Defining How to Split Instructions#
There are two cases where you should specify how to split a pattern into multiple insns. On machines that have instructions requiring delay slots (see Delay Slot Scheduling) or that have instructions whose output is not available for multiple cycles (see Specifying processor pipeline description), the compiler phases that optimize these cases need to be able to move insns into one-instruction delay slots. However, some insns may generate more than one machine instruction. These insns cannot be placed into a delay slot.
Often you can rewrite the single insn as a list of individual insns, each corresponding to one machine instruction. The disadvantage of doing so is that it will cause the compilation to be slower and require more space. If the resulting insns are too complex, it may also suppress some optimizations. The compiler splits the insn if there is a reason to believe that it might improve instruction or delay slot scheduling.
The insn combiner phase also splits putative insns. If three insns are
merged into one insn with a complex expression that cannot be matched by
some define_insn pattern, the combiner phase attempts to split
the complex pattern into two insns that are recognized. Usually it can
break the complex pattern into two patterns by splitting out some
subexpression. However, in some other cases, such as performing an
addition of a large constant in two insns on a RISC machine, the way to
split the addition into two insns is machine-dependent.
The define_split definition tells the compiler how to split a
complex insn into several simpler insns. It looks like this:
(define_split
[insn-pattern]
"condition"
[new-insn-pattern-1
new-insn-pattern-2
...]
"preparation-statements")
insn-pattern is a pattern that needs to be split and
condition is the final condition to be tested, as in a
define_insn. When an insn matching insn-pattern and
satisfying condition is found, it is replaced in the insn list
with the insns given by new-insn-pattern-1,
new-insn-pattern-2, etc.
The preparation-statements are similar to those statements that
are specified for define_expand (see Defining RTL Sequences for Code Generation)
and are executed before the new RTL is generated to prepare for the
generated code or emit some insns whose pattern is not fixed. Unlike
those in define_expand, however, these statements must not
generate any new pseudo-registers. Once reload has completed, they also
must not allocate any space in the stack frame.
There are two special macros defined for use in the preparation statements:
DONE and FAIL. Use them with a following semicolon,
as a statement.
- DONE#
Use the
DONEmacro to end RTL generation for the splitter. The only RTL insns generated as replacement for the matched input insn will be those already emitted by explicit calls toemit_insnwithin the preparation statements; the replacement pattern is not used.
- FAIL#
Make the
define_splitfail on this occasion. When adefine_splitfails, it means that the splitter was not truly available for the inputs it was given, and the input insn will not be split.
If the preparation falls through (invokes neither DONE nor
FAIL), then the define_split uses the replacement
template.
Patterns are matched against insn-pattern in two different
circumstances. If an insn needs to be split for delay slot scheduling
or insn scheduling, the insn is already known to be valid, which means
that it must have been matched by some define_insn and, if
reload_completed is nonzero, is known to satisfy the constraints
of that define_insn. In that case, the new insn patterns must
also be insns that are matched by some define_insn and, if
reload_completed is nonzero, must also satisfy the constraints
of those definitions.
As an example of this usage of define_split, consider the following
example from a29k.md, which splits a sign_extend from
HImode to SImode into a pair of shift insns:
(define_split
[(set (match_operand:SI 0 "gen_reg_operand" "")
(sign_extend:SI (match_operand:HI 1 "gen_reg_operand" "")))]
""
[(set (match_dup 0)
(ashift:SI (match_dup 1)
(const_int 16)))
(set (match_dup 0)
(ashiftrt:SI (match_dup 0)
(const_int 16)))]
"
{ operands[1] = gen_lowpart (SImode, operands[1]); }")
When the combiner phase tries to split an insn pattern, it is always the
case that the pattern is not matched by any define_insn.
The combiner pass first tries to split a single set expression
and then the same set expression inside a parallel, but
followed by a clobber of a pseudo-reg to use as a scratch
register. In these cases, the combiner expects exactly one or two new insn
patterns to be generated. It will verify that these patterns match some
define_insn definitions, so you need not do this test in the
define_split (of course, there is no point in writing a
define_split that will never produce insns that match).
