame.h */ >> >> As we've discussed before, I think the cache idea is nice if we have >> to deal with targets with multiple CFA's (in our case, we have either >> 4 SP's or 2 SP's, plus aliases). >> >> DWARF doesn't seem to support this at the moment, and the function >> HOW for DWARF is not smart enough to remember a previously-fetched >> value. So it seems we have room >> for some improvement, unless there is enough reason elsewhere about >> why we shouldn't have a cache. > > > This patch does not provide a cache or anything, it just provides a > way for the callback function to save some additional data between > calls for the same frame. > The code above is generic in the way that it has one data object per > function and frame. The reason for this implementation is that it's > rather easy to ensure that the data object is okay for the function > that uses it without any inter-dependencies with some other function > that might be called for some other register on the same frame. You > could even consider having a shared function to be a callback > function. In the case of a shared function, that would mean that the > object would be large and public and then it would simply make more > sense to make the dwarf2 object public and extend it instead. > My approach ensures that each callback function has its own data and > the data structure is "private" to the function. It's possible to > share the struct for the data object between 2 functions, but it's not > possible to share the instance of the struct between 2 functions. Sorry, the per-function-pointers looks like an overkill to me. Maybe I'm just an old school programmer and don't like associative arrays... - frame unwinders use a generic pointer and ensuring the proper type cast is fully in responsibility of the implementation. - we need just to replicate the similar functionality for architecture dependent handling of dwarf2 frames - functions assigned to a dwarf2 frame by how = DWARF2_FRAME_REG_FN are never isolated functions from different parts of code: a gdbarch can set only one initializer by dwarf2_frame_set_init_reg() and it sets all functions - if we ever have more than one function assigned in one dwarf2 frame, seems me likely that all functions would prefer a single cache over isolated ones > > >> It would be nice to have some opinions from others, so we can >> potentially shape this in a way that makes it useful for the general >> case. > > Yes. Please give me some more feedback to work on! > > Kind regards, > Torbjörn regards Tomas
Hi, On 2022-11-29 17:24, Tomas Vanek wrote: > Hi Torbjorn, > > On 29/11/2022 16:19, Torbjorn SVENSSON wrote: >> Hi, >> >> I've had a long discussion with Luis on IRC regarding the points >> mentioned here, but I'll reply to the list now in order to get more >> eyes on the topic. >> >> >> On 2022-11-21 22:16, Luis Machado wrote: >>> Hi, >>> >>> On 11/18/22 15:52, Torbjörn SVENSSON wrote: >>>> When there is no dwarf2 data for a register, a function can be called >>>> to provide the value of this register. In some situations, it might >>>> not be trivial to determine the value to return and it would cause a >>>> performance bottleneck to do the computation each time. >>>> >>>> This patch allows the called function to have a "cache" object that it >>>> can use to store some metadata between calls to reduce the performance >>>> impact of the complex logic. >>>> >>>> The cache object is unique for each function and frame, so if there are >>>> more than one function pointer stored in the dwarf2_frame_cache->reg >>>> array, then the appropriate pointer will be supplied (the type is not >>>> known by the dwarf2 implementation). >>>> >>>> dwarf2_frame_get_fn_data can be used to retrieve the function unique >>>> cache object. >>>> dwarf2_frame_allocate_fn_data can be used to allocate and retrieve the >>>> function unqiue cache object. >>> >>> unqiue -> unique >>> >>>> >>>> Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com> >>>> Signed-off-by: Yvan Roux <yvan.roux@foss.st.com> >>>> --- >>>> gdb/dwarf2/frame.c | 48 >>>> ++++++++++++++++++++++++++++++++++++++++++++++ >>>> gdb/dwarf2/frame.h | 20 +++++++++++++++++-- >>>> 2 files changed, 66 insertions(+), 2 deletions(-) >>>> >>>> diff --git a/gdb/dwarf2/frame.c b/gdb/dwarf2/frame.c >>>> index 3f884abe1d5..bff3b706e7e 100644 >>>> --- a/gdb/dwarf2/frame.c >>>> +++ b/gdb/dwarf2/frame.c >>>> @@ -831,6 +831,14 @@ dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, >>>> CORE_ADDR pc, >>>> } >>>> >>>> +struct dwarf2_frame_fn_data >>>> +{ >>>> + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >>>> + int regnum); >>>> + void *data; >>>> + struct dwarf2_frame_fn_data* next; >>>> +}; >>>> + >>> >>> I'm wondering if we really need to have a function pointer here. >>> Isn't the cache supposed to be frame-wide and not >>> function-specific? >>> >>> If we don't need it, the cache just becomes an opaque data pointer. >>> >>>> struct dwarf2_frame_cache >>>> { >>>> /* DWARF Call Frame Address. */ >>>> @@ -862,6 +870,8 @@ struct dwarf2_frame_cache >>>> dwarf2_tailcall_frame_unwind unwinder so this field does not >>>> apply for >>>> them. */ >>>> void *tailcall_cache; >>>> + >>>> + struct dwarf2_frame_fn_data *fn_data; >>>> }; >>>> static struct dwarf2_frame_cache * >>>> @@ -1221,6 +1231,44 @@ dwarf2_frame_prev_register (frame_info_ptr >>>> this_frame, void **this_cache, >>>> } >>>> } >>>> +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void >>>> **this_cache, >>>> + fn_prev_register fn) >>>> +{ >>>> + struct dwarf2_frame_fn_data *fn_data = nullptr; >>>> + struct dwarf2_frame_cache *cache >>>> + = dwarf2_frame_cache (this_frame, this_cache); >>>> + >>>> + /* Find the object for the function. */ >>>> + for (fn_data = cache->fn_data; fn_data; fn_data = fn_data->next) >>>> + if (fn_data->fn == fn) >>>> + return fn_data->data; >>>> + >>>> + return nullptr; >>>> +} >>>> + >>>> +void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >>>> + void **this_cache, >>>> + fn_prev_register fn, unsigned long size) >>>> +{ >>>> + struct dwarf2_frame_fn_data *fn_data = nullptr; >>>> + struct dwarf2_frame_cache *cache >>>> + = dwarf2_frame_cache (this_frame, this_cache); >>>> + >>>> + /* First try to find an existing object. */ >>>> + void *data = dwarf2_frame_get_fn_data (this_frame, this_cache, fn); >>>> + if (data) >>>> + return data; >>>> + >>>> + /* No object found, lets create a new instance. */ >>>> + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); >>>> + fn_data->fn = fn; >>>> + fn_data->data = frame_obstack_zalloc (size); >>>> + fn_data->next = cache->fn_data; >>>> + cache->fn_data = fn_data; >>>> + >>>> + return fn_data->data; >>>> +} >>> >>> And if we only have a data pointer, we can return a reference to it >>> through the argument, and then DWARF can cache it. >>> >>> We could even have a destructor/cleanup that can get called once the >>> frames are destroyed. >> >> I don't think we can do that without introducing a lot more changes to >> the common code. My changes are designed in a way that would only have >> an impact on arm (as they are the only users for the functionality >> right now) and not for every target out there that GDB supports. If >> going for a simpler solution, it would mean that every target needs to >> be re-tested in order to get the confirmation that the implementation >> would not break some other target. >> >> >>> >>>> + >>>> /* Proxy for tailcall_frame_dealloc_cache for bottom frame of a >>>> virtual tail >>>> call frames chain. */ >>>> diff --git a/gdb/dwarf2/frame.h b/gdb/dwarf2/frame.h >>>> index 06c8a10c178..444afd9f8eb 100644 >>>> --- a/gdb/dwarf2/frame.h >>>> +++ b/gdb/dwarf2/frame.h >>>> @@ -66,6 +66,9 @@ enum dwarf2_frame_reg_rule >>>> /* Register state. */ >>>> +typedef struct value *(*fn_prev_register) (frame_info_ptr this_frame, >>>> + void **this_cache, int regnum); >>>> + >>>> struct dwarf2_frame_state_reg >>>> { >>>> /* Each register save state can be described in terms of a CFA >>>> slot, >>>> @@ -78,8 +81,7 @@ struct dwarf2_frame_state_reg >>>> const gdb_byte *start; >>>> ULONGEST len; >>>> } exp; >>>> - struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >>>> - int regnum); >>>> + fn_prev_register fn; >>>> } loc; >>>> enum dwarf2_frame_reg_rule how; >>>> }; >>>> @@ -262,4 +264,18 @@ extern int dwarf2_fetch_cfa_info (struct >>>> gdbarch *gdbarch, CORE_ADDR pc, >>>> const gdb_byte **cfa_start_out, >>>> const gdb_byte **cfa_end_out); >>>> + >>>> +/* Allocate a new instance of the function unique data. */ >>>> + >>>> +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >>>> + void **this_cache, >>>> + fn_prev_register fn, >>>> + unsigned long size); >>>> + >>>> +/* Retrieve the function unique data for this frame. */ >>>> + >>>> +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, >>>> + void **this_cache, >>>> + fn_prev_register fn); >>>> + >>>> #endif /* dwarf2-frame.h */ >>> >>> As we've discussed before, I think the cache idea is nice if we have >>> to deal with targets with multiple CFA's (in our case, we have either >>> 4 SP's or 2 SP's, plus aliases). >>> >>> DWARF doesn't seem to support this at the moment, and the function >>> HOW for DWARF is not smart enough to remember a previously-fetched >>> value. So it seems we have room >>> for some improvement, unless there is enough reason elsewhere about >>> why we shouldn't have a cache. >> >> >> This patch does not provide a cache or anything, it just provides a >> way for the callback function to save some additional data between >> calls for the same frame. >> The code above is generic in the way that it has one data object per >> function and frame. The reason for this implementation is that it's >> rather easy to ensure that the data object is okay for the function >> that uses it without any inter-dependencies with some other function >> that might be called for some other register on the same frame. You >> could even consider having a shared function to be a callback >> function. In the case of a shared function, that would mean that the >> object would be large and public and then it would simply make more >> sense to make the dwarf2 object public and extend it instead. >> My approach ensures that each callback function has its own data and >> the data structure is "private" to the function. It's possible to >> share the struct for the data object between 2 functions, but it's not >> possible to share the instance of the struct between 2 functions. > > Sorry, the per-function-pointers looks like an overkill to me. > Maybe I'm just an old school programmer and don't like associative > arrays... > - frame unwinders use a generic pointer and ensuring the proper type > cast is fully in responsibility of the implementation. > - we need just to replicate the similar functionality for architecture > dependent handling of dwarf2 frames > - functions assigned to a dwarf2 frame by how = DWARF2_FRAME_REG_FN are > never isolated functions from different parts > of code: a gdbarch can set only one initializer by > dwarf2_frame_set_init_reg() and it sets all functions > - if we ever have more than one function assigned in one dwarf2 frame, > seems me likely that all functions would prefer a single cache over > isolated ones Based on the points above, can please answer the below questions? a) Why is there a function pointer per register if it's always going to be a single one? It would have made more sense to simply have a single bool telling if the function should be called or not if your suggestion would make sense. b) If you want a shared cache object, then why is all the cache types private to the compile unit? You can't have something shared across 2 compile units that are just defined in the .c file. Don't get me wrong, I'm not against defining the cache types in the public namespace. By doing so, it will have a large impact on the GDB code base and it will require a lot more testing than the change I'm proposing. - Is it work the extra risk? - Who can actually do the testing on all the targets to make sure that nothing broke? Kind regards, Torbjörn > >> >> >>> It would be nice to have some opinions from others, so we can >>> potentially shape this in a way that makes it useful for the general >>> case. >> >> Yes. Please give me some more feedback to work on! >> >> Kind regards, >> Torbjörn > > regards > Tomas
On 11/30/22 10:16, Torbjorn SVENSSON wrote: > Hi, > > On 2022-11-29 17:24, Tomas Vanek wrote: >> Hi Torbjorn, >> >> On 29/11/2022 16:19, Torbjorn SVENSSON wrote: >>> Hi, >>> >>> I've had a long discussion with Luis on IRC regarding the points mentioned here, but I'll reply to the list now in order to get more eyes on the topic. >>> >>> >>> On 2022-11-21 22:16, Luis Machado wrote: >>>> Hi, >>>> >>>> On 11/18/22 15:52, Torbjörn SVENSSON wrote: >>>>> When there is no dwarf2 data for a register, a function can be called >>>>> to provide the value of this register. In some situations, it might >>>>> not be trivial to determine the value to return and it would cause a >>>>> performance bottleneck to do the computation each time. >>>>> >>>>> This patch allows the called function to have a "cache" object that it >>>>> can use to store some metadata between calls to reduce the performance >>>>> impact of the complex logic. >>>>> >>>>> The cache object is unique for each function and frame, so if there are >>>>> more than one function pointer stored in the dwarf2_frame_cache->reg >>>>> array, then the appropriate pointer will be supplied (the type is not >>>>> known by the dwarf2 implementation). >>>>> >>>>> dwarf2_frame_get_fn_data can be used to retrieve the function unique >>>>> cache object. >>>>> dwarf2_frame_allocate_fn_data can be used to allocate and retrieve the >>>>> function unqiue cache object. >>>> >>>> unqiue -> unique >>>> >>>>> >>>>> Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com> >>>>> Signed-off-by: Yvan Roux <yvan.roux@foss.st.com> >>>>> --- >>>>> gdb/dwarf2/frame.c | 48 ++++++++++++++++++++++++++++++++++++++++++++++ >>>>> gdb/dwarf2/frame.h | 20 +++++++++++++++++-- >>>>> 2 files changed, 66 insertions(+), 2 deletions(-) >>>>> >>>>> diff --git a/gdb/dwarf2/frame.c b/gdb/dwarf2/frame.c >>>>> index 3f884abe1d5..bff3b706e7e 100644 >>>>> --- a/gdb/dwarf2/frame.c >>>>> +++ b/gdb/dwarf2/frame.c >>>>> @@ -831,6 +831,14 @@ dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, CORE_ADDR pc, >>>>> } >>>>> >>>>> +struct dwarf2_frame_fn_data >>>>> +{ >>>>> + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >>>>> + int regnum); >>>>> + void *data; >>>>> + struct dwarf2_frame_fn_data* next; >>>>> +}; >>>>> + >>>> >>>> I'm wondering if we really need to have a function pointer here. Isn't the cache supposed to be frame-wide and not >>>> function-specific? >>>> >>>> If we don't need it, the cache just becomes an opaque data pointer. >>>> >>>>> struct dwarf2_frame_cache >>>>> { >>>>> /* DWARF Call Frame Address. */ >>>>> @@ -862,6 +870,8 @@ struct dwarf2_frame_cache >>>>> dwarf2_tailcall_frame_unwind unwinder so this field does not apply for >>>>> them. */ >>>>> void *tailcall_cache; >>>>> + >>>>> + struct dwarf2_frame_fn_data *fn_data; >>>>> }; >>>>> static struct dwarf2_frame_cache * >>>>> @@ -1221,6 +1231,44 @@ dwarf2_frame_prev_register (frame_info_ptr this_frame, void **this_cache, >>>>> } >>>>> } >>>>> +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void **this_cache, >>>>> + fn_prev_register fn) >>>>> +{ >>>>> + struct dwarf2_frame_fn_data *fn_data = nullptr; >>>>> + struct dwarf2_frame_cache *cache >>>>> + = dwarf2_frame_cache (this_frame, this_cache); >>>>> + >>>>> + /* Find the object for the function. */ >>>>> + for (fn_data = cache->fn_data; fn_data; fn_data = fn_data->next) >>>>> + if (fn_data->fn == fn) >>>>> + return fn_data->data; >>>>> + >>>>> + return nullptr; >>>>> +} >>>>> + >>>>> +void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >>>>> + void **this_cache, >>>>> + fn_prev_register fn, unsigned long size) >>>>> +{ >>>>> + struct dwarf2_frame_fn_data *fn_data = nullptr; >>>>> + struct dwarf2_frame_cache *cache >>>>> + = dwarf2_frame_cache (this_frame, this_cache); >>>>> + >>>>> + /* First try to find an existing object. */ >>>>> + void *data = dwarf2_frame_get_fn_data (this_frame, this_cache, fn); >>>>> + if (data) >>>>> + return data; >>>>> + >>>>> + /* No object found, lets create a new instance. */ >>>>> + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); >>>>> + fn_data->fn = fn; >>>>> + fn_data->data = frame_obstack_zalloc (size); >>>>> + fn_data->next = cache->fn_data; >>>>> + cache->fn_data = fn_data; >>>>> + >>>>> + return fn_data->data; >>>>> +} >>>> >>>> And if we only have a data pointer, we can return a reference to it through the argument, and then DWARF can cache it. >>>> >>>> We could even have a destructor/cleanup that can get called once the frames are destroyed. >>> >>> I don't think we can do that without introducing a lot more changes to the common code. My changes are designed in a way that would only have an impact on arm (as they are the only users for the functionality right now) and not for every target out there that GDB supports. If going for a simpler solution, it would mean that every target needs to be re-tested in order to get the confirmation that the implementation would not break some other target. >>> >>> >>>> >>>>> + >>>>> /* Proxy for tailcall_frame_dealloc_cache for bottom frame of a virtual tail >>>>> call frames chain. */ >>>>> diff --git a/gdb/dwarf2/frame.h b/gdb/dwarf2/frame.h >>>>> index 06c8a10c178..444afd9f8eb 100644 >>>>> --- a/gdb/dwarf2/frame.h >>>>> +++ b/gdb/dwarf2/frame.h >>>>> @@ -66,6 +66,9 @@ enum dwarf2_frame_reg_rule >>>>> /* Register state. */ >>>>> +typedef struct value *(*fn_prev_register) (frame_info_ptr this_frame, >>>>> + void **this_cache, int regnum); >>>>> + >>>>> struct dwarf2_frame_state_reg >>>>> { >>>>> /* Each register save state can be described in terms of a CFA slot, >>>>> @@ -78,8 +81,7 @@ struct dwarf2_frame_state_reg >>>>> const gdb_byte *start; >>>>> ULONGEST len; >>>>> } exp; >>>>> - struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >>>>> - int regnum); >>>>> + fn_prev_register fn; >>>>> } loc; >>>>> enum dwarf2_frame_reg_rule how; >>>>> }; >>>>> @@ -262,4 +264,18 @@ extern int dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, CORE_ADDR pc, >>>>> const gdb_byte **cfa_start_out, >>>>> const gdb_byte **cfa_end_out); >>>>> + >>>>> +/* Allocate a new instance of the function unique data. */ >>>>> + >>>>> +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >>>>> + void **this_cache, >>>>> + fn_prev_register fn, >>>>> + unsigned long size); >>>>> + >>>>> +/* Retrieve the function unique data for this frame. */ >>>>> + >>>>> +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, >>>>> + void **this_cache, >>>>> + fn_prev_register fn); >>>>> + >>>>> #endif /* dwarf2-frame.h */ >>>> >>>> As we've discussed before, I think the cache idea is nice if we have to deal with targets with multiple CFA's (in our case, we have either 4 SP's or 2 SP's, plus aliases). >>>> >>>> DWARF doesn't seem to support this at the moment, and the function HOW for DWARF is not smart enough to remember a previously-fetched value. So it seems we have room >>>> for some improvement, unless there is enough reason elsewhere about why we shouldn't have a cache. >>> >>> >>> This patch does not provide a cache or anything, it just provides a way for the callback function to save some additional data between calls for the same frame. >>> The code above is generic in the way that it has one data object per function and frame. The reason for this implementation is that it's rather easy to ensure that the data object is okay for the function that uses it without any inter-dependencies with some other function that might be called for some other register on the same frame. You could even consider having a shared function to be a callback function. In the case of a shared function, that would mean that the object would be large and public and then it would simply make more sense to make the dwarf2 object public and extend it instead. >>> My approach ensures that each callback function has its own data and the data structure is "private" to the function. It's possible to share the struct for the data object between 2 functions, but it's not possible to share the instance of the struct between 2 functions. >> >> Sorry, the per-function-pointers looks like an overkill to me. >> Maybe I'm just an old school programmer and don't like associative arrays... >> - frame unwinders use a generic pointer and ensuring the proper type cast is fully in responsibility of the implementation. >> - we need just to replicate the similar functionality for architecture dependent handling of dwarf2 frames >> - functions assigned to a dwarf2 frame by how = DWARF2_FRAME_REG_FN are never isolated functions from different parts >> of code: a gdbarch can set only one initializer by dwarf2_frame_set_init_reg() and it sets all functions >> - if we ever have more than one function assigned in one dwarf2 frame, seems me likely that all functions would prefer a single cache over isolated ones > > > Based on the points above, can please answer the below questions? > > a) Why is there a function pointer per register if it's always going to be a single one? It would have made more sense to simply have a single bool telling if the function should be called or not if your suggestion would make sense. > > b) If you want a shared cache object, then why is all the cache types private to the compile unit? You can't have something shared across 2 compile units that are just defined in the .c file. > > > Don't get me wrong, I'm not against defining the cache types in the public namespace. By doing so, it will have a large impact on the GDB code base and it will require a lot more testing than the change I'm proposing. > - Is it work the extra risk? > - Who can actually do the testing on all the targets to make sure that nothing broke? > What I see is an area of gdb where we can make some general improvements to help all targets. Testing wouldn't be that much more complicated, specially if we create a new HOW and leave the old one as-is. > Kind regards, > Torbjörn > > >> >>> >>> >>>> It would be nice to have some opinions from others, so we can potentially shape this in a way that makes it useful for the general case. >>> >>> Yes. Please give me some more feedback to work on! >>> >>> Kind regards, >>> Torbjörn >> >> regards >> Tomas
On 11/30/22 10:16, Torbjorn SVENSSON wrote: > Hi, > > On 2022-11-29 17:24, Tomas Vanek wrote: >> Hi Torbjorn, >> >> On 29/11/2022 16:19, Torbjorn SVENSSON wrote: >>> Hi, >>> >>> I've had a long discussion with Luis on IRC regarding the points mentioned here, but I'll reply to the list now in order to get more eyes on the topic. >>> >>> >>> On 2022-11-21 22:16, Luis Machado wrote: >>>> Hi, >>>> >>>> On 11/18/22 15:52, Torbjörn SVENSSON wrote: >>>>> When there is no dwarf2 data for a register, a function can be called >>>>> to provide the value of this register. In some situations, it might >>>>> not be trivial to determine the value to return and it would cause a >>>>> performance bottleneck to do the computation each time. >>>>> >>>>> This patch allows the called function to have a "cache" object that it >>>>> can use to store some metadata between calls to reduce the performance >>>>> impact of the complex logic. >>>>> >>>>> The cache object is unique for each function and frame, so if there are >>>>> more than one function pointer stored in the dwarf2_frame_cache->reg >>>>> array, then the appropriate pointer will be supplied (the type is not >>>>> known by the dwarf2 implementation). >>>>> >>>>> dwarf2_frame_get_fn_data can be used to retrieve the function unique >>>>> cache object. >>>>> dwarf2_frame_allocate_fn_data can be used to allocate and retrieve the >>>>> function unqiue cache object. >>>> >>>> unqiue -> unique >>>> >>>>> >>>>> Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com> >>>>> Signed-off-by: Yvan Roux <yvan.roux@foss.st.com> >>>>> --- >>>>> gdb/dwarf2/frame.c | 48 ++++++++++++++++++++++++++++++++++++++++++++++ >>>>> gdb/dwarf2/frame.h | 20 +++++++++++++++++-- >>>>> 2 files changed, 66 insertions(+), 2 deletions(-) >>>>> >>>>> diff --git a/gdb/dwarf2/frame.c b/gdb/dwarf2/frame.c >>>>> index 3f884abe1d5..bff3b706e7e 100644 >>>>> --- a/gdb/dwarf2/frame.c >>>>> +++ b/gdb/dwarf2/frame.c >>>>> @@ -831,6 +831,14 @@ dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, CORE_ADDR pc, >>>>> } >>>>> >>>>> +struct dwarf2_frame_fn_data >>>>> +{ >>>>> + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >>>>> + int regnum); >>>>> + void *data; >>>>> + struct dwarf2_frame_fn_data* next; >>>>> +}; >>>>> + >>>> >>>> I'm wondering if we really need to have a function pointer here. Isn't the cache supposed to be frame-wide and not >>>> function-specific? >>>> >>>> If we don't need it, the cache just becomes an opaque data pointer. >>>> >>>>> struct dwarf2_frame_cache >>>>> { >>>>> /* DWARF Call Frame Address. */ >>>>> @@ -862,6 +870,8 @@ struct dwarf2_frame_cache >>>>> dwarf2_tailcall_frame_unwind unwinder so this field does not apply for >>>>> them. */ >>>>> void *tailcall_cache; >>>>> + >>>>> + struct dwarf2_frame_fn_data *fn_data; >>>>> }; >>>>> static struct dwarf2_frame_cache * >>>>> @@ -1221,6 +1231,44 @@ dwarf2_frame_prev_register (frame_info_ptr this_frame, void **this_cache, >>>>> } >>>>> } >>>>> +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void **this_cache, >>>>> + fn_prev_register fn) >>>>> +{ >>>>> + struct dwarf2_frame_fn_data *fn_data = nullptr; >>>>> + struct dwarf2_frame_cache *cache >>>>> + = dwarf2_frame_cache (this_frame, this_cache); >>>>> + >>>>> + /* Find the object for the function. */ >>>>> + for (fn_data = cache->fn_data; fn_data; fn_data = fn_data->next) >>>>> + if (fn_data->fn == fn) >>>>> + return fn_data->data; >>>>> + >>>>> + return nullptr; >>>>> +} >>>>> + >>>>> +void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >>>>> + void **this_cache, >>>>> + fn_prev_register fn, unsigned long size) >>>>> +{ >>>>> + struct dwarf2_frame_fn_data *fn_data = nullptr; >>>>> + struct dwarf2_frame_cache *cache >>>>> + = dwarf2_frame_cache (this_frame, this_cache); >>>>> + >>>>> + /* First try to find an existing object. */ >>>>> + void *data = dwarf2_frame_get_fn_data (this_frame, this_cache, fn); >>>>> + if (data) >>>>> + return data; >>>>> + >>>>> + /* No object found, lets create a new instance. */ >>>>> + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); >>>>> + fn_data->fn = fn; >>>>> + fn_data->data = frame_obstack_zalloc (size); >>>>> + fn_data->next = cache->fn_data; >>>>> + cache->fn_data = fn_data; >>>>> + >>>>> + return fn_data->data; >>>>> +} >>>> >>>> And if we only have a data pointer, we can return a reference to it through the argument, and then DWARF can cache it. >>>> >>>> We could even have a destructor/cleanup that can get called once the frames are destroyed. >>> >>> I don't think we can do that without introducing a lot more changes to the common code. My changes are designed in a way that would only have an impact on arm (as they are the only users for the functionality right now) and not for every target out there that GDB supports. If going for a simpler solution, it would mean that every target needs to be re-tested in order to get the confirmation that the implementation would not break some other target. >>> >>> >>>> >>>>> + >>>>> /* Proxy for tailcall_frame_dealloc_cache for bottom frame of a virtual tail >>>>> call frames chain. */ >>>>> diff --git a/gdb/dwarf2/frame.h b/gdb/dwarf2/frame.h >>>>> index 06c8a10c178..444afd9f8eb 100644 >>>>> --- a/gdb/dwarf2/frame.h >>>>> +++ b/gdb/dwarf2/frame.h >>>>> @@ -66,6 +66,9 @@ enum dwarf2_frame_reg_rule >>>>> /* Register state. */ >>>>> +typedef struct value *(*fn_prev_register) (frame_info_ptr this_frame, >>>>> + void **this_cache, int regnum); >>>>> + >>>>> struct dwarf2_frame_state_reg >>>>> { >>>>> /* Each register save state can be described in terms of a CFA slot, >>>>> @@ -78,8 +81,7 @@ struct dwarf2_frame_state_reg >>>>> const gdb_byte *start; >>>>> ULONGEST len; >>>>> } exp; >>>>> - struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >>>>> - int regnum); >>>>> + fn_prev_register fn; >>>>> } loc; >>>>> enum dwarf2_frame_reg_rule how; >>>>> }; >>>>> @@ -262,4 +264,18 @@ extern int dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, CORE_ADDR pc, >>>>> const gdb_byte **cfa_start_out, >>>>> const gdb_byte **cfa_end_out); >>>>> + >>>>> +/* Allocate a new instance of the function unique data. */ >>>>> + >>>>> +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >>>>> + void **this_cache, >>>>> + fn_prev_register fn, >>>>> + unsigned long size); >>>>> + >>>>> +/* Retrieve the function unique data for this frame. */ >>>>> + >>>>> +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, >>>>> + void **this_cache, >>>>> + fn_prev_register fn); >>>>> + >>>>> #endif /* dwarf2-frame.h */ >>>> >>>> As we've discussed before, I think the cache idea is nice if we have to deal with targets with multiple CFA's (in our case, we have either 4 SP's or 2 SP's, plus aliases). >>>> >>>> DWARF doesn't seem to support this at the moment, and the function HOW for DWARF is not smart enough to remember a previously-fetched value. So it seems we have room >>>> for some improvement, unless there is enough reason elsewhere about why we shouldn't have a cache. >>> >>> >>> This patch does not provide a cache or anything, it just provides a way for the callback function to save some additional data between calls for the same frame. >>> The code above is generic in the way that it has one data object per function and frame. The reason for this implementation is that it's rather easy to ensure that the data object is okay for the function that uses it without any inter-dependencies with some other function that might be called for some other register on the same frame. You could even consider having a shared function to be a callback function. In the case of a shared function, that would mean that the object would be large and public and then it would simply make more sense to make the dwarf2 object public and extend it instead. >>> My approach ensures that each callback function has its own data and the data structure is "private" to the function. It's possible to share the struct for the data object between 2 functions, but it's not possible to share the instance of the struct between 2 functions. >> >> Sorry, the per-function-pointers looks like an overkill to me. >> Maybe I'm just an old school programmer and don't like associative arrays... >> - frame unwinders use a generic pointer and ensuring the proper type cast is fully in responsibility of the implementation. >> - we need just to replicate the similar functionality for architecture dependent handling of dwarf2 frames >> - functions assigned to a dwarf2 frame by how = DWARF2_FRAME_REG_FN are never isolated functions from different parts >> of code: a gdbarch can set only one initializer by dwarf2_frame_set_init_reg() and it sets all functions >> - if we ever have more than one function assigned in one dwarf2 frame, seems me likely that all functions would prefer a single cache over isolated ones > > > Based on the points above, can please answer the below questions? I'll try to clarify some of them from my personal perspective. > > a) Why is there a function pointer per register if it's always going to be a single one? It would have made more sense to simply have a single bool telling if the function should be called or not if your suggestion would make sense. That's correct. Each register (dwarf register column) gets to define a HOW, so you can have different functions for different registers. It may not be useful to have multiple different functions for this purpose though. > > b) If you want a shared cache object, then why is all the cache types private to the compile unit? You can't have something shared across 2 compile units that are just defined in the .c file. There is probably some historic reason for doing it that way. My guess is that we want to isolate