Am 05.09.2013 14:00, schrieb Jay Foad: >> diff --git a/tci.c b/tci.c >> index 18c888e..94b7851 100644 >> --- a/tci.c >> +++ b/tci.c >> @@ -952,8 +952,16 @@ uintptr_t tcg_qemu_tb_exec(CPUArchState *env, uint8_t >> *tb_ptr) >> break; >> #if TCG_TARGET_HAS_rot_i64 >> case INDEX_op_rotl_i64: >> + t0 = *tb_ptr++; >> + t1 = tci_read_ri64(&tb_ptr); >> + t2 = tci_read_ri64(&tb_ptr); >> + tci_write_reg64(t0, (t1 << t2) | (t1 >> (64 - t2))); >> + break; >> case INDEX_op_rotr_i64: >> - TODO(); >> + t0 = *tb_ptr++; >> + t1 = tci_read_ri64(&tb_ptr); >> + t2 = tci_read_ri64(&tb_ptr); >> + tci_write_reg64(t0, (t1 >> t2) | (t1 << (64 - t2))); > << (64 - t2) is undefined behaviour in C when t2 is 0. How about << (-t2 & > 63) ? > > Jay.
A short test confirms that the behaviour for (t1 << 64) is indeed unexpected. I added assertions for (t2 > 0) and (t2 < 64). They never raised an abort. Are those cases possible? We already have similar code for 32 bit shifts, and tcg/optimize.c also includes an implementation which is identical to my rotl_i64, rotr_i64. Therefore I think my patch can be applied as it is. Stefan
