STR in Thumb 32 decode problem

[JIANG Muhui]

Hi

It seems that QEMU does not have a proper check on the STR instruction in Thumb32 mode.

Specifically, the machine code is 0xf84f0ddd, which is 0b1111 1000 0100 1111 0000 1101 1101 1101.
This is an STR (immediate, Thumb) instruction with a T4 encoding scheme.

The symbols is

Rn = 1111
Rt = 0000
P = 1
U = 0
W = 1

The decode ASL is below:

if P == ‘1’ && U == ‘1’ && W == ‘0’ then SEE STRT;
if Rn == ‘1101’ && P == ‘1’ && U == ‘0’ && W == ‘1’ && imm8 == ‘00000100’ then SEE PUSH;
if Rn == ‘1111’ || (P == ‘0’ && W == ‘0’) then UNDEFINED;
t = UInt(Rt); n = UInt(Rn); imm32 = ZeroExtend(imm8, 32);
index = (P == ‘1’); add = (U == ‘1’); wback = (W == ‘1’);
if t == 15 || (wback && n == t) then UNPREDICTABLE;

When Rn == 1111, it should be an undefined instruction, which should raise SEGILL signal. However, it seems that QEMU does not check this constraint, which should be a bug. Many thanks

Regards
Muhui

[Peter Maydell (pmaydell)]

NB that for Arm mode the equivalent case is UNPREDICTABLE only in the writeback case (and has defined behaviour for non-writeback), so we need to enforce the UNDEF on rn==15 for Thumb decode only.

[Peter Maydell (pmaydell)]

Patch sent to list: if you could test it against whatever your test case was that would be helpful.
https://<email address hidden>/

PS: out of interest, why/how were you checking should-UNDEF cases ?

[JIANG Muhui (muhui)]

We just test the patched version. It looks well. Now QEMU would raise SEGILL signals, which should be the right behavior.

We are not checking should-UNDEF cases in particular. This is a case we observed and checked manually when doing a research project with QEMU.

[Thomas Huth (th-huth)]

Patch has been merged:
https://gitlab.com/qemu-project/qemu/-/commit/8196fe9d83d6519128b5