A few more shift-by-constant improvements on the H8.
For H8/300H arithmetic right shift 15 bits, we'd conceptually like to
use the shift-by-16 idiom where we move half-words around. Of course
that loses a bit. But we can save away that bit into C, shift-by-16,
sign-extend, then rotate through the carry one bit the other way. That
saves nearly 100 cycles and is the same size. We were already using
this concept elsewhere.
For H8/300H SImode arithmetic rightshift by 28, 29 or 30 bits we use the
shift-by-16 + shift-by-8 sequences, sign extend, then handle the
residuals inline. This saves on the order of 200 cycles vs a loop. THe
sequences are larger, but we're stil talking about burning at most 10
bytes of instruction space to save 200 cycles. Obviously with -Os we
revert to the loop sequence.
For H8/S HImode logical shift by 12 (right or left) we switch from using
a shift-by-8 idiom, clear bits, shift-by-2, shift-by-2 to a 2 rotate by
2 positions + masking sequence which is the same size/speed on the
hardware, but should simulate faster as it's one less instruction.
For H8/S SImode arithmetic right by 15 bits, we use the same trick as on
the H8/300H.
For H8/S SImode logical shifts by 27 bits we can shave off 2 more bytes
and 2 cycles by using a rotation based sequence over shift-by-16,
shift-by-8, shift-by-2, shift-by-1 style sequence.
This has gone through the usual simulator testing without regressions.
Jeff
commit 5f80c6270de6ac79d819de50048b32351a6b97c3
Author: Jeff Law <jlaw@localhost.localdomain>
Date: Fri Aug 20 11:19:05 2021 -0400
Further improvements to constant shifts for the H8
gcc/
* config/h8300/h8300.c (shift_alg_hi): Improve arithmetic shift
right
by 15 bits for H8/300H and H8/S. Improve logical shifts by 12
bits for H8/S.
(shift_alg_si): Improve arithmetic right shift by 28-30 bits for
H8/300H. Improve arithmetic shift right by 15 bits for H8/S.
Improve logical shifts by 27 bits for H8/S.
(get_shift_alg): Corresponding changes.
(h8300_option_override): Revert to loops for -Os when profitable.
diff --git a/gcc/config/h8300/h8300.c b/gcc/config/h8300/h8300.c
index 0c4e5089791..8ccacecba79 100644
--- a/gcc/config/h8300/h8300.c
+++ b/gcc/config/h8300/h8300.c
@@ -213,9 +213,9 @@ static enum shift_alg shift_alg_hi[2][3][16] = {
/* 0 1 2 3 4 5 6 7 */
/* 8 9 10 11 12 13 14 15 */
{ INL, INL, INL, INL, INL, INL, INL, INL,
- SPC, SPC, SPC, SPC, SPC, ROT, ROT, ROT }, /* SHIFT_ASHIFT */
+ SPC, SPC, SPC, SPC, ROT, ROT, ROT, ROT }, /* SHIFT_ASHIFT */
{ INL, INL, INL, INL, INL, INL, INL, INL,
- SPC, SPC, SPC, SPC, SPC, ROT, ROT, ROT }, /* SHIFT_LSHIFTRT */
+ SPC, SPC, SPC, SPC, ROT, ROT, ROT, ROT }, /* SHIFT_LSHIFTRT */
{ INL, INL, INL, INL, INL, INL, INL, INL,
SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC }, /* SHIFT_ASHIFTRT */
}
@@ -237,9 +237,9 @@ static enum shift_alg shift_alg_si[2][3][32] = {
SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC,
SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC }, /* SHIFT_LSHIFTRT */
{ INL, INL, INL, INL, INL, INL, INL, LOP,
- SPC, LOP, LOP, LOP, LOP, LOP, LOP, LOP,
+ SPC, LOP, LOP, LOP, LOP, LOP, LOP, SPC,
SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC,
- SPC, SPC, SPC, SPC, LOP, LOP, LOP, SPC }, /* SHIFT_ASHIFTRT */
+ SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC }, /* SHIFT_ASHIFTRT */
},
{
/* TARGET_H8300S */
@@ -256,7 +256,7 @@ static enum shift_alg shift_alg_si[2][3][32] = {
SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC,
SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC }, /* SHIFT_LSHIFTRT */
{ INL, INL, INL, INL, INL, INL, INL, INL,
- INL, INL, INL, INL, INL, INL, INL, LOP,
+ INL, INL, INL, INL, INL, INL, INL, SPC,
SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC,
SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC }, /* SHIFT_ASHIFTRT */
}
@@ -372,6 +372,9 @@ h8300_option_override (void)
shift_alg_si[H8_300H][SHIFT_ASHIFTRT][25] = SHIFT_LOOP;
shift_alg_si[H8_300H][SHIFT_ASHIFTRT][26] = SHIFT_LOOP;
shift_alg_si[H8_300H][SHIFT_ASHIFTRT][27] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFTRT][28] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFTRT][29] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFTRT][30] = SHIFT_LOOP;
/* H8S */
shift_alg_hi[H8_S][SHIFT_ASHIFTRT][14] = SHIFT_LOOP;
@@ -3830,6 +3833,10 @@ get_shift_alg (enum shift_type shift_type, enum
shift_mode shift_mode,
}
else if (count == 15)
{
+ /* The basic idea here is to use the shift-by-16 idiom to make things
+ small and efficient. Of course, that loses one bit that we need,
+ so we stuff the bit into C, shift by 16, then rotate the bit
+ back in. */
switch (shift_type)
{
case SHIFT_ASHIFT:
@@ -3841,7 +3848,9 @@ get_shift_alg (enum shift_type shift_type, enum
shift_mode shift_mode,
info->cc_special = OLD_CC_SET_ZNV;
goto end;
case SHIFT_ASHIFTRT:
- gcc_unreachable ();
+ info->special =
"shll.w\t%f0\n\tmov.w\t%e0,%f0\n\texts.l\t%S0\n\trotxl.l\t%S0";
+ info->cc_special = OLD_CC_SET_ZNV;
+ goto end;
}
}
else if (count >= 16 && count <= 23)
@@ -3863,6 +3872,23 @@ get_shift_alg (enum shift_type shift_type, enum
shift_mode shift_mode,
goto end;
}
}
+ else if (TARGET_H8300S && count == 27)
+ {
+ switch (shift_type)
+ {
+ case SHIFT_ASHIFT:
+ info->special =
"sub.w\t%e0,%e0\n\trotr.l\t#2,%S0\n\trotr.l\t#2,%S0\n\trotr.l\t%S0\n\tsub.w\t%f0,%f0";
+ goto end;
+ case SHIFT_LSHIFTRT:
+ info->special =
"sub.w\t%f0,%f0\n\trotl.l\t#2,%S0\n\trotl.l\t#2,%S0\n\trotl.l\t%S0\n\textu.l\t%S0";
+ goto end;
+ case SHIFT_ASHIFTRT:
+ info->remainder = count - 24;
+ info->special =
"mov.w\t%e0,%f0\n\tmov.b\t%t0,%s0\n\texts.w\t%f0\n\texts.l\t%S0";
+ info->cc_special = OLD_CC_SET_ZNV;
+ goto end;
+ }
+ }
else if (count >= 24 && count <= 27)
{
info->remainder = count - 24;
