I recently had CI start failing for some firmware I use to make sure Rust still compiles msp430 code correctly. It appears that as of #98582, Rust has stopped optimizing out dead panic code; there is no I/O device to send the panic strings to, and so panic strings- and setting up arguments to access panic strings- should be considered dead code.
Normally, the firmware size should take 1994 (1992+2) bytes:
cargo build --release -Zbuild-std=core --target=msp430-none-elf
msp430-elf-size target/msp430-none-elf/release/at2xt
text data bss dec hex filename
1992 2 48 2042 7fa target/msp430-none-elf/release/at2xt
However, the firmware size had gone up 100 bytes (2048+46 = 2094- 5% regression) before #98582 was reverted by #99495:
cargo build --release -Zbuild-std=core --target=msp430-none-elf
error: linking with `msp430-elf-gcc` failed: exit status: 1
|
= note: "msp430-elf-gcc" "/home/william/Projects/embedded/msp430/AT2XT/target/msp430-none-elf/release/deps/at2xt-4706a59318f01763.at2xt.aa970193-cgu.0.rcgu.o" "-L" "/home/william/Projects/embedded/msp430/AT2XT/target/msp430-none-elf/release/deps" "-L" "/home/william/Projects/embedded/msp430/AT2XT/target/release/deps" "-L" "/home/william/Projects/embedded/msp430/AT2XT/target/msp430-none-elf/release/build/msp430-rt-587a8251c4c89199/out" "-L" "/home/william/Projects/embedded/msp430/AT2XT/target/msp430-none-elf/release/build/msp430g2211-1199852debc08ead/out" "-L" "/home/william/Projects/toolchains/build-llvm-toolchain/build-rust/build/x86_64-unknown-linux-gnu/stage1/lib/rustlib/msp430-none-elf/lib" "-Wl,-Bstatic" "/tmp/rustcy1Mond/libmsp430_rt-19edab0518267d0f.rlib" "-Wl,--start-group" "-Wl,--end-group" "/home/william/Projects/embedded/msp430/AT2XT/target/msp430-none-elf/release/deps/libcompiler_builtins-1ef8666a5bfcc3f4.rlib" "-Wl,-Bdynamic" "-L" "/home/william/Projects/toolchains/build-llvm-toolchain/build-rust/build/x86_64-unknown-linux-gnu/stage1/lib/rustlib/msp430-none-elf/lib" "-o" "/home/william/Projects/embedded/msp430/AT2XT/target/msp430-none-elf/release/deps/at2xt-4706a59318f01763" "-nodefaultlibs" "-Tlink.x" "-mcpu=msp430" "-nostartfiles" "-lmul_none" "-lgcc"
= note: /home/william/.local/bin/../lib/gcc/msp430-elf/9.2.0/../../../../msp430-elf/bin/ld: /home/william/Projects/embedded/msp430/AT2XT/target/msp430-none-elf/release/deps/at2xt-4706a59318f01763 section `.rodata' will not fit in region `ROM'
/home/william/.local/bin/../lib/gcc/msp430-elf/9.2.0/../../../../msp430-elf/bin/ld: section .rodata VMA wraps around address space
/home/william/.local/bin/../lib/gcc/msp430-elf/9.2.0/../../../../msp430-elf/bin/ld: section .vector_table LMA [000000000000ffe0,000000000000ffff] overlaps section .rodata LMA [000000000000ff48,000000000001000b]
/home/william/.local/bin/../lib/gcc/msp430-elf/9.2.0/../../../../msp430-elf/bin/ld: region `ROM' overflowed by 46 bytes
collect2: error: ld returned 1 exit status
error: could not compile `at2xt` due to previous error
error: Recipe `timer` failed on line 10 with exit code 101
I've created an MVCE from the failing CI to illustrate:
Code
Instructions
-
Make sure the msp430-elf-gcc toolchain is installed. Optionally install just for convenience.
-
git clone https://github.com/cr1901/msp430-size. Use commit e44cf66 specifically.
-
Compile rustc and add your new rustc to rustup using rustup toolchain link override-name /path/to/override. Use this override for both rustc commits in the next step.
