Re: middle-end/5100: Wrong register class selected by find_valid_class

["Marco Altieri" <maltieri at nergal dot it>]

In a simplified example we have only 4 registers (named 0, 1, 2, 3)
and 3 classes of registers:
EVEN_REGS = { 0, 2 }
ODD_REGS = { 1, 3 }
GENERAL_REGS = { 0, 1, 2, 3 }

A pointer can be reloaded only on even/odd pair. In this simplified
example the possible pair are only (0, 1) and (2, 3).

The macro HARD_REGNO_MODE_OK have to return a not null
value for the Pmode mode only for registers 0 and 2:
HARD_REGNO_MODE_OK(0, Pmode) = 1
HARD_REGNO_MODE_OK(1, Pmode) = 0
HARD_REGNO_MODE_OK(2, Pmode) = 1
HARD_REGNO_MODE_OK(3, Pmode) = 0

find_valid_class have to return GENERAL_REGS because EVEN_REGS
and ODD_REGS have not any even/odd pair.
But the algorithm, used in find_valid_class,  rejects GENERAL_REGS
and accepts EVEN_REGS.

Infact when
  (regno == 0 || regno == 2) &&
  m1 == Pmode &&
  class == GENERAL_REGISTER &&
  n == 1

we have:
  HARD_REGNO_MODE_OK (regno + n, m1) == 0
  TEST_HARD_REG_BIT (reg_class_contents[class], regno) == 1
  TEST_HARD_REG_BIT (reg_class_contents[class], regno + n) == 1

and when
  (regno == 0 || regno == 2) &&
  m1 == Pmode &&
  class == EVEN_REGS &&
  n == 1

we have:
  HARD_REGNO_MODE_OK (regno + n, m1) == 0
  TEST_HARD_REG_BIT (reg_class_contents[class], regno) == 1
  TEST_HARD_REG_BIT (reg_class_contents[class], regno + n) == 0


So, the algorithm in find_valid_class selects EVEN_REGS for a
mode == Pmode and n==1.

thanks, Marco Altieri