introduced by previous commit.
- SelectCode now returns a SDNode*. If it is not null, the selected node
produces the same number of results as the input node. The seletion loop
is responsible for calling ReplaceAllUsesWith() to replace the input node
with the output target node. For other cases, e.g. when load is folded,
the selection code is responsible for calling ReplaceAllUsesOfValueWith()
and SelectCode returns NULL.
- Other clean ups.
llvm-svn: 29602
in the start of an array and a count of operands where applicable. In many
cases, the number of operands is known, so this static array can be allocated
on the stack, avoiding the heap. In many other cases, a SmallVector can be
used, which has the same benefit in the common cases.
I updated a lot of code calling getNode that takes a vector, but ran out of
time. The rest of the code should be updated, and these methods should be
removed.
We should also do the same thing to eliminate the methods that take a
vector of MVT::ValueTypes.
It would be extra nice to convert the dagiselemitter to avoid creating vectors
for operands when calling getTargetNode.
llvm-svn: 29566
per possible ValueType of the node. e.g. Select_add is split into Select_add_i8,
Select_add_i16, etc.
For opcodes which do not produce a non-chain result, it is split on the
ValueType of its first non-chain operand. e.g. Select_store.
On X86 / Mac OS X, Select_store used to be the largest function. It had a stack
frame size of 8.5k. Now the largest one is Store_i32 with a frame size of 3.1k.
llvm-svn: 29404
instructions not handled would have a case value of #0, throwing things off.
This marginally shrinks the X86 asmprinter, but shrinks the sparc asmwriter
by 25 lines.
llvm-svn: 29187
series of identical commands, handle them all with one switch. In the case
of the x86 at&t asm printer, only 3 switches are needed for all instructions.
llvm-svn: 29184
based and less switch-statements-with-hundreds-of-cases based. This shrinks
the x86 asmprinters to about 1/3 their previous size.
Other improvements coming.
llvm-svn: 29177
code that emit target specific nodes into emit functions that are uniquified
and shared among selection routines.
e.g. This reduces X86ISelDAGToDAG.o (release) from ~2M to ~1.5M. Stack frame
size of Select_store from ~13k down to ~8k.
This is the first step. Further work to enable more sharing will follow.
llvm-svn: 29158
and index into it, instead of emitting it like this:
static const char * const OpStrs[] = {
"PHINODE\n", // PHI
0, // INLINEASM
"adc ", // ADC32mi
"adc ", // ADC32mi8
...
The old way required thousands of relocations that slows down link time and
dynamic load times.
This also cuts about 10K off each of the X86 asmprinters, and should shrink
the others as well.
llvm-svn: 29152
because information about one can help refine the other. This allows us to
write:
def : Pat<(i32 (extload xaddr:$src, i8)),
(LBZX xaddr:$src)>;
as:
def : Pat<(extload xaddr:$src, i8),
(LBZX xaddr:$src)>;
because tblgen knows LBZX returns i32.
llvm-svn: 28865
instruction, and the result type of the instruction to refine the pattern.
This allows us to write things like this:
def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
as:
def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (VR128:$src)>
and fixes a ppc64 issue.
llvm-svn: 28863
operands. e.g.
def CALL32r : I<0xFF, MRM2r, (ops GR32:$dst, variable_ops),
"call {*}$dst", [(X86call GR32:$dst)]>;
TableGen should emit operand informations for the "required" operands.
Added a target instruction info flag M_VARIABLE_OPS to indicate the target
instruction may have more operands in addition to the minimum required
operands.
llvm-svn: 28791
a cycle. This increase the search space and will increase compile time (in
practice it appears to be small, e.g. 176.gcc goes from 62 sec to 65 sec)
that will be addressed later.
llvm-svn: 28476
patterns that look like this:
def : Pat<(i32 (X86Wrapper tconstpool :$dst)), (MOV32ri tconstpool :$dst)>;
InsertOneTypeCheck should copy the type from the resolved pattern to the
unresolved one as long as there types are different.
llvm-svn: 28389
1. Use expects a chain output.
2. Node is expanded into multiple target ops.
3. One of the inner node produces a chain, the outer most one doesn't.
llvm-svn: 28209
x86 and ppc for 100% dense switch statements when relocations are non-PIC.
