Add runtime markers for a full-width 32x32 multiply (0x12345678 * 0x9abcdef0
= 0x242d2080) and for a wide-dividend remainder, so a 16-bit-truncated MPYS or
a wrong FS1 cannot still produce the expected small results. Compute the 16x16
unsigned product in unsigned arithmetic to remove the signed-overflow
annotation, and reject a regressed nuw nsw multiply in the optimized IR.
Add a runtime marker test for 32x32 multiply, the 16x16 MPYU path, signed and
unsigned divide/remainder, and for sumto/squares compiled through the ordinary
tms-clang -O2 driver. The linked assembly is also checked for MPYS/MPYU/
DIVS/DIVU and for a surviving sumto loop.
Scripts/tms-clang runs the generic opt pipeline without a target machine, so
the target cannot veto IndVarSimplify's exit-value replacement through TTI.
That rewrite turns sumto's closed form into an i33 multiply, which llc cannot
select. Disable only that rewrite with -replexitval=never and document why.
Preserve the dis34 output on a failed round-trip instead of claiming a
deleted temp file is kept, require GSPA "No Warnings" in addition to
"No Errors", and correct the six-function and M1 pipeline comments.
Raw clang IR has byte GEP offsets that llc treats as bits, so addresses
are 8x too small; always run tms-clang/tms-bitaddr --scale first.
Document that M1 supports only zero-initialized globals and rejects
.data/string literals loudly, and defer the spec's .data initializer
probe to M2.
--normalize lowers GEP offsets as already-scaled bit offsets. A typed
GEP still carries a DataLayout byte offset and a vector GEP has no
scalar lowering, so accepting either produced addresses 8x too small.
Check all instruction and constant-expression GEPs and fail clearly;
--scale still scales typed GEPs normally. Add positive and negative
tool tests.
arti