As I've mentioned before, the H8/300H can only shift a single bit
position at a time. Naturally this means many shifts are implemented as
loops. There's a variety of special cases that we can do without loops
by using rotates, sub-word moves, etc. The general guidance for the
port has been to only use inline or special sequences if they're shorter
or just one instruction longer than the loop.
This was pretty reasonable guidance for QI/HI mode. It was relaxed a
bit about 10 years ago for HImode in particular where the kpit team
realized they could save 50-100 cycles for some shifts by allowing 2
instructions of code growth over the loop implementation.
But they only re-tuned HImode shifts. There's even bigger benefits for
re-tuning SImode shifts. There's cases where we can save close to 200
cycles by allowing 2 additional instructions.
This patch re-tunes SImode shifts on the H8/300H primarily by inlining
more often or using a special sequence + inlining for residuals. Both
cases were already supported and this just uses those existing
capabilities more often, so it was trivial to implement. I think
there's some cases were entirely new special sequences could be used,
but I haven't tried those yet.
There'll be a similar follow-up for the H8/S. The gains aren't as
spectacular as the H8/S gained shift-by-2 instructions, but they should
still be significant.
Committed to the trunk after the usual testing and no regressions.
Jeff
commit 882f1d58bfa56737ff2de84c3cd1e0acfc318b86
Author: Jeff Law <jlaw@localhost.localdomain>
Date: Sun Aug 15 00:13:23 2021 -0400
Improve many SImode shifts on the H8/300H
As I've mentioned before, the H8/300H can only shift a single bit position
at a time. Naturally this means many shifts are implemented as loops. There's
a variety of special cases that we can do without loops by using rotates,
sub-word moves, etc. The general guidance for the port has been to only use
inline or special sequences if they're shorter or just one instruction longer
than the loop.
This was pretty reasonable guidance for QI/HI mode. It was relaxed a bit
about 10 years ago for HImode in particular where the kpit team realized they
could save 50-100 cycles for some shifts by allowing 2 instructions of code
growth over the loop implementation.
But they only re-tuned HImode shifts. There's even bigger benefits for
re-tuning SImode shifts. There's cases where we can save close to 200 cycles
by allowing 2 additional instructions.
This patch re-tunes SImode shifts on the H8/300H primarily by inlining more
often or using a special sequence + inlining for residuals. Both cases were
already supported and this just uses those existing capabilities more often, so
it was trivial to implement. I think there's some cases were entirely new
special sequences could be used, but I haven't tried those yet.
gcc/
* config/h8300/h8300.c (shift_alg_si): Retune H8/300H shifts
to allow a bit more code growth, saving many dozens of cycles.
(h8300_option_override): Adjus shift_alg_si if optimizing for
code size.
(get_shift_alg): Use special + inline shifts for residuals
in more cases.
diff --git a/gcc/config/h8300/h8300.c b/gcc/config/h8300/h8300.c
index d2f6548a265..7959ad1e276 100644
--- a/gcc/config/h8300/h8300.c
+++ b/gcc/config/h8300/h8300.c
@@ -228,18 +228,18 @@ static enum shift_alg shift_alg_si[2][3][32] = {
/* 8 9 10 11 12 13 14 15 */
/* 16 17 18 19 20 21 22 23 */
/* 24 25 26 27 28 29 30 31 */
- { INL, INL, INL, INL, INL, LOP, LOP, LOP,
+ { INL, INL, INL, INL, INL, INL, INL, LOP,
SPC, LOP, LOP, LOP, LOP, LOP, LOP, SPC,
- SPC, SPC, SPC, SPC, LOP, LOP, LOP, LOP,
- SPC, LOP, LOP, LOP, SPC, SPC, SPC, SPC }, /* SHIFT_ASHIFT */
- { INL, INL, INL, INL, INL, LOP, LOP, LOP,
+ SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC,
+ SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC }, /* SHIFT_ASHIFT */
+ { INL, INL, INL, INL, INL, INL, INL, LOP,
SPC, LOP, LOP, LOP, LOP, LOP, LOP, SPC,
- SPC, SPC, SPC, SPC, LOP, LOP, LOP, LOP,
- SPC, LOP, LOP, LOP, SPC, SPC, SPC, SPC }, /* SHIFT_LSHIFTRT */
- { INL, INL, INL, INL, INL, LOP, LOP, LOP,
+ 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, SPC, SPC, SPC, LOP, LOP, LOP, LOP,
- SPC, LOP, LOP, LOP, LOP, LOP, LOP, SPC }, /* SHIFT_ASHIFTRT */
+ SPC, SPC, SPC, SPC, SPC, SPC, SPC, SPC,
+ SPC, SPC, SPC, SPC, LOP, LOP, LOP, SPC }, /* SHIFT_ASHIFTRT */
},
{
/* TARGET_H8300S */
@@ -343,6 +343,36 @@ h8300_option_override (void)
shift_alg_hi[H8_300H][SHIFT_ASHIFTRT][13] = SHIFT_LOOP;
shift_alg_hi[H8_300H][SHIFT_ASHIFTRT][14] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFT][5] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFT][6] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFT][20] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFT][21] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFT][22] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFT][23] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFT][25] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFT][26] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFT][27] = SHIFT_LOOP;
+
+ shift_alg_si[H8_300H][SHIFT_LSHIFTRT][5] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_LSHIFTRT][6] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_LSHIFTRT][20] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_LSHIFTRT][21] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_LSHIFTRT][22] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_LSHIFTRT][23] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_LSHIFTRT][25] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_LSHIFTRT][26] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_LSHIFTRT][27] = SHIFT_LOOP;
+
+ shift_alg_si[H8_300H][SHIFT_ASHIFTRT][5] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFTRT][6] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFTRT][20] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFTRT][21] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFTRT][22] = SHIFT_LOOP;
+ shift_alg_si[H8_300H][SHIFT_ASHIFTRT][23] = SHIFT_LOOP;
+ 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;
+
/* H8S */
shift_alg_hi[H8_S][SHIFT_ASHIFTRT][14] = SHIFT_LOOP;
}
@@ -3784,7 +3814,7 @@ get_shift_alg (enum shift_type shift_type, enum
shift_mode shift_mode,
gcc_unreachable ();
}
}
- else if ((TARGET_H8300H && count >= 16 && count <= 19)
+ else if ((TARGET_H8300H && count >= 16 && count <= 23)
|| (TARGET_H8300S && count >= 16 && count <= 21))
{
info->remainder = count - 16;
@@ -3804,7 +3834,7 @@ get_shift_alg (enum shift_type shift_type, enum
shift_mode shift_mode,
goto end;
}
}
- else if ((TARGET_H8300H && count == 24)
+ else if ((TARGET_H8300H && count >= 24 || count <= 27)
|| (TARGET_H8300S && count >= 24 && count <= 25))
{
info->remainder = count - 24;