the dwarf unwinding machinery as much as possible from the rest of the code. Having other code influencing the dwarf unwinding logic may not be desirable. But then we have ways to do it, like the callback HOW. With that said, I agree with Tomas that we should have a single data pointer/cache per frame. If we need to touch a bit more generic code, that is fine as long as it is done in a flexible future-proof way. We already have a good mechanism for storing register values, and that's the trad_frame_saved_reg structure that we use for most of the prologue unwinder/analyzer code. That structure allocates things on the frame obstack, so memory release is done automatically when the frame data gets flushed. A getter function is needed so we can return the trad_frame_saved_reg array contained in the opaque (to the arch-specific code) dwarf2_frame_cache struct. With the trad_frame_saved_reg array in hand, the arch-specific dwarf2 unwinder can cache some useful data for registers that don't have dwarf id's. > > > Don't get me wrong, I'm not against defining the cache types in the public namespace. By doing so, it will have a large impact on the GDB code base and it will require a lot more testing than the change I'm proposing. If we define a new HOW for this, the changes will be easier to test. Code using the non-cached version of the callback HOW will behave the same as before. We could potentially have a single callback HOW and teach gdb how to detect that an arch-specific cache is being used. But I think that is beyond the scope of fixing the performance issue here. > - Is it work the extra risk? > - Who can actually do the testing on all the targets to make sure that nothing broke? > > Kind regards, > Torbjörn > > >> >>> >>> >>>> It would be nice to have some opinions from others, so we can potentially shape this in a way that makes it useful for the general case. >>> >>> Yes. Please give me some more feedback to work on! >>> >>> Kind regards, >>> Torbjörn >> >> regards >> Tomas
Ping! :) Can someone (Simon?, Andrew?, Pedro?, Joel?) take a look at the proposed patch? I would like to conclude something for the GDB13 release... Kind regards, Torbjörn On 2022-11-21 22:16, Luis Machado wrote: > Hi, > > On 11/18/22 15:52, Torbjörn SVENSSON wrote: >> When there is no dwarf2 data for a register, a function can be called >> to provide the value of this register. In some situations, it might >> not be trivial to determine the value to return and it would cause a >> performance bottleneck to do the computation each time. >> >> This patch allows the called function to have a "cache" object that it >> can use to store some metadata between calls to reduce the performance >> impact of the complex logic. >> >> The cache object is unique for each function and frame, so if there are >> more than one function pointer stored in the dwarf2_frame_cache->reg >> array, then the appropriate pointer will be supplied (the type is not >> known by the dwarf2 implementation). >> >> dwarf2_frame_get_fn_data can be used to retrieve the function unique >> cache object. >> dwarf2_frame_allocate_fn_data can be used to allocate and retrieve the >> function unqiue cache object. > > unqiue -> unique > >> >> Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com> >> Signed-off-by: Yvan Roux <yvan.roux@foss.st.com> >> --- >> gdb/dwarf2/frame.c | 48 ++++++++++++++++++++++++++++++++++++++++++++++ >> gdb/dwarf2/frame.h | 20 +++++++++++++++++-- >> 2 files changed, 66 insertions(+), 2 deletions(-) >> >> diff --git a/gdb/dwarf2/frame.c b/gdb/dwarf2/frame.c >> index 3f884abe1d5..bff3b706e7e 100644 >> --- a/gdb/dwarf2/frame.c >> +++ b/gdb/dwarf2/frame.c >> @@ -831,6 +831,14 @@ dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, >> CORE_ADDR pc, >> } >> >> +struct dwarf2_frame_fn_data >> +{ >> + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >> + int regnum); >> + void *data; >> + struct dwarf2_frame_fn_data* next; >> +}; >> + > > I'm wondering if we really need to have a function pointer here. Isn't > the cache supposed to be frame-wide and not > function-specific? > > If we don't need it, the cache just becomes an opaque data pointer. > >> struct dwarf2_frame_cache >> { >> /* DWARF Call Frame Address. */ >> @@ -862,6 +870,8 @@ struct dwarf2_frame_cache >> dwarf2_tailcall_frame_unwind unwinder so this field does not >> apply for >> them. */ >> void *tailcall_cache; >> + >> + struct dwarf2_frame_fn_data *fn_data; >> }; >> static struct dwarf2_frame_cache * >> @@ -1221,6 +1231,44 @@ dwarf2_frame_prev_register (frame_info_ptr >> this_frame, void **this_cache, >> } >> } >> +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void >> **this_cache, >> + fn_prev_register fn) >> +{ >> + struct dwarf2_frame_fn_data *fn_data = nullptr; >> + struct dwarf2_frame_cache *cache >> + = dwarf2_frame_cache (this_frame, this_cache); >> + >> + /* Find the object for the function. */ >> + for (fn_data = cache->fn_data; fn_data; fn_data = fn_data->next) >> + if (fn_data->fn == fn) >> + return fn_data->data; >> + >> + return nullptr; >> +} >> + >> +void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >> + void **this_cache, >> + fn_prev_register fn, unsigned long size) >> +{ >> + struct dwarf2_frame_fn_data *fn_data = nullptr; >> + struct dwarf2_frame_cache *cache >> + = dwarf2_frame_cache (this_frame, this_cache); >> + >> + /* First try to find an existing object. */ >> + void *data = dwarf2_frame_get_fn_data (this_frame, this_cache, fn); >> + if (data) >> + return data; >> + >> + /* No object found, lets create a new instance. */ >> + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); >> + fn_data->fn = fn; >> + fn_data->data = frame_obstack_zalloc (size); >> + fn_data->next = cache->fn_data; >> + cache->fn_data = fn_data; >> + >> + return fn_data->data; >> +} > > And if we only have a data pointer, we can return a reference to it > through the argument, and then DWARF can cache it. > > We could even have a destructor/cleanup that can get called once the > frames are destroyed. > >> + >> /* Proxy for tailcall_frame_dealloc_cache for bottom frame of a >> virtual tail >> call frames chain. */ >> diff --git a/gdb/dwarf2/frame.h b/gdb/dwarf2/frame.h >> index 06c8a10c178..444afd9f8eb 100644 >> --- a/gdb/dwarf2/frame.h >> +++ b/gdb/dwarf2/frame.h >> @@ -66,6 +66,9 @@ enum dwarf2_frame_reg_rule >> /* Register state. */ >> +typedef struct value *(*fn_prev_register) (frame_info_ptr this_frame, >> + void **this_cache, int regnum); >> + >> struct dwarf2_frame_state_reg >> { >> /* Each register save state can be described in terms of a CFA slot, >> @@ -78,8 +81,7 @@ struct dwarf2_frame_state_reg >> const gdb_byte *start; >> ULONGEST len; >> } exp; >> - struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >> - int regnum); >> + fn_prev_register fn; >> } loc; >> enum dwarf2_frame_reg_rule how; >> }; >> @@ -262,4 +264,18 @@ extern int dwarf2_fetch_cfa_info (struct gdbarch >> *gdbarch, CORE_ADDR pc, >> const gdb_byte **cfa_start_out, >> const gdb_byte **cfa_end_out); >> + >> +/* Allocate a new instance of the function unique data. */ >> + >> +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >> + void **this_cache, >> + fn_prev_register fn, >> + unsigned long size); >> + >> +/* Retrieve the function unique data for this frame. */ >> + >> +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, >> + void **this_cache, >> + fn_prev_register fn); >> + >> #endif /* dwarf2-frame.h */ > > As we've discussed before, I think the cache idea is nice if we have to > deal with targets with multiple CFA's (in our case, we have either 4 > SP's or 2 SP's, plus aliases). > > DWARF doesn't seem to support this at the moment, and the function HOW > for DWARF is not smart enough to remember a previously-fetched value. So > it seems we have room > for some improvement, unless there is enough reason elsewhere about why > we shouldn't have a cache. > > It would be nice to have some opinions from others, so we can > potentially shape this in a way that makes it useful for the general case.
