Bootstrap classloader is stuck in a busy loop on x86-64 about jato

[penberg]

The -Xtrace:trampoline switch shows that we end up in:

[main] jit_magic_trampoline: ret0=0x41e7fea3, ret1=0x41e82ef4: java/util/zip/InflaterDynHeader.decode #2

which has an infinite loop:

public boolean decode(StreamManipulator input) throws DataFormatException
{
decode_loop:
  for (;;)
    {

Given that peekBits and dropBits from StreamManipulator are involved, my first guess is that it's a 32-bit vs 64-bit issue (possibly so that the higher 32-bits are not zeroed properly).

[penberg]

The problem is definitely related to 32-bit vs 64-bit issues. The following patch "fixes" the busy loop problem:

diff --git a/arch/x86/insn-selector_64.brg b/arch/x86/insn-selector_64.brg
index ce75c52..dc8d9fa 100644
--- a/arch/x86/insn-selector_64.brg
+++ b/arch/x86/insn-selector_64.brg
@@ -2275,7 +2275,7 @@ emulate_op_64(struct _MBState *state, struct basic_block *s,
        struct var_info *eax;

        eax = get_fixed_var(s->b_parent, MACH_REG_xAX);
-       state->reg1 = get_var(s->b_parent, J_INT);
+       state->reg1 = get_var(s->b_parent, J_LONG);

        rsi = get_fixed_var(s->b_parent, MACH_REG_RSI);
        select_insn(s, tree, reg_reg_insn(INSN_MOV_REG_REG, state->right->reg1, rsi));

but it breaks ControlTransferTest. A quick look reveals various potential places where we mix up 32-bit and 64-bit registers.