Instruction cache bug in Pentium III?

I was getting some strange crashes in a program compiled under MSVC on
Windows 98, Pentium III processor. The following instruction turned
out not to be working as advertised:

8D 0C 80 LEA ECX,[EAX+EAX*4]

I could make this happen repeatedly, even single-stepping in the
debugger, and after trying a lot of input values I eventually deduced
that the instruction that was actually being *executed* was

8D 0C CC LEA ECX,[ESP+ECX*8]

-- in other words, the third byte was different from what the
disassembly display showed. The interesting thing is that the
instruction was at location 004C4ABE, which means that the scrambled
byte was at location 004C4AC0, which is a multiple of 64... which is
why I'm wondering about caches.

For the avoidance of doubt... although CC is INT 3 and is used to
implement debugger breakpoints, the same thing happened even if I put
a breakpoint somewhat later in the code, so the debugger isn't likely
to be responsible *directly*.

What is puzzling is that Visual Studio's disassembly display continued
to show 8D 0C 80 even though the CPU was executing 8D 0C CC.

Has anyone ever seen anything resembling this? Is it likely to be the
CPU, or something in the memory cache??

- Martin Kochanski.