>>>>> Torbjörn SVENSSON via Gdb-patches <gdb-patches@sourceware.org> writes: > When there is no dwarf2 data for a register, a function can be called > to provide the value of this register. In some situations, it might > not be trivial to determine the value to return and it would cause a > performance bottleneck to do the computation each time. Thanks for the patch. > This patch allows the called function to have a "cache" object that it > can use to store some metadata between calls to reduce the performance > impact of the complex logic. > The cache object is unique for each function and frame, so if there are > more than one function pointer stored in the dwarf2_frame_cache->reg > array, then the appropriate pointer will be supplied (the type is not > known by the dwarf2 implementation). Does this ever happen? If not perhaps a simpler approach would be better. > +struct dwarf2_frame_fn_data > +{ New type should have a comment. > + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, > + int regnum); Shouldn't this use the fn_prev_register typedef? > + void *data; > + struct dwarf2_frame_fn_data* next; Wrong placement of the "*", but really a lot of the code isn't following the GNU / gdb style. > +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void **this_cache, > + fn_prev_register fn) Normally new functions get a comment referring to the header where they are declared. > + > +/* Allocate a new instance of the function unique data. */ > + > +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, > + void **this_cache, > + fn_prev_register fn, > + unsigned long size); > + > +/* Retrieve the function unique data for this frame. */ > + > +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, I think these comments could perhaps be expanded a bit. Tom
>>>>> "Torbjorn" == Torbjorn SVENSSON via Gdb-patches <gdb-patches@sourceware.org> writes: >> + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); >> + fn_data->fn = fn; >> + fn_data->data = frame_obstack_zalloc (size); Torbjorn> Since these 2 blocks (fn_data and fn_data->fn) are allocated on the Torbjorn> obstack, do we need to release them in dwarf2_frame_dealloc_cache() or Torbjorn> just leave them to the generic garbage collection? They can just be left. The obstack as a whole is destroyed at once. However, you probably should consider the lifetime management of the data that's attached to the frame. That said, it's fine if the answer is that they should be allocated on the obstack as well. Tom
On 2022-12-20 22:02, Tom Tromey wrote: >>>>>> Torbjörn SVENSSON via Gdb-patches <gdb-patches@sourceware.org> writes: > >> When there is no dwarf2 data for a register, a function can be called >> to provide the value of this register. In some situations, it might >> not be trivial to determine the value to return and it would cause a >> performance bottleneck to do the computation each time. > > Thanks for the patch. > >> This patch allows the called function to have a "cache" object that it >> can use to store some metadata between calls to reduce the performance >> impact of the complex logic. > >> The cache object is unique for each function and frame, so if there are >> more than one function pointer stored in the dwarf2_frame_cache->reg >> array, then the appropriate pointer will be supplied (the type is not >> known by the dwarf2 implementation). > > Does this ever happen? If not perhaps a simpler approach would be > better. Right now; I don't know, but as the fn member in the dwarf2_frame_state_reg struct contains one function pointer per register, the architecture allows more than one function pointer per frame. If we went with a simpler solution, to only have one data block per frame, regardless of what function that is "owning" the data, then it could lead to nasty surprises if there is some unwinder that expects to be able to use more than data type... If we move the function pointer from the register scope to the frame scope, then I agree that only one data object would be needed. If we stick to having the function pointer per register, I could accept to have one data block, but somewhere, an assert should added so that the wrongful assumption mentioned above would be caught early rather than leading to strange bugs. This would mean that the type needs to be stored in the dwarf2_frame_cache struct somehow, but the type is currently internal to another compile unit. This is basically the reason why I went with the decoupled solution that is provided in this patch. > >> +struct dwarf2_frame_fn_data >> +{ > > New type should have a comment. Okay, I'll add comments, but I would like to know if this is the way to go or if there should be some alternative implementation before spending more time on this. > >> + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >> + int regnum); > > Shouldn't this use the fn_prev_register typedef? Indeed. > >> + void *data; >> + struct dwarf2_frame_fn_data* next; > > Wrong placement of the "*", but really a lot of the code isn't following > the GNU / gdb style. Don't know why the contrib/check_GNU_style.py in the GCC source tree did not flag this. Anyway, will be fixed in v3. >> +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void **this_cache, >> + fn_prev_register fn) > > Normally new functions get a comment referring to the header where they > are declared. Can you point me to an example and I will do something similar for these new functions if we decide to go this way. >> + >> +/* Allocate a new instance of the function unique data. */ >> + >> +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >> + void **this_cache, >> + fn_prev_register fn, >> + unsigned long size); >> + >> +/* Retrieve the function unique data for this frame. */ >> + >> +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, > > I think these comments could perhaps be expanded a bit. What more detail would you like to include? > > Tom Kind regards, Torbjörn
Hi, Any comments on my last reply? Kind regards, Torbjörn On 2022-12-28 17:16, Torbjorn SVENSSON wrote: > > On 2022-12-20 22:02, Tom Tromey wrote: >>>>>>> Torbjörn SVENSSON via Gdb-patches <gdb-patches@sourceware.org> >>>>>>> writes: >> >>> When there is no dwarf2 data for a register, a function can be called >>> to provide the value of this register. In some situations, it might >>> not be trivial to determine the value to return and it would cause a >>> performance bottleneck to do the computation each time. >> >> Thanks for the patch. >> >>> This patch allows the called function to have a "cache" object that it >>> can use to store some metadata between calls to reduce the performance >>> impact of the complex logic. >> >>> The cache object is unique for each function and frame, so if there are >>> more than one function pointer stored in the dwarf2_frame_cache->reg >>> array, then the appropriate pointer will be supplied (the type is not >>> known by the dwarf2 implementation). >> >> Does this ever happen? If not perhaps a simpler approach would be >> better. > > Right now; I don't know, but as the fn member in the > dwarf2_frame_state_reg struct contains one function pointer per > register, the architecture allows more than one function pointer per frame. > If we went with a simpler solution, to only have one data block per > frame, regardless of what function that is "owning" the data, then it > could lead to nasty surprises if there is some unwinder that expects to > be able to use more than data type... > If we move the function pointer from the register scope to the frame > scope, then I agree that only one data object would be needed. > If we stick to having the function pointer per register, I could accept > to have one data block, but somewhere, an assert should added so that > the wrongful assumption mentioned above would be caught early rather > than leading to strange bugs. This would mean that the type needs to be > stored in the dwarf2_frame_cache struct somehow, but the type is > currently internal to another compile unit. > This is basically the reason why I went with the decoupled solution that > is provided in this patch. > >> >>> +struct dwarf2_frame_fn_data >>> +{ >> >> New type should have a comment. > > Okay, I'll add comments, but I would like to know if this is the way to > go or if there should be some alternative implementation before spending > more time on this. > >> >>> + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >>> + int regnum); >> >> Shouldn't this use the fn_prev_register typedef? > > Indeed. > >> >>> + void *data; >>> + struct dwarf2_frame_fn_data* next; >> >> Wrong placement of the "*", but really a lot of the code isn't following >> the GNU / gdb style. > > Don't know why the contrib/check_GNU_style.py in the GCC source tree did > not flag this. Anyway, will be fixed in v3. > >>> +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void >>> **this_cache, >>> + fn_prev_register fn) >> >> Normally new functions get a comment referring to the header where they >> are declared. > > Can you point me to an example and I will do something similar for these > new functions if we decide to go this way. > >>> + >>> +/* Allocate a new instance of the function unique data. */ >>> + >>> +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >>> + void **this_cache, >>> + fn_prev_register fn, >>> + unsigned long size); >>> + >>> +/* Retrieve the function unique data for this frame. */ >>> + >>> +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, >> >> I think these comments could perhaps be expanded a bit. > > What more detail would you like to include? > >> >> Tom > > Kind regards, > Torbjörn
On Wed, Dec 28, 2022 at 05:16:03PM +0100, Torbjorn SVENSSON via Gdb-patches wrote: > > On 2022-12-20 22:02, Tom Tromey wrote: > > > > > > > Torbjörn SVENSSON via Gdb-patches <gdb-patches@sourceware.org> writes: > > > > > When there is no dwarf2 data for a register, a function can be called > > > to provide the value of this register. In some situations, it might > > > not be trivial to determine the value to return and it would cause a > > > performance bottleneck to do the computation each time. > > > > Thanks for the patch. > > > > > This patch allows the called function to have a "cache" object that it > > > can use to store some metadata between calls to reduce the performance > > > impact of the complex logic. > > > > > The cache object is unique for each function and frame, so if there are > > > more than one function pointer stored in the dwarf2_frame_cache->reg > > > array, then the appropriate pointer will be supplied (the type is not > > > known by the dwarf2 implementation). > > > > Does this ever happen? If not perhaps a simpler approach would be > > better. > > Right now; I don't know, but as the fn member in the dwarf2_frame_state_reg > struct contains one function pointer per register, the architecture allows > more than one function pointer per frame. > If we went with a simpler solution, to only have one data block per frame, > regardless of what function that is "owning" the data, then it could lead to > nasty surprises if there is some unwinder that expects to be able to use > more than data type... > If we move the function pointer from the register scope to the frame scope, > then I agree that only one data object would be needed. > If we stick to having the function pointer per register, I could accept to > have one data block, but somewhere, an assert should added so that the > wrongful assumption mentioned above would be caught early rather than > leading to strange bugs. This would mean that the type needs to be stored in > the dwarf2_frame_cache struct somehow, but the type is currently internal to > another compile unit. > This is basically the reason why I went with the decoupled solution that is > provided in this patch. > > > > > > +struct dwarf2_frame_fn_data > > > +{ > > > > New type should have a comment. > > Okay, I'll add comments, but I would like to know if this is the way to go > or if there should be some alternative implementation before spending more > time on this. > > > > > > + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, > > > + int regnum); > > > > Shouldn't this use the fn_prev_register typedef? > > Indeed. > > > > > > + void *data; > > > + struct dwarf2_frame_fn_data* next; > > > > Wrong placement of the "*", but really a lot of the code isn't following > > the GNU / gdb style. > > Don't know why the contrib/check_GNU_style.py in the GCC source tree did not > flag this. Anyway, will be fixed in v3. > > > > +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void **this_cache, > > > + fn_prev_register fn) > > > > Normally new functions get a comment referring to the header where they > > are declared. > > Can you point me to an example and I will do something similar for these new > functions if we decide to go this way. See for instance is_fixed_point_type's declaration in gdbtypes.h: /* Return True if TYPE is a TYPE_CODE_FIXED_POINT or if TYPE is a range type whose base type is a TYPE_CODE_FIXED_POINT. */ extern bool is_fixed_point_type (struct type *type); ... where we provide a description of what the function does, including information about all the parameters. And then the corresponding implementation in gdbtypes.c: /* See gdbtypes.h. */ bool is_fixed_point_type (struct type *type) { The goal here is to provide the function's documentation at the API level, rather than at the implementation level, for all situations where it is declared like so. For situations where the function is declared static, we put the function's documentation next to the implementation, as there isn't always a declaration. The "See xxx.h" comment is there to confirm where the function's documentation can be found.
>>>>> Torbjörn SVENSSON via Gdb-patches <gdb-patches@sourceware.org> writes: > When there is no dwarf2 data for a register, a function can be called > to provide the value of this register. In some situations, it might > not be trivial to determine the value to return and it would cause a > performance bottleneck to do the computation each time. > This patch allows the called function to have a "cache" object that it > can use to store some metadata between calls to reduce the performance > impact of the complex logic. > +struct dwarf2_frame_fn_data > +{ This struct should have some kind of introductory comment, as should the fields. > + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, > + int regnum); Seems like this should use the fn_prev_register typedef. (But see below.) > + void *data; > + struct dwarf2_frame_fn_data* next; Wrong "*" placement. > +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void **this_cache, > + fn_prev_register fn) Wrong formatting. > +void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, > + void **this_cache, > + fn_prev_register fn, unsigned long size) Wrong formatting. > +{ > + struct dwarf2_frame_fn_data *fn_data = nullptr; > + struct dwarf2_frame_cache *cache > + = dwarf2_frame_cache (this_frame, this_cache); > + > + /* First try to find an existing object. */ > + void *data = dwarf2_frame_get_fn_data (this_frame, this_cache, fn); > + if (data) > + return data; > + > + /* No object found, lets create a new instance. */ > + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); > + fn_data->fn = fn; > + fn_data->data = frame_obstack_zalloc (size); > + fn_data->next = cache->fn_data; > + cache->fn_data = fn_data; > + > + return fn_data->data; This API seems a bit weird to me. It seems like the 'fn' parameter is never really used. It's maybe just a sort of cookie. But if so, I think it would be better to just use a 'const void *' or 'const char *' or something like that. (A string is nice because then it can also be seen in the debugger and give a clue where it came from.) Ok, I dug up the follow-up patch and indeed it is just a cookie. I think naming it as such and changing the type would make this more clear. Also in the follow-up I see that it calls dwarf2_frame_fn_data first. So if you're going to go that route, then it seems that dwarf2_frame_allocate_fn_data does not need to find an existing object -- it can just assert there isn't one. > + > +/* Allocate a new instance of the function unique data. */ > + > +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, > + void **this_cache, > + fn_prev_register fn, > + unsigned long size); > + > +/* Retrieve the function unique data for this frame. */ > + > +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, > + void **this_cache, > + fn_prev_register fn); IMO both of these could use a longer comment. From this it's impossible to tell what the point of them is. thanks, Tom
