On 6/12/21 3:21 AM, Philippe Mathieu-Daudé wrote:
On 6/1/21 5:00 PM, Richard Henderson wrote:
This removes all of the problems with unaligned accesses
to the bytecode stream.
With an 8-bit opcode at the bottom, we have 24 bits remaining,
which are generally split into 6 4-bit slots. This fits well
with the maximum length opcodes, e.g. INDEX_op_add2_i32, which
have 6 register operands.
We have, in previous patches, rearranged things such that there
are no operations with a label which have more than one other
operand. Which leaves us with a 20-bit field in which to encode
a label, giving us a maximum TB size of 512k -- easily large.
Change the INDEX_op_tci_movi_{i32,i64} opcodes to tci_mov[il].
The former puts the immediate in the upper 20 bits of the insn,
like we do for the label displacement. The later uses a label
to reference an entry in the constant pool. Thus, in the worst
case we still have a single memory reference for any constant,
but now the constants are out-of-line of the bytecode and can
be shared between different moves saving space.
Change INDEX_op_call to use a label to reference a pair of
pointers in the constant pool. This removes the only slightly
dodgy link with the layout of struct TCGHelperInfo.
The re-encode cannot be done in pieces.
Tested-by: Philippe Mathieu-Daudé <f4...@amsat.org>
Signed-off-by: Richard Henderson <richard.hender...@linaro.org>
---
include/tcg/tcg-opc.h | 4 +-
tcg/tci/tcg-target.h | 3 +-
tcg/tci.c | 541 +++++++++++++++------------------------
tcg/tci/tcg-target.c.inc | 379 ++++++++++++---------------
tcg/tci/README | 20 +-
5 files changed, 384 insertions(+), 563 deletions(-)
[ ... ]
+ case 0:
+ /* tcg_out_nop_fill uses zeros */
+ if (insn == 0) {
+ info->fprintf_func(info->stream, "align");
"nop"?
You can't execute it as a nop. It's just alignment.
Otherwise,
Reviewed-by: Philippe Mathieu-Daudé <f4...@amsat.org>
+ break;
+ }
+ /* fall through */
+
default:
info->fprintf_func(info->stream, "illegal opcode %d", op);
break;
}