This support will be extended and enhanced in the coming days to support
PIC, and less dense forms of jump tables.
llvm-svn: 27947
validate the prototype of intrinsic functions. This prevents GCC from going
crazy and inlining too much stuff, eventually running out of memory.
llvm-svn: 27283
independently, batch up checks so that identically typed intrinsics share
verifier code. This dramatically reduces the size of the verifier function,
which should help avoid GCC running out of memory compiling Verifier.cpp.
llvm-svn: 27281
mismatch against the enum table.
This is a part of Sabre's master plan to drive me nuts with subtle bugs that
happens to only affect x86 be. :-)
llvm-svn: 27237
tblgen: In STVEBX: Intrinsic 'llvm.ppc.altivec.stvebx' expects 3 operands, not 2 operands!
instead of like this:
tblgen: In STVEBX: Intrinsic 'intrinsic_void expects 3 operands, not 2 operands!
llvm-svn: 27185
intrinsics that don't take pointer arguments now work. For example, we can
compile this:
int test3( __m128d *A) {
return _mm_movemask_pd(*A);
}
int test4( __m128 *A) {
return _mm_movemask_ps(*A);
}
to this:
_test3:
movl 4(%esp), %eax
movapd (%eax), %xmm0
movmskpd %xmm0, %eax
ret
_test4:
movl 4(%esp), %eax
movaps (%eax), %xmm0
movmskps %xmm0, %eax
ret
llvm-svn: 27090
The instruction patterns do not contain enough information to resolve the
exact type of the destination if it of a generic vector type.
llvm-svn: 26892
if (N1.getOpcode() == ISD::ADD &&
...)
if (... &&
(N1.getNumOperands() == 1 || !isNonImmUse(N1.Val, N10.Val))) &&
...)
TableGen knows N1 must have more than one operand.
llvm-svn: 26592
us to avoid creating lots of "Operand" types with different printers, instead
we can fold several together and use modifiers. For example, we can now use:
${target:call} to say that the operand should be printed like a 'call' operand.
llvm-svn: 26024
due to ordering issue. i.e. they were selected for chain use first.
Now at load select time, check if it is being selected for a chain use and if
it has only a single real use. If so, return a HANDLENODE (with the load as
its operand) in its place and record it.
When it is folded or the load is selected for a real use, the isel records it
as the replacement for the HANDLENODE. The replacement is done when all nodes
are selected.
This scheme exposed a couple of problems where cycles can happen. (See comments
in EmitMatchCode() for descriptions of the problems and their workaround /
solutions.) These problems have been resolved with a small compile time
penality.
llvm-svn: 25995
Chain is initially set to the chain operand of store node, when it reaches
load, if it matches the load then Chain is set to the chain operand of the
load.
However, if the matching code that follows this fails, isel moves on to the
next pattern but it does not restore Chain to the chain operand of the store.
So when it tries to match the next store / op / load pattern it would fail on
the Chain == load.getOperand(0) test.
The solution is for each chain operand to get a unique name. e.g. Chain10.
llvm-svn: 25931
if (N1.getOpcode() == ISD::LOAD &&
N1.hasOneUse() &&
!CodeGenMap.count(N1.getValue(0)) &&
!CodeGenMap.count(N1.getValue(1))) {
instead of this:
if (N1.getOpcode() == ISD::LOAD) {
if (N1.hasOneUse()) {
if (!CodeGenMap.count(N1.getValue(0))) {
if (!CodeGenMap.count(N1.getValue(1))) {
llvm-svn: 25763
"if" statements (indenting it appropriately, of course) instead of using goto's.