On 2023-01-18 19:47, Tom Tromey wrote: >>>>>> Torbjörn SVENSSON via Gdb-patches <gdb-patches@sourceware.org> writes: > >> When there is no dwarf2 data for a register, a function can be called >> to provide the value of this register. In some situations, it might >> not be trivial to determine the value to return and it would cause a >> performance bottleneck to do the computation each time. > >> This patch allows the called function to have a "cache" object that it >> can use to store some metadata between calls to reduce the performance >> impact of the complex logic. > >> +struct dwarf2_frame_fn_data >> +{ > > This struct should have some kind of introductory comment, as should the > fields. Fixed in v3. > >> + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, >> + int regnum); > > Seems like this should use the fn_prev_register typedef. (But see below.) > >> + void *data; >> + struct dwarf2_frame_fn_data* next; > > Wrong "*" placement. Fixed in v3. > >> +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void **this_cache, >> + fn_prev_register fn) > > Wrong formatting. > >> +void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >> + void **this_cache, >> + fn_prev_register fn, unsigned long size) > > Wrong formatting. I think the format is correct, but it looks strange in the mail as the '+' sign is "eaten" by the first tab. Looking in my editor for the source code, it looks correct. >> +{ >> + struct dwarf2_frame_fn_data *fn_data = nullptr; >> + struct dwarf2_frame_cache *cache >> + = dwarf2_frame_cache (this_frame, this_cache); >> + >> + /* First try to find an existing object. */ >> + void *data = dwarf2_frame_get_fn_data (this_frame, this_cache, fn); >> + if (data) >> + return data; >> + >> + /* No object found, lets create a new instance. */ >> + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); >> + fn_data->fn = fn; >> + fn_data->data = frame_obstack_zalloc (size); >> + fn_data->next = cache->fn_data; >> + cache->fn_data = fn_data; >> + >> + return fn_data->data; > > This API seems a bit weird to me. > > It seems like the 'fn' parameter is never really used. It's maybe just > a sort of cookie. But if so, I think it would be better to just use a > 'const void *' or 'const char *' or something like that. (A string is > nice because then it can also be seen in the debugger and give a clue > where it came from.) > > Ok, I dug up the follow-up patch and indeed it is just a cookie. I > think naming it as such and changing the type would make this more > clear. Yes, it's the cookie/key to identify the right object for the prev_register function. In v3, I've renamed the variable (and the member) to be called "cookie" instead if that's better. The "fn" name came from the struct where it was originally defined. The cookie/key could be a string, but then it should be auto-generated when calling the functions rather than letting the user type it. The reason is that if the user types it, it's less clear what the consequences of reusing it will be. Also, if we go for a string, it would consume more memory than having a function pointer (not that memory should be an issue, but anyway...). If there is a function pointer, doesn't GDB lookup the symbol for the function pointer if there are debug symbols in the application? > > Also in the follow-up I see that it calls dwarf2_frame_fn_data first. > So if you're going to go that route, then it seems that > dwarf2_frame_allocate_fn_data does not need to find an existing > object -- it can just assert there isn't one. What happens if there are 2 prev_register functions that both wants some custom data? With the approach I have (looping over the list and returning the matching data for the cookie/key) would allow us to reuse the existing object rather than create a new one. I could change the block that checks if there was a match and returns it to an assert, but what would the benefit be? Instead of reusing the existing object, we would crash GDB. Is this better? > >> + >> +/* Allocate a new instance of the function unique data. */ >> + >> +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, >> + void **this_cache, >> + fn_prev_register fn, >> + unsigned long size); >> + >> +/* Retrieve the function unique data for this frame. */ >> + >> +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, >> + void **this_cache, >> + fn_prev_register fn); > > IMO both of these could use a longer comment. From this it's impossible > to tell what the point of them is. Improved in v3. > > thanks, > Tom Kind regards, Torbjörn + Yvan
diff --git a/gdb/dwarf2/frame.c b/gdb/dwarf2/frame.c index 3f884abe1d5..bff3b706e7e 100644 --- a/gdb/dwarf2/frame.c +++ b/gdb/dwarf2/frame.c @@ -831,6 +831,14 @@ dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, CORE_ADDR pc, } +struct dwarf2_frame_fn_data +{ + struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, + int regnum); + void *data; + struct dwarf2_frame_fn_data* next; +}; + struct dwarf2_frame_cache { /* DWARF Call Frame Address. */ @@ -862,6 +870,8 @@ struct dwarf2_frame_cache dwarf2_tailcall_frame_unwind unwinder so this field does not apply for them. */ void *tailcall_cache; + + struct dwarf2_frame_fn_data *fn_data; }; static struct dwarf2_frame_cache * @@ -1221,6 +1231,44 @@ dwarf2_frame_prev_register (frame_info_ptr this_frame, void **this_cache, } } +void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, void **this_cache, + fn_prev_register fn) +{ + struct dwarf2_frame_fn_data *fn_data = nullptr; + struct dwarf2_frame_cache *cache + = dwarf2_frame_cache (this_frame, this_cache); + + /* Find the object for the function. */ + for (fn_data = cache->fn_data; fn_data; fn_data = fn_data->next) + if (fn_data->fn == fn) + return fn_data->data; + + return nullptr; +} + +void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, + void **this_cache, + fn_prev_register fn, unsigned long size) +{ + struct dwarf2_frame_fn_data *fn_data = nullptr; + struct dwarf2_frame_cache *cache + = dwarf2_frame_cache (this_frame, this_cache); + + /* First try to find an existing object. */ + void *data = dwarf2_frame_get_fn_data (this_frame, this_cache, fn); + if (data) + return data; + + /* No object found, lets create a new instance. */ + fn_data = FRAME_OBSTACK_ZALLOC (struct dwarf2_frame_fn_data); + fn_data->fn = fn; + fn_data->data = frame_obstack_zalloc (size); + fn_data->next = cache->fn_data; + cache->fn_data = fn_data; + + return fn_data->data; +} + /* Proxy for tailcall_frame_dealloc_cache for bottom frame of a virtual tail call frames chain. */ diff --git a/gdb/dwarf2/frame.h b/gdb/dwarf2/frame.h index 06c8a10c178..444afd9f8eb 100644 --- a/gdb/dwarf2/frame.h +++ b/gdb/dwarf2/frame.h @@ -66,6 +66,9 @@ enum dwarf2_frame_reg_rule /* Register state. */ +typedef struct value *(*fn_prev_register) (frame_info_ptr this_frame, + void **this_cache, int regnum); + struct dwarf2_frame_state_reg { /* Each register save state can be described in terms of a CFA slot, @@ -78,8 +81,7 @@ struct dwarf2_frame_state_reg const gdb_byte *start; ULONGEST len; } exp; - struct value *(*fn) (frame_info_ptr this_frame, void **this_cache, - int regnum); + fn_prev_register fn; } loc; enum dwarf2_frame_reg_rule how; }; @@ -262,4 +264,18 @@ extern int dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, CORE_ADDR pc, const gdb_byte **cfa_start_out, const gdb_byte **cfa_end_out); + +/* Allocate a new instance of the function unique data. */ + +extern void *dwarf2_frame_allocate_fn_data (frame_info_ptr this_frame, + void **this_cache, + fn_prev_register fn, + unsigned long size); + +/* Retrieve the function unique data for this frame. */ + +extern void *dwarf2_frame_get_fn_data (frame_info_ptr this_frame, + void **this_cache, + fn_prev_register fn); + #endif /* dwarf2-frame.h */