Here is an example of this use of define_split, taken from
rs6000.md:
(define_split
[(set (match_operand:SI 0 "gen_reg_operand" "")
(plus:SI (match_operand:SI 1 "gen_reg_operand" "")
(match_operand:SI 2 "non_add_cint_operand" "")))]
""
[(set (match_dup 0) (plus:SI (match_dup 1) (match_dup 3)))
(set (match_dup 0) (plus:SI (match_dup 0) (match_dup 4)))]
"
{
int low = INTVAL (operands[2]) & 0xffff;
int high = (unsigned) INTVAL (operands[2]) >> 16;
if (low & 0x8000)
high++, low |= 0xffff0000;
operands[3] = GEN_INT (high << 16);
operands[4] = GEN_INT (low);
}")
Here the predicate non_add_cint_operand matches any
const_int that is not a valid operand of a single add
insn. The add with the smaller displacement is written so that it
can be substituted into the address of a subsequent operation.
An example that uses a scratch register, from the same file, generates an equality comparison of a register and a large constant:
(define_split
[(set (match_operand:CC 0 "cc_reg_operand" "")
(compare:CC (match_operand:SI 1 "gen_reg_operand" "")
(match_operand:SI 2 "non_short_cint_operand" "")))
(clobber (match_operand:SI 3 "gen_reg_operand" ""))]
"find_single_use (operands[0], insn, 0)
&& (GET_CODE (*find_single_use (operands[0], insn, 0)) == EQ
|| GET_CODE (*find_single_use (operands[0], insn, 0)) == NE)"
[(set (match_dup 3) (xor:SI (match_dup 1) (match_dup 4)))
(set (match_dup 0) (compare:CC (match_dup 3) (match_dup 5)))]
"
{
/* Get the constant we are comparing against, C, and see what it
looks like sign-extended to 16 bits. Then see what constant
could be XOR'ed with C to get the sign-extended value. */
int c = INTVAL (operands[2]);
int sextc = (c << 16) >> 16;
int xorv = c ^ sextc;
operands[4] = GEN_INT (xorv);
operands[5] = GEN_INT (sextc);
}")
To avoid confusion, don’t write a single define_split that
accepts some insns that match some define_insn as well as some
insns that don’t. Instead, write two separate define_split
definitions, one for the insns that are valid and one for the insns that
are not valid.
The splitter is allowed to split jump instructions into sequence of jumps or create new jumps in while splitting non-jump instructions. As the control flow graph and branch prediction information needs to be updated, several restriction apply.
Splitting of jump instruction into sequence that over by another jump
instruction is always valid, as compiler expect identical behavior of new
jump. When new sequence contains multiple jump instructions or new labels,
more assistance is needed. Splitter is required to create only unconditional
jumps, or simple conditional jump instructions. Additionally it must attach a
REG_BR_PROB note to each conditional jump. A global variable
split_branch_probability holds the probability of the original branch in case
it was a simple conditional jump, -1 otherwise. To simplify
recomputing of edge frequencies, the new sequence is required to have only
forward jumps to the newly created labels.
For the common case where the pattern of a define_split exactly matches the
pattern of a define_insn, use define_insn_and_split. It looks like
this:
(define_insn_and_split
[insn-pattern]
"condition"
"output-template"
"split-condition"
[new-insn-pattern-1
new-insn-pattern-2
...]
"preparation-statements"
[insn-attributes])
insn-pattern, condition, output-template, and
insn-attributes are used as in define_insn. The
new-insn-pattern vector and the preparation-statements are used as
in a define_split. The split-condition is also used as in
define_split, with the additional behavior that if the condition starts
with &&, the condition used for the split will be the constructed as a
logical ‘and’ of the split condition with the insn condition. For example,
from i386.md:
(define_insn_and_split "zero_extendhisi2_and"
[(set (match_operand:SI 0 "register_operand" "=r")
(zero_extend:SI (match_operand:HI 1 "register_operand" "0")))
(clobber (reg:CC 17))]
"TARGET_ZERO_EXTEND_WITH_AND && !optimize_size"
"#"
"&& reload_completed"
[(parallel [(set (match_dup 0)
(and:SI (match_dup 0) (const_int 65535)))
(clobber (reg:CC 17))])]
""
[(set_attr "type" "alu1")])
In this case, the actual split condition will be
TARGET_ZERO_EXTEND_WITH_AND && !optimize_size && reload_completed.
