Minor bugs

When passing AND receiving arguments on the stack (win64 passes only 4 args in registers, and requires backing stack store), the RA math was off. Bug triggered need for a stack-stack spill, now done. Inlining also hides the bug, since there are no large-argument calls afterwards.
x86 store-immediate encoded an 8byte imm but the op only takes a 4byte imm.
Making a 1-reg regmask with reg>64 math wrong.
"popular constant" RA split would trigger when spanning calls; visit pattern for multi-users would then skip a user.
Encoding would pad fun headers to 16bytes, then expand x86 short jumps to long jumps, messing up the padding.
Call/End/Node visiting in RA would double-visit trailing projections.
GCM not treating Calls as memory kills.

0 files changed, 0 insertions, 0 deletions