-
Notifications
You must be signed in to change notification settings - Fork 12.9k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Include frame pointer for bare metal RISC-V targets #61675
Conversation
(rust_highfive has picked a reviewer for you, use r? to override) |
r? @nagisa |
I’m trying to understand the motivation better here, please bear with me :)
Looking at the LLVM code (here and here, it seems that LLVM will not reserve
My understanding and experience in embedded scene suggests that the ideal approach is indeed to load an ELF executable in your debugger and let it handle all sorts of mappings that may be necessary between the loaded ELF and binary code loaded into ROM. Reading this, it looks to me that you’re looking to generate stack traces internally (a-la
On the other hand, won’t landing this increase the size of generated code by default? Do RISC-V (unlike ARM) targets have an excess of storage for machine code? cc @rkruppe |
I'm happy to explain.
I might be completely wrong on this. My claim was based on compiling a bunch of Rust code and observing the output from rustc seemed to never include
Ah, I think this the source of confusion: You are thinking of tiny embedded chips with a tiny amount of ROM and even smaller amount of RAM, whereas I've been thinking about writing kernels for systems with gigabytes of RAM (like the HiFive Unleashed) which are powerful enough to run a full desktop OS. |
Fair. It still seems to me that these targets might be used for embedded chips, but I have no idea if those even exist, so @bors r+ |
📌 Commit 493d1b4 has been approved by |
RV32IM(A)C cores with tiny program memories for embedded applications certainly exist, I think they vastly outnumber the big linux-capable cores both in number of designs available and units shipped. I don't know whether that is reflected in the usage statistics of the Rust triples, but maybe it would be justified to keep frame pointers disabled at least on the smallest (riscv32imc) target? HiFive Unleashed is 64 bit anyway (with no way to run 32 bit code), so if @fintelia's targeting that, they wouldn't be affected even if they didn't use the riscv64gc triple (e.g. riscv64imac to disable hard float). |
Include frame pointer for bare metal RISC-V targets This changes the default setting to enable the use of the frame pointer register when targeting RISC-V. On that architecture there is a dedicated frame pointer register which LLVM would otherwise never use so there is no increase in register pressure. Further, since these are bare metal targets, getting backtraces without the frame pointer is considerably more difficult (you can't just ask the OS to load the ELF executable and parse DWARF symbols). It is true that this setting can also be changed with the `-C force-frame-pointers` flag but that won't impact the compilation of the standard library, meaning that backtraces from, say, a panic handler would be useless.
Include frame pointer for bare metal RISC-V targets This changes the default setting to enable the use of the frame pointer register when targeting RISC-V. On that architecture there is a dedicated frame pointer register which LLVM would otherwise never use so there is no increase in register pressure. Further, since these are bare metal targets, getting backtraces without the frame pointer is considerably more difficult (you can't just ask the OS to load the ELF executable and parse DWARF symbols). It is true that this setting can also be changed with the `-C force-frame-pointers` flag but that won't impact the compilation of the standard library, meaning that backtraces from, say, a panic handler would be useless.
Rollup of 9 pull requests Successful merges: - #60376 (Stabilize Option::xor) - #61398 (Stabilize copy_within) - #61629 (Hygienize macros in the standard library) - #61675 (Include frame pointer for bare metal RISC-V targets) - #61750 (Fix x.py install) - #61761 (Add an alias for x86_64-sun-solaris target tuple) - #61762 (rustbuild: fix libtest_stamp) - #61763 (ci: fix ci stats upload condition) - #61776 (Fix typos in error_codes) Failed merges: r? @ghost
@rkruppe What you are describing seems reasonable. Slightly longer term rust-lang/rfcs#2663 or similar should hopefully resolve the issue of having to pick only one set of compile options for the standard library of each target. |
We have been seeing some very inefficient code that went away when using `-Cforce-frame-pointers=no`. For instance `core::ptr::drop_in_place` at `-Oz` was compiled into a function which consisted entirely of saving registers to the stack, then using the frame pointer to restore the same registers (without any instructions between the prolog and epilog). The RISC-V LLVM backend supports frame pointer elimination, so it makes sense to allow this to happen when using Rust. It's not clear to me that frame pointers have ever been required in the general case. In rust-lang#61675 it was pointed out that this made reassembling stack traces easier, which is true, but there is a code generation option for forcing frame pointers, and I feel the default should not be to require frame pointers, given it demonstrably makes code size worse (around 10% in some embedded applications). The kinds of targets mentioned in rust-lang#61675 are popular, but should not dictate that code generation should be worse for all RISC-V targets, especially as there is a way to use CFI information to reconstruct the stack when the frame pointer is eliminated. It is also a misconception that `fp` is always used for the frame pointer. `fp` is an ABI name for `x8` (aka `s0`), and if no frame pointer is required, `x8` may be used for other callee-saved values. This commit does ensure that the standard library is built with unwind tables, so that users do not need to rebuild the standard library in order to get a backtrace that includes standard library calls (which is the original reason for forcing frame pointers).