This inverts the logic for all of the if statements, which makes things simpler
to understand in addition to making the generated code easier to read.
llvm-svn: 25757
directly to the output file. This makes things simple because the code doesn't
have to worry about indentation or the case when there is no goto. It also
allows us to indent the code better without touching everything :)
llvm-svn: 25756
If store's chain operand is load, then use load's chain operand instead. If
it isn't (likely a TokenFactor), then do not allow the folding.
llvm-svn: 25708
get the order, don't compute it ourselves.
Don't emit stuff like (14<<0), emit 14 instead.
Don't attempt to get target properties for builtin instructions.
llvm-svn: 25672
has already been selected. The number of use check is not strong enough since
a node can be replaced with newly created target node. e.g. If the original
node has two uses, when it is selected for one of the uses it is replaced with
another. Each node now has a single use but isel still should not fold it.
llvm-svn: 25651
instruction to produce a result. e.g MUL8m, the instruction does not
produce a explicit result. However it produces an implicit result in
AL which would be copied to a temp. The root operator of the matching
pattern is a mul so the use would expect it to produce a result.
llvm-svn: 25458
SNDPOutFlag to DAG nodes. These properties do not belong to target specific
instructions.
* Added DAG node property SNDPOptInFlag. It's same as SNDPInFlag except it's
optional. Used by ret / call, etc.
llvm-svn: 25154
SDOperand Tmp0,Tmp1,Tmp2,Tmp3,;
GCC has a bug (24907) in which is fails to catch this, but VC++ correctly
notes its illegality, so tblgen must be taught to only generate legal C++.
llvm-svn: 25075
Currently tblgen cannot tell which operands in the operand list are results so
it assumes the first one is a result. This is bad. Ideally we would fix this
by separating results from inputs, e.g. (res R32:$dst),
(ops R32:$src1, R32:$src2). But that's a more distruptive change. Adding
'let noResults = 1' is the workaround to tell tblgen that the instruction does
not produces a result. It works for now since tblgen does not support
instructions which produce multiple results.
llvm-svn: 25017
use too much stack space, overflowing the stack for large functions. Instead
of emitting new SDOperands in each match block, we emit some common ones at
the top of SelectCode then reuse them when possible.
This reduces the stack size of SelectCode from 28K to 21K. Note that GCC
compiles it to 512 bytes :-/
I've filed GCC PR 25505 to track this.
llvm-svn: 24882
if (!N.Val->hasOneUse()) {
std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(N);
if (CGMI != CodeGenMap.end()) return CGMI->second;
}
Suppose a DAG like this:
X
^ ^
/ \
USE1 USE2
Suppose USE1 is being selected first and during which X is selected and
returned a new node. After this, USE1 is no longer an use of X. During USE2
selection, X will be selected again since it has only one use!
The fix is to always query CodeGenMap.
llvm-svn: 24679
it has more than one real use (non-chain uses).
* Record folded chain producing node in CodeGenMap.
* Do not fold a chain producing node if it has already been selected as an
operand of a chain use.
llvm-svn: 24647
matching code that is not currently auto-generated by tblgen, e.g. X86
addressing mode. Selection routines for complex patterns can return multiple operands, e.g. X86 addressing mode returns 4.
llvm-svn: 24634
* Enhanced tblgen to handle instructions which have chain operand and writes a
chain result.
* Enhanced tblgen to handle instructions which produces no results. Part of
the change is a temporary hack which relies on instruction property (e.g.
isReturn, isBranch). The proper fix would be to change the .td syntax to
separate results dag from ops dag.
llvm-svn: 24587
def SHL8rCL : I<0xD2, MRM4r, (ops R8 :$dst, R8 :$src),
"shl{b} {%cl, $dst|$dst, %CL}",
[(set R8:$dst, (shl R8:$src, CL))]>, Imp<[CL],[]>;
This generates a CopyToReg operand and added its 2nd result to the shl as
a flag operand.
llvm-svn: 24557
write things like this:
def : Pat<(add GPRC:$in, 12),
(ADD12 GPRC:$in)>;
Andrew: if this isn't enough or doesn't work for you, please lemme know.
llvm-svn: 23819
a pattern match, make sure to emit the (minimal number of) type checks that
verify the pattern matches this specific instruction. This allows FMA32
patterns to not match double expressions for example.
llvm-svn: 23748
type constraint. This lets tblgen realize that it doesn't need any dynamic
type checks for fextend/fround on PPC (and many other targets), because there
are only two fp types.
llvm-svn: 23730
1. If an operation has to be int or fp and the target only supports one
int or fp type, relize that the op has to have that type.
