thumbv4t-none-eabi target doesn't work

[jonathanpallant]

In examples/versatileab run:

cargo run --target thumbv4t-none-eabi --bin hello -Zbuild-std

qemu-system-arm will hang without printing anything. I think there's an issue with the branches I added as part of the SMP support where we end up running Arm code in Thumb mode (or vice-versa).

This doesn't seem to happen with the thumbv5te-none-eabi target, probably because that architecture has a richer set of instructions from moving between Arm and Thumb modes. IIRC in Armv4T there is BL and there is BX but there is no BLX, meaning it has to be emulated with a BL to a thunk routine that does a BX (or something like that).

This blocks #211

[jonathanpallant]

Also, why didn't CI fail? Is it my qemu version? I am using 11.1.0 locally.

[thejpster]

Ugh, it's a debug profile vs release profile thing. The CI always runs --release.

[thejpster]

OK on thumbv4-none-eabi, I observed a global_asm! function assembled as A32 but with an alignment of 2. All A32 functions must have an alignment of 4, so this caused a prefetch abort.

$ cd examples/versatileab
$ cargo build --target thumbv4t-none-eabi --bin hello -Zbuild-std  && arm-none-eabi-objdump -d ./target/thumbv4t-none-eabi/debug/hello | rustfilt | grep _asm_core_start -A4
   Compiling aarch32-rt v0.4.0 (/Volumes/Macintosh SSD/jonathan/Documents/github/rust-embedded/aarch32/aarch32-rt)
   Compiling versatileab v0.0.0 (/Volumes/Macintosh SSD/jonathan/Documents/github/rust-embedded/aarch32/examples/versatileab)
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 0.44s
00002326 <_asm_core_start>:
    2326:	e1a0c000 	mov	ip, r0
    232a:	eb000012 	bl	237a <_asm_core_start+0x54>
    232e:	e3a01000 	mov	r1, #0
    2332:	e3a02000 	mov	r2, #0
    2336:	e3a03000 	mov	r3, #0
    233a:	e3a04000 	mov	r4, #0
--
    2366:	1a000001 	bne	2372 <_asm_core_start+0x4c>
    236a:	eb005fa7 	bl	1a20e <_default_start+0xa>
    236e:	eafffffe 	b	236e <_asm_core_start+0x48>
    2372:	eb005fa8 	bl	1a21a <__ARMv4ABSLongBXThunk_kmain+0xa>
    2376:	eafffffe 	b	2376 <_asm_core_start+0x50>
    237a:	Address 0x237a is out of bounds.


0000237c <_asm_stack_setup_preallocated>:

First up, it's weird that on thumb targets the global_asm! was being assembled as A32. To avoid confusion, I will mark all global_asm! functions with .arm.

Second, I think we've just been extremely lucky so far with our alignment. The asm emitted by rustc always includes .p2align 2 for A32 functions. I should add that directive to all of the global_asm! functions.

[thejpster]

Fixed by #219