The bad situation reproduced as a miscompilation of some prof forms at *opt-level* 5 or above. The basic idea is that there is a situation like this prof t2 ... profiled code here producing value in t8 mov t2 t8 end t2 end t2 The code block produces a value in t8, which is copied into t2, and executes the end instruction. This instruction does not fall through to the next one but passes control back to the prof instruction. The prof instruction then stores the result value, which came from t2, back into the t2 register and resumes the program at the end t2. The first bad thing that happens is that the end instructions get merged together into one basic block. The optimizer then treats them without regard for the prof instruction, as if they were a linear sequence. It looks like the register move mov t2 t8 is wasteful and so it eliminates it, rewriting the end instruction to: end t8 end t8 Of course, the second instruction is now wrong because prof is still producing the result in t2. To fix this without changing the instruction set, I'm introducing another pseudo-op that represents end, called xend. This is similar to jend, except that jend is regarded as an unconditional branch whereas xend isn't. The special thing about xend is that a basic block in which it occcurs is marked as non-joinable. It will not be joined with the following basic block. * stdlib/asm.tl (xend): New alias opcode for end. * stdlib/compiler.tl (comp-prof): Use xend to end prof fragment, rather than plain end. * stdlib/optimize.tl (basic-block): New slot, nojoin. If true, block cannot be joined with next one. (basic-blocks jump-ops): Add xend to list of jump ops, so that a basic block will terminate on xend. (basic-blocks link-graph): Set the nojoin flag on a basic block which contains (and thus ends with) xend. (basic-blocks local-liveness): Add xend to the case in def-ref that handles end. (basic-blocks (peephole, join-blocks)): Refuse to join blocks marked nojoin. * tests/019/comp-bugs.tl: New file with miscompiled test case that was returning 42 instead of (42 0 0 0) as a result of the wrong register's value being returned.