target/i386: improve code generation for BT

target/i386: improve code generation for BT

Because BT does not write back to the source operand, it can modify it to
ensure that one of the operands of TSTNE is a constant (after either gen_BT
or the optimizer's constant propagation).  This produces better and more
optimizable TCG ops.  For example, the sequence

  movl $0x60013f, %ebx
  btl %ecx, %ebx

becomes just

  and_i32 tmp1,ecx,$0x1f                   dead: 1 2  pref=0xffff
  shr_i32 tmp0,$0x60013f,tmp1              dead: 1 2  pref=0xffff
  and_i32 tmp16,tmp0,$0x1                  dead: 1  pref=0xbf80

On s390x, it can use four instructions to isolate bit 0 of 0x60013f >> (ecx & 31):

  nilf     %r12, 0x1f
  lgfi     %r11, 0x60013f
  srlk     %r12, %r11, 0(%r12)
  nilf     %r12, 1

Previously, it used five instructions to build 1 << (ecx & 31) and compute
TSTEQ, and also needed two more to construct the result of setcond:

  nilf     %r12, 0x1f
  lghi     %r11, 1
  sllk     %r12, %r11, 0(%r12)
  lgfi     %r9, 0x60013f
  nrk      %r0, %r12, %r9
  lghi     %r12, 0
  locghilh %r12, 1

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>