The define_insn_and_split construction provides exactly the same
functionality as two separate define_insn and define_split
patterns. It exists for compactness, and as a maintenance tool to prevent
having to ensure the two patterns’ templates match.
It is sometimes useful to have a define_insn_and_split
that replaces specific operands of an instruction but leaves the
rest of the instruction pattern unchanged. You can do this directly
with a define_insn_and_split, but it requires a
new-insn-pattern-1 that repeats most of the original insn-pattern.
There is also the complication that an implicit parallel in
insn-pattern must become an explicit parallel in
new-insn-pattern-1, which is easy to overlook.
A simpler alternative is to use define_insn_and_rewrite, which
is a form of define_insn_and_split that automatically generates
new-insn-pattern-1 by replacing each match_operand
in insn-pattern with a corresponding match_dup, and each
match_operator in the pattern with a corresponding match_op_dup.
The arguments are otherwise identical to define_insn_and_split :
(define_insn_and_rewrite
[insn-pattern]
"condition"
"output-template"
"split-condition"
"preparation-statements"
[insn-attributes])
The match_dup s and match_op_dup s in the new
instruction pattern use any new operand values that the
preparation-statements store in the operands array,
as for a normal define_insn_and_split. preparation-statements
can also emit additional instructions before the new instruction.
They can even emit an entirely different sequence of instructions and
use DONE to avoid emitting a new form of the original
instruction.
The split in a define_insn_and_rewrite is only intended
to apply to existing instructions that match insn-pattern.
split-condition must therefore start with &&,
so that the split condition applies on top of condition.
Here is an example from the AArch64 SVE port, in which operand 1 is known to be equivalent to an all-true constant and isn’t used by the output template:
(define_insn_and_rewrite "*while_ult<GPI:mode><PRED_ALL:mode>_cc"
[(set (reg:CC CC_REGNUM)
(compare:CC
(unspec:SI [(match_operand:PRED_ALL 1)
(unspec:PRED_ALL
[(match_operand:GPI 2 "aarch64_reg_or_zero" "rZ")
(match_operand:GPI 3 "aarch64_reg_or_zero" "rZ")]
UNSPEC_WHILE_LO)]
UNSPEC_PTEST_PTRUE)
(const_int 0)))
(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
(unspec:PRED_ALL [(match_dup 2)
(match_dup 3)]
UNSPEC_WHILE_LO))]
"TARGET_SVE"
"whilelo\t%0.<PRED_ALL:Vetype>, %<w>2, %<w>3"
;; Force the compiler to drop the unused predicate operand, so that we
;; don't have an unnecessary PTRUE.
"&& !CONSTANT_P (operands[1])"
{
operands[1] = CONSTM1_RTX (<MODE>mode);
}
)
The splitter in this case simply replaces operand 1 with the constant
value that it is known to have. The equivalent define_insn_and_split
would be:
(define_insn_and_split "*while_ult<GPI:mode><PRED_ALL:mode>_cc"
[(set (reg:CC CC_REGNUM)
(compare:CC
(unspec:SI [(match_operand:PRED_ALL 1)
(unspec:PRED_ALL
[(match_operand:GPI 2 "aarch64_reg_or_zero" "rZ")
(match_operand:GPI 3 "aarch64_reg_or_zero" "rZ")]
UNSPEC_WHILE_LO)]
UNSPEC_PTEST_PTRUE)
(const_int 0)))
(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
(unspec:PRED_ALL [(match_dup 2)
(match_dup 3)]
UNSPEC_WHILE_LO))]
"TARGET_SVE"
"whilelo\t%0.<PRED_ALL:Vetype>, %<w>2, %<w>3"
;; Force the compiler to drop the unused predicate operand, so that we
;; don't have an unnecessary PTRUE.
"&& !CONSTANT_P (operands[1])"
[(parallel
[(set (reg:CC CC_REGNUM)
(compare:CC
(unspec:SI [(match_dup 1)
(unspec:PRED_ALL [(match_dup 2)
(match_dup 3)]
UNSPEC_WHILE_LO)]
UNSPEC_PTEST_PTRUE)
(const_int 0)))
(set (match_dup 0)
(unspec:PRED_ALL [(match_dup 2)
(match_dup 3)]
UNSPEC_WHILE_LO))])]
{
operands[1] = CONSTM1_RTX (<MODE>mode);
}
)