BUGS

1. No "RETURN" functionality.  So the output value in MSK is always
unchanged by definition.  The code doesn't allow BREAK at top level
(which would/should work to set a MSK output, but really isn't the
right solution).

------------------------------------------------------------------------

2. Copy propagation is pretty borked.  I'm seeing this in itermul and a
few other places, which should be caught by copy propagation but
isn't:
  R2    = MOV(R4/IN0) # R2 is unused elsewhere in the code!
  R37   = MOV(R2)

Ditto here, where R19 is needless:
  R19   = MOV(R5/IMM0)
  R34   = MOV(R19)

Likewise, in this situation R27 (again, unused except here) isn't
required, R37 can take the assignment itself.
  R27   = FSUB(R37, R8/IMM1)
  R37   = MOV(R27)
  R13   = FLE(R27, R5/IMM0)

And yet again here, this sequence at the bottom of itermul assigns the
same value to four different variables.  Only R33 and R25 are used
outside this sequence, so R29/R31 should be eliminated.
  R29   = ANDN(R33, R28)
{removed}
  BREAK()
ELSE()
  R31   = MOV(R29)
  R35   = MOV(R31)
  R33   = MOV(R31)
FI()
POOL(R29)

------------------------------------------------------------------------

3. MOV in loops doesn't hoist, even if the RHS is invariant and the
instruction comes before any BREAKs (i.e. it would always be
executed).  The reason is that after predication, the code is always
predicated on MSK, and MSK is variant.  Do a non-SSA hoist upstream
before predication?  LOAD/STORE has a similar issue: there the MSK
dependency is explicit, but they could still hoist if the other values
were invariant (and the instruction came before the first break).

------------------------------------------------------------------------

4. My god the MUX/BLENDPS argument order is a disaster.  How many
conventions do we have?  There are four registers: a destination, a
selector, the value placed in the destination if the selector is true,
and the value if it is false.

+ The mux() function looks like ternary:  DST = SEL ?     IF_TRUE : IF_FALSE
+ The MUX bytecode and avxgen API use:    DST,  IF_FALSE, IF_TRUE,  SEL
+ The Intel BLENDVPS spec looks like:     DST,  IF_TRUE,  IF_FALSE, SEL
+ The GNU binutils spits out:             SEL,  IF_TRUE,  IF_FALSE, DST

So the bytecode is just wrong, everything else looks like C, modulo
whether the desination or selector comes first (and the other last)