[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [PATCH v4 62/64] tcg: Use ctpop to generate ctz if need
From: |
Richard Henderson |
Subject: |
Re: [Qemu-devel] [PATCH v4 62/64] tcg: Use ctpop to generate ctz if needed |
Date: |
Fri, 9 Dec 2016 08:48:49 -0800 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Thunderbird/45.5.1 |
On 12/09/2016 08:07 AM, Alex Bennée wrote:
>
> Richard Henderson <address@hidden> writes:
>
>> Particularly when andc is also available, this is two insns
>> shorter than using clz to compute ctz.
>>
>> Signed-off-by: Richard Henderson <address@hidden>
>> ---
>> tcg/tcg-op.c | 107
>> ++++++++++++++++++++++++++++++++++++-----------------------
>> 1 file changed, 65 insertions(+), 42 deletions(-)
>>
>> diff --git a/tcg/tcg-op.c b/tcg/tcg-op.c
>> index 6f4b1b6..d1debde 100644
>> --- a/tcg/tcg-op.c
>> +++ b/tcg/tcg-op.c
>> @@ -497,43 +497,46 @@ void tcg_gen_ctz_i32(TCGv_i32 ret, TCGv_i32 arg1,
>> TCGv_i32 arg2)
> <snip>
>> } else {
>> - gen_helper_ctz_i32(ret, arg1, arg2);
>> + TCGv_i32 z, t;
>> + if (TCG_TARGET_HAS_ctpop_i32 && TCG_TARGET_HAS_andc_i32) {
>> + t = tcg_temp_new_i32();
>> + tcg_gen_subi_i32(t, arg1, 1);
>> + tcg_gen_andc_i32(t, t, arg1);
>> + tcg_gen_ctpop_i32(t, t);
>> + do_movc:
>
> Hmmm and...
>
> <snip>
>> void tcg_gen_clrsb_i32(TCGv_i32 ret, TCGv_i32 arg)
>> @@ -1842,18 +1845,29 @@ void tcg_gen_ctz_i64(TCGv_i64 ret, TCGv_i64 arg1,
>> TCGv_i64 arg2)
>> {
>> if (TCG_TARGET_HAS_ctz_i64) {
>> tcg_gen_op3_i64(INDEX_op_ctz_i64, ret, arg1, arg2);
> <snip>
>> } else {
>> - gen_helper_ctz_i64(ret, arg1, arg2);
>> + TCGv_i64 z, t;
>> + if (TCG_TARGET_HAS_ctpop_i64 && TCG_TARGET_HAS_andc_i64) {
>> + t = tcg_temp_new_i64();
>> + tcg_gen_subi_i64(t, arg1, 1);
>> + tcg_gen_andc_i64(t, t, arg1);
>> + tcg_gen_ctpop_i64(t, t);
>> + do_movc:
>
> Hmmm.
>
> So I'm not a goto hater as it makes sense for a bunch of things. But
> this seems just a little too liberal usage to my eyes. What's wrong with
> a little extra nesting (seeing the compiler sorts it all out in the
> end):
>
> if ((TCG_TARGET_HAS_ctpop_i32 && TCG_TARGET_HAS_andc_i32)
> || TCG_TARGET_HAS_clz_i32 || TCG_TARGET_HAS_clz_i64) {
>
> TCGv_i32 z, t;
>
> if (TCG_TARGET_HAS_ctpop_i32 && TCG_TARGET_HAS_andc_i32) {
> t = tcg_temp_new_i32();
> tcg_gen_subi_i32(t, arg1, 1);
> tcg_gen_andc_i32(t, t, arg1);
> tcg_gen_ctpop_i32(t, t);
> } else if (TCG_TARGET_HAS_clz_i32 || TCG_TARGET_HAS_clz_i64) {
> /* Since all non-x86 hosts have clz(0) == 32, don't fight it.
> */
> t = tcg_temp_new_i32();
> tcg_gen_neg_i32(t, arg1);
> tcg_gen_and_i32(t, t, arg1);
> tcg_gen_clzi_i32(t, t, 32);
> tcg_gen_xori_i32(t, t, 31);
> }
> /* final movc */
> z = tcg_const_i32(0);
> tcg_gen_movcond_i32(TCG_COND_EQ, ret, arg1, z, arg2, t);
> tcg_temp_free_i32(t);
> tcg_temp_free_i32(z);
> } else {
> gen_helper_ctz_i32(ret, arg1, arg2);
> }
That does look better. Thanks,
r~