o SHIFT_LOOP: Emit a loop using one (or two on H8S) bit shifts.
- Here are some thoughts on what the absolutely positively best code
- is. "Best" here means some rational trade-off between code size
- and speed, where speed is more preferred but not at the expense of
- generating 20 insns.
-
- Below, a trailing '*' after the shift count indicates the "best"
- mode isn't implemented. We only describe SHIFT_SPECIAL cases to
- simplify the table. For other cases, refer to shift_alg_[qhs]i.
-
- H8/300 QImode shifts
- 7 - ASHIFTRT: shll, subx (propagate carry bit to all bits)
-
- H8/300 HImode shifts
- 7 - shift 2nd half other way into carry.
- copy 1st half into 2nd half
- rotate 2nd half other way with carry
- rotate 1st half other way (no carry)
- mask off bits in 1st half (ASHIFT | LSHIFTRT).
- sign extend 1st half (ASHIFTRT)
- 8 - move byte, zero (ASHIFT | LSHIFTRT) or sign extend other (ASHIFTRT)
- 9-12 - do shift by 8, inline remaining shifts
- 15 - ASHIFTRT: shll, subx, set other byte
-
- H8/300 SImode shifts
- 7* - shift other way once, move bytes into place,
- move carry into place (possibly with sign extension)
- 8 - move bytes into place, zero or sign extend other
- 15* - shift other way once, move word into place, move carry into place
- 16 - move word, zero or sign extend other
- 24* - move bytes into place, zero or sign extend other
- 31 - ASHIFTRT: shll top byte, subx, copy to other bytes
-
- H8/300H QImode shifts (same as H8/300 QImode shifts)
- 7 - ASHIFTRT: shll, subx (propagate carry bit to all bits)
-
- H8/300H HImode shifts
- 7 - shift 2nd half other way into carry.
- copy 1st half into 2nd half
- rotate entire word other way using carry
- mask off remaining bits (ASHIFT | LSHIFTRT)
- sign extend remaining bits (ASHIFTRT)
- 8 - move byte, zero (ASHIFT | LSHIFTRT) or sign extend other (ASHIFTRT)
- 9-12 - do shift by 8, inline remaining shifts
- 15 - ASHIFTRT: shll, subx, set other byte
-
- H8/300H SImode shifts
- (These are complicated by the fact that we don't have byte level access to
- the top word.)
- A word is: bytes 3,2,1,0 (msb -> lsb), word 1,0 (msw -> lsw)
- 15* - shift other way once, move word into place, move carry into place
- (with sign extension for ASHIFTRT)
- 16 - move word into place, zero or sign extend other
- 17-20 - do 16bit shift, then inline remaining shifts
- 24* - ASHIFT: move byte 0(msb) to byte 1, zero byte 0,
- move word 0 to word 1, zero word 0
- LSHIFTRT: move word 1 to word 0, move byte 1 to byte 0,
- zero word 1, zero byte 1
- ASHIFTRT: move word 1 to word 0, move byte 1 to byte 0,
- sign extend byte 0, sign extend word 0
- 25-27* - either loop, or
- do 24 bit shift, inline rest
- 31 - shll, subx byte 0, sign extend byte 0, sign extend word 0
-
- H8S QImode shifts
- 7 - ASHIFTRT: shll, subx (propagate carry bit to all bits)
-
- H8S HImode shifts
- 8 - move byte, zero (ASHIFT | LSHIFTRT) or sign extend other (ASHIFTRT)
- 9-12 - do shift by 8, inline remaining shifts
- 15 - ASHIFTRT: shll, subx, set other byte
-
- H8S SImode shifts
- (These are complicated by the fact that we don't have byte level access to
- the top word.)
- A word is: bytes 3,2,1,0 (msb -> lsb), word 1,0 (msw -> lsw)
- 15* - shift other way once, move word into place, move carry into place
- (with sign extension for ASHIFTRT)
- 16 - move word into place, zero or sign extend other
- 17-20 - do 16bit shift, then inline remaining shifts
- 24* - ASHIFT: move byte 0(msb) to byte 1, zero byte 0,
- move word 0 to word 1, zero word 0
- LSHIFTRT: move word 1 to word 0, move byte 1 to byte 0,
- zero word 1, zero byte 1
- ASHIFTRT: move word 1 to word 0, move byte 1 to byte 0,
- sign extend byte 0, sign extend word 0
- 25-27* - either loop, or
- do 24 bit shift, inline rest
- 31 - shll, subx byte 0, sign extend byte 0, sign extend word 0
-
- Panic!!! */
+ For each shift count, we try to use code that has no trade-off
+ between code size and speed whenever possible.
+
+ If the trade-off is unavoidable, we try to be reasonable.
+ Specifically, the fastest version is one instruction longer than
+ the shortest version, we take the fastest version. We also provide
+ the use a way to switch back to the shortest version with -Os.
+
+ For the details of the shift algorithms for various shift counts,
+ refer to shift_alg_[qhs]i. */
int
nshift_operator (x, mode)
/* Given SHIFT_TYPE, SHIFT_MODE, and shift count COUNT, determine the
best algorithm for doing the shift. The assembler code is stored
- in the pointers in INFO. We don't achieve maximum efficiency in
- all cases, but the hooks are here to do so.
-
- For now we just use lots of switch statements. Since we don't even come
- close to supporting all the cases, this is simplest. If this function ever
- gets too big, perhaps resort to a more table based lookup. Of course,
- at this point you may just wish to do it all in rtl.
-
- WARNING: The constraints on insns shiftbyn_QI/HI/SI assume shifts of
- 1,2,3,4 will be inlined (1,2 for SI). */
+ in the pointers in INFO. We achieve the maximum efficiency in most
+ cases when !TARGET_H8300. In case of TARGET_H8300, shifts in
+ SImode in particular have a lot of room to optimize.
+
+ We first determine the strategy of the shift algorithm by a table
+ lookup. If that tells us to use a hand crafted assembly code, we
+ go into the big switch statement to find what that is. Otherwise,
+ we resort to a generic way, such as inlining. In either case, the
+ result is returned through INFO. */
static void
get_shift_alg (shift_type, shift_mode, count, info)
OSDN Git Service