…=hanna-kruppe,Mark-Simulacrum [RISC-V] Do not force frame pointers We have been seeing some very inefficient code that went away when using `-Cforce-frame-pointers=no`. For instance `core::ptr::drop_in_place` at `-Oz` was compiled into a function which consisted entirely of saving registers to the stack, then using the frame pointer to restore the same registers (without any instructions between the prolog and epilog). The RISC-V LLVM backend supports frame pointer elimination, so it makes sense to allow this to happen when using Rust. It's not clear to me that frame pointers have ever been required in the general case. In rust-lang#61675 it was pointed out that this made reassembling stack traces easier, which is true, but there is a code generation option for forcing frame pointers, and I feel the default should not be to require frame pointers, given it demonstrably makes code size worse (around 10% in some embedded applications). The kinds of targets mentioned in rust-lang#61675 are popular, but should not dictate that code generation should be worse for all RISC-V targets, especially as there is a way to use CFI information to reconstruct the stack when the frame pointer is eliminated. It is also a misconception that `fp` is always used for the frame pointer. `fp` is an ABI name for `x8` (aka `s0`), and if no frame pointer is required, `x8` may be used for other callee-saved values. --- I am partly posting this to get feedback from @fintelia who introduced the change to require frame pointers, and @hanna-kruppe who had issues with the original PR. I would understand if we wanted to remove this setting on only a subset of RISC-V targets, but my preference would be to remove this setting everywhere. There are more details on the code size savings seen in Tock here: tock/tock#1660
… r=hanna-kruppe,Mark-Simulacrum [RISC-V] Do not force frame pointers We have been seeing some very inefficient code that went away when using `-Cforce-frame-pointers=no`. For instance `core::ptr::drop_in_place` at `-Oz` was compiled into a function which consisted entirely of saving registers to the stack, then using the frame pointer to restore the same registers (without any instructions between the prolog and epilog). The RISC-V LLVM backend supports frame pointer elimination, so it makes sense to allow this to happen when using Rust. It's not clear to me that frame pointers have ever been required in the general case. In rust-lang#61675 it was pointed out that this made reassembling stack traces easier, which is true, but there is a code generation option for forcing frame pointers, and I feel the default should not be to require frame pointers, given it demonstrably makes code size worse (around 10% in some embedded applications). The kinds of targets mentioned in rust-lang#61675 are popular, but should not dictate that code generation should be worse for all RISC-V targets, especially as there is a way to use CFI information to reconstruct the stack when the frame pointer is eliminated. It is also a misconception that `fp` is always used for the frame pointer. `fp` is an ABI name for `x8` (aka `s0`), and if no frame pointer is required, `x8` may be used for other callee-saved values. --- I am partly posting this to get feedback from @fintelia who introduced the change to require frame pointers, and @hanna-kruppe who had issues with the original PR. I would understand if we wanted to remove this setting on only a subset of RISC-V targets, but my preference would be to remove this setting everywhere. There are more details on the code size savings seen in Tock here: tock/tock#1660
… r=hanna-kruppe,Mark-Simulacrum [RISC-V] Do not force frame pointers We have been seeing some very inefficient code that went away when using `-Cforce-frame-pointers=no`. For instance `core::ptr::drop_in_place` at `-Oz` was compiled into a function which consisted entirely of saving registers to the stack, then using the frame pointer to restore the same registers (without any instructions between the prolog and epilog). The RISC-V LLVM backend supports frame pointer elimination, so it makes sense to allow this to happen when using Rust. It's not clear to me that frame pointers have ever been required in the general case. In rust-lang#61675 it was pointed out that this made reassembling stack traces easier, which is true, but there is a code generation option for forcing frame pointers, and I feel the default should not be to require frame pointers, given it demonstrably makes code size worse (around 10% in some embedded applications). The kinds of targets mentioned in rust-lang#61675 are popular, but should not dictate that code generation should be worse for all RISC-V targets, especially as there is a way to use CFI information to reconstruct the stack when the frame pointer is eliminated. It is also a misconception that `fp` is always used for the frame pointer. `fp` is an ABI name for `x8` (aka `s0`), and if no frame pointer is required, `x8` may be used for other callee-saved values. --- I am partly posting this to get feedback from @fintelia who introduced the change to require frame pointers, and @hanna-kruppe who had issues with the original PR. I would understand if we wanted to remove this setting on only a subset of RISC-V targets, but my preference would be to remove this setting everywhere. There are more details on the code size savings seen in Tock here: tock/tock#1660
This changes the default setting to enable the use of the frame pointer register when targeting RISC-V. On that architecture there is a dedicated frame pointer register which LLVM would otherwise never use so there is no increase in register pressure. Further, since these are bare metal targets, getting backtraces without the frame pointer is considerably more difficult (you can't just ask the OS to load the ELF executable and parse DWARF symbols). It is true that this setting can also be changed with the
-C force-frame-pointers
flag but that won't impact the compilation of the standard library, meaning that backtraces from, say, a panic handler would be useless.