-
Run the below command twice (using different target dirs):
cargo +override-name rustc --manifest-path=./take-api/Cargo.toml --release -Zbuild-std=core --example=once -- --emit=obj=target/msp430-none-elf/release/examples/once.o,llvm-ir=target/msp430-none-elf/release/examples/once.ll,asm=target/msp430-none-elf/release/examples/once.s
The above command needs to be run with two different compilers:
-
Extra compiler artifacts can be generated like so:
msp430-elf-objdump -Cd target/msp430-none-elf/release/examples/once
msp430-elf-readelf -a --wide target/msp430-none-elf/release/examples/once
msp430-elf-objdump -Cd target/msp430-none-elf/release/examples/once.o
msp430-elf-readelf -a --wide target/msp430-none-elf/release/examples/once.o
msp430-elf-size target/msp430-none-elf/release/examples/once
Expected Results
For both Rust compiler commits, I expected to see the following output from msp430-elf-size, with no panic strings emitted:
william@xubuntu-dtrain:~/Projects/embedded/msp430/msp430-size$ msp430-elf-size -A target/msp430-none-elf/release/examples/once
target/msp430-none-elf/release/examples/once :
section size addr
.vector_table 32 65504
.text 168 49152
.rodata 4 49320
.bss 2 512
.data 0 514
.MSP430.attributes 23 0
Total 229
strings target/msp430-none-elf/release/examples/once | grep called
Instead, the compiler which had PR #98582 had a different size, and a panic string for Option::unwrap() failure was emitted into the binary:
william@xubuntu-dtrain:~/Projects/embedded/msp430/msp430-size$ msp430-elf-size -A target/msp430-none-elf/release/examples/once
target/msp430-none-elf/release/examples/once :
section size addr
.vector_table 32 65504
.text 200 49152
.rodata 48 49352
.bss 2 512
.data 0 514
.MSP430.attributes 23 0
Total 305
strings target/msp430-none-elf/release/examples/once | grep called
?called `Option::unwrap()` on a `None` value
Regression Commits
Based on git bisect, this regression was introduced by PR #98582. Specifically 728c7e8 introduced the regression. This regression was fixed by #99495. However, @oli-obk asked me to open an issue anyway.
@rustbot modify labels: +regression-from-stable-to-nightly -regression-untriaged
Other Context
Although I don't have numbers handy, this regression also affects libcores compiled with the panic_immediate_abort feature for less marked size regression.
I recently had CI start failing for some firmware I use to make sure Rust still compiles msp430 code correctly. It appears that as of #98582, Rust has stopped optimizing out dead panic code; there is no I/O device to send the panic strings to, and so panic strings- and setting up arguments to access panic strings- should be considered dead code.
Normally, the firmware size should take 1994 (1992+2) bytes:
However, the firmware size had gone up 100 bytes (2048+46 = 2094- 5% regression) before #98582 was reverted by #99495:
I've created an MVCE from the failing CI to illustrate:
Code
Instructions
Make sure the
msp430-elf-gcctoolchain is installed. Optionally installjustfor convenience.git clone https://github.com/cr1901/msp430-size. Use commit e44cf66 specifically.Compile
rustcand add your newrustctorustupusingrustup toolchain link override-name /path/to/override. Use this override for bothrustccommits in the next step.Run the below command twice (using different
targetdirs):The above command needs to be run with two different compilers:
Extra compiler artifacts can be generated like so:
Expected Results
For both Rust compiler commits, I expected to see the following output from
msp430-elf-size, with no panic strings emitted:Instead, the compiler which had PR #98582 had a different size, and a panic string for
Option::unwrap()failure was emitted into the binary:Regression Commits
Based on
git bisect, this regression was introduced by PR #98582. Specifically 728c7e8 introduced the regression. This regression was fixed by #99495. However, @oli-obk asked me to open an issue anyway.@rustbot modify labels: +regression-from-stable-to-nightly -regression-untriaged
Other Context
Although I don't have numbers handy, this regression also affects
libcores compiled with thepanic_immediate_abortfeature for less marked size regression.