The compiler's per-compile pad halfword is no evidence #649
Loading…
Reference in a new issue
No description provided.
Delete branch "rpg-9-ao-pad-halfword-20261028"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
The compiler's per-compile pad halfword is no evidence
A program with output specifications has a static initialization that stores
a halfword whose high byte the compiler takes from uninitialized memory. The
byte differs from one compile of a source to the next (0x57, 0x17, 0x34, 0x54
over five compiles of
r6e-const), so two compiles of one source compared asdifferent, and the 32 fixed O-specification rows of the loop could not reachmatchwhatever the emitter wrote.The byte appears twice. In the code it is
li rN,0xXX00andsth rN,D(r30);for XX of 0x80 and above,
rldicl rN,rN,0,48stands between them, which addsone word. In the literals it is the
XX 00pair of the initial value image.The comparison now treats the store as one pattern whatever the byte: the pair
counts as two instructions, the
rldiclword counts as none, and every lateroffset, branch destination and the range size count without it (the same
instruction counts and bytes on both sides). A store is that pattern only when its byte is corroborated: the
XX 00pair is the only one in the literals, and the literal byte is masked.With a second pair, or none (a store of zero, an ordinary constant), nothing is
located, the store stays ordinary code and its bytes stay evidence. Only the
exact form is normalized: the base register r30, a zero low byte of the immediate, a
rldiclof the same register. Another displacement or register, or any otherinstruction, still differs.
A constant that is also the only
XX 00pair of the literals (a field withinzof 4096 beside such a store) would still compare equal to a differentone of the same form; class-audit flags a lucky match.
RECOMPILE/r6e-constis a second compile of the publicr6e-constprobe; itscomparison with the saved twin is the regression test.
Gates