IAR LDR{,B,H}T fix

[RobinKastberg]

Inline assembler constraints were incorrect.

[github-actions]

Test Results

   264 files   -   108     264 suites   - 108   0s ⏱️ - 8m 25s
    56 tests +    7      54 ✅ + 10      2 💤  -     3  0 ❌ ±0 
14 292 runs   - 3 936  12 124 ✅ +160  2 168 💤  - 4 096  0 ❌ ±0 

Results for commit 456dc6f. ± Comparison against base commit 0fa5d9f.

This pull request removes 49 and adds 56 tests. Note that renamed tests count towards both.

CMSIS-Core.src ‑ apsr.c
CMSIS-Core.src ‑ basepri.c
CMSIS-Core.src ‑ bkpt.c
CMSIS-Core.src ‑ clrex.c
CMSIS-Core.src ‑ clz.c
CMSIS-Core.src ‑ control.c
CMSIS-Core.src ‑ cp15.c
CMSIS-Core.src ‑ cpsr.c
CMSIS-Core.src ‑ dmb.c
CMSIS-Core.src ‑ dsb.c
…

TC_CML1Cache_CleanDCacheByAddrWhileDisabled
TC_CML1Cache_EnDisableDCache
TC_CML1Cache_EnDisableICache
TC_CoreFunc_APSR
TC_CoreFunc_BASEPRI
TC_CoreFunc_Control
TC_CoreFunc_EnDisIRQ
TC_CoreFunc_EncDecIRQPrio
TC_CoreFunc_FAULTMASK
TC_CoreFunc_FPSCR
…

This pull request removes 5 skipped tests and adds 2 skipped tests. Note that renamed tests count towards both.

CMSIS-Core.src ‑ lda.c
CMSIS-Core.src ‑ ldaex.c
CMSIS-Core.src ‑ stl.c
CMSIS-Core.src ‑ stlex.c
CMSIS-Core.src ‑ systick.c

TC_CoreInstr_WFE
TC_CoreInstr_WFI

♻️ This comment has been updated with latest results.

[RobinKastberg]

@JonatanAntoni I think this issue affects the other toolchains as well.. just that they got more luck with register allocation. however I hadnt set up the others to test with. Maybe worth a look?

[JonatanAntoni]

@RobinKastberg, any chance to detect issues like this in a test? E.g., by inspecting the assembly output (disassembly) as done in the Core tests using lit?

I.e., you added ampersand output operand (=&). I read this as a specific instruction to the compiler to not reuse the input register for output value. I am far from being an compiler expert to tell if this is required or compilers may be clever enough.

[RobinKastberg]

@RobinKastberg, any chance to detect issues like this in a test? E.g., by inspecting the assembly output (disassembly) as done in the Core tests using lit?

I.e., you added ampersand output operand (=&). I read this as a specific instruction to the compiler to not reuse the input register for output value. I am far from being an compiler expert to tell if this is required or compilers may be clever enough.

If you can trigger it you will get a compilation/assemblation error and dont need a test (I would assume, haven't checked) The hard part is making the compiler register allocation put the input register into the same register as output. And if we try too hard we will succeed and the compilation will fail. Maybe there is some happy middle ground.

I can experiment a bit.