We expect the backend to require register pairs in host-endian ordering, thus for big-endian the first register of a pair contains the high part. We were forcing R0 to contain the low part for calls.
Signed-off-by: Richard Henderson <richard.hender...@linaro.org> --- tcg/tci.c | 21 +++++++++++---------- 1 file changed, 11 insertions(+), 10 deletions(-) diff --git a/tcg/tci.c b/tcg/tci.c index 05a24163d3..eeccdde8bc 100644 --- a/tcg/tci.c +++ b/tcg/tci.c @@ -520,27 +520,28 @@ uintptr_t QEMU_DISABLE_CFI tcg_qemu_tb_exec(CPUArchState *env, ffi_call(pptr[1], pptr[0], stack, call_slots); } - /* Any result winds up "left-aligned" in the stack[0] slot. */ switch (len) { case 0: /* void */ break; case 1: /* uint32_t */ /* + * The result winds up "left-aligned" in the stack[0] slot. * Note that libffi has an odd special case in that it will * always widen an integral result to ffi_arg. */ - if (sizeof(ffi_arg) == 4) { - regs[TCG_REG_R0] = *(uint32_t *)stack; - break; - } - /* fall through */ - case 2: /* uint64_t */ - if (TCG_TARGET_REG_BITS == 32) { - tci_write_reg64(regs, TCG_REG_R1, TCG_REG_R0, stack[0]); + if (sizeof(ffi_arg) == 8) { + regs[TCG_REG_R0] = (uint32_t)stack[0]; } else { - regs[TCG_REG_R0] = stack[0]; + regs[TCG_REG_R0] = *(uint32_t *)stack; } break; + case 2: /* uint64_t */ + /* + * For TCG_TARGET_REG_BITS == 32, the register pair + * must stay in host memory order. + */ + memcpy(®s[TCG_REG_R0], stack, 8); + break; default: g_assert_not_reached(); } -- 2.34.1