Incorrect logic on multiple SMSTART calls

[FinnWilkinson]

As part of the SME AArch64 extension, Streaming SVE Mode was introduced. This is enabled via SMSTART and disabled by SMSTOP. For any SME instruction to be executed the core must be in Streaming SVE mode.

SMSTART/SMSTOP is composed of 2 parts: SM and ZA (not to be confused with the register za). When SM goes from an enabled state (=1) to a disabled state (=0) or vice versa then all Vector (z, and subsequently v, b, h, s, d regs) and Predicate (p) registers are set to 0. Similarly, when ZA goes from an enabled state (=1) to a disabled state (=0) or vice versa then the za matrix register is zeroed out.

Currently within SimEng, on any execution of SMSTART or SMSTOP the register clearing functionality is performed for SM or ZA or both; which ever is appropriate for the instruction variation executed. However, this zeroing functionality should only be performed when the state changes from 1 to 0 or 0 to 1, not on any call.

For example, the trace below is what currently would happen in SimEng

SMSTART                    // Enter Streaming SVE mode. Zero all vector, fp, scalar, neon, predicate, and za matrix registers
dup z0.s, #3               // Duplicate the number 3 in all elements of z0 :  `z0.s = {3, 3, 3, 3}`
dup z1.s, #2               // Duplicate the number 2 in all elements of z1 : `z1.s = {2, 2, 2, 2}`
SMSTART                  // Enter Streaming SVE mode, zero all vector, fp, scalar, neon, predicate, and za matrix registers
add z2.s, z1.s, z0.s   // Add z0.s and z1.s element wise : `z2.s = {0, 0, 0, 0}`
SMSTOP                    // Exit Streaming mode. Zero all vector, fp, scalar, neon, predicate, and za matrix registers

However, the following should happen instead - effectively making the second SMSTART have no affect:

SMSTART                    // Enter Streaming SVE mode. Zero all vector, fp, scalar, neon, predicate, and za matrix registers
dup z0.s, #3               // Duplicate the number 3 in all elements of z0 :  `z0.s = {3, 3, 3, 3}`
dup z1.s, #2               // Duplicate the number 2 in all elements of z1 : `z1.s = {2, 2, 2, 2}`
SMSTART                  // Already in Streaming SVE mode. Do nothing
add z2.s, z1.s, z0.s   // Add z0.s and z1.s element wise : `z2.s = {5, 5, 5, 5}`
SMSTOP                    // Exit Streaming mode. Zero all vector, fp, scalar, neon, predicate, and za matrix registers

Whilst such a scenario (or a variation of it where single SMSTART and SMSTOP pairs aren't used in a code block) is unlikely, it would be a valid execution path and as such could cause an unexpected bug in the future.

[FinnWilkinson]

Issue fixed via commit e0bd27e into dev branch