2. If a target has operations on multiple types, do not emit matching code
for patterns involving those operators, since we do not emit the code to
check for them yet. This prevents PPC from generating FP ops currently.
Also move some code around into more logical places.
llvm-svn: 23724
doesn't have to specify them manually. It currently handles associativity,
e.g. knowing that (X*Y)+Z also matches X+(Y*Z) and will be extended in
the future.
It is smart enough to not introduce duplicate patterns or patterns that can
never match.
llvm-svn: 23526
Currently we check that immediate values live on the RHS of commutative
operators. Defining ORI like this, for example:
def ORI : DForm_4<24, (ops GPRC:$dst, GPRC:$src1, u16imm:$src2),
"ori $dst, $src1, $src2",
[(set GPRC:$dst, (or immZExt16:$src2, GPRC:$src1))]>;
results in:
tblgen: In ORI: Instruction can never match: Immediate values must be on the RHS of commutative operators!
llvm-svn: 23501
This does not check that types match yet, but PPC only has one integer type
;-).
This also doesn't have the code to build the resultant dag.
llvm-svn: 23414
constraints defined in the DAG node definitions in the .td files. This
allows us to infer (and check!) the types for all nodes in the current
ppc .td file. For example, instead of:
Inst pattern EQV: (set GPRC:i32:$rT, (xor (xor GPRC:i32:$rA, GPRC:i32:$rB), (imm)<<Predicate_immAllOnes>>))
we now fully infer:
Inst pattern EQV: (set:void GPRC:i32:$rT, (xor:i32 (xor:i32 GPRC:i32:$rA, GPRC:i32:$rB), (imm:i32)<<Predicate_immAllOnes>>))
from: (set GPRC:$rT, (not (xor GPRC:$rA, GPRC:$rB)))
llvm-svn: 23284
progress. It correctly parses instructions and pattern fragments and glues
together pattern fragments into instructions.
The only code it generates currently is some boilerplate code for things
like the EntryNode.
llvm-svn: 23261
These changes modify the makefiles so that the output of flex and bison are
placed in the SRC directory, not the OBJ directory. It is intended that they
be checked in as any other LLVM source so that platforms without convenient
access to flex/bison can be compiled. From now on, if you change a .y or
.l file you *must* also commit the generated .cpp and .h files.
llvm-svn: 23115
anonymous regclass definition renaming.
Change the generated code to emit register classes as properly namespace
qualified entities like everything else.
expose the actual class definition as an object, though this isn't quite
usable yet.
llvm-svn: 22920
instruction defined, actually emit this to the InstrInfoDescriptor, which
allows an assert in the machineinstrbuilder to do some checking for us,
and is required by the dag->dag emitter
llvm-svn: 22895
LLVM is able to merge identical static const globals, GCC isn't, and this caused
some bloat in the generated data. This has a marginal effect on PPC, shrinking
the implicit sets from 10->4, but shrinks X86 from 179 to 23, a much bigger
reduction.
This should speed up the register allocator as well by reducing the dcache
footprint for this static data.
llvm-svn: 22879
printed as part of the opcode. This allows something like
cmp${cc}ss in the x86 backed to be printed as cmpltss, cmpless, etc.
depending on what the value of $cc is.
llvm-svn: 22439
finished up, only resolve fully when the def is defined. This allows things
to be changed and all uses to be propagated through. This implements
TableGen/LazyChange.td and fixes TemplateArgRename.td in the process.
None of the .td files used in LLVM backends are changed at all by this
patch.
llvm-svn: 21344
differences, which means that identical instructions (after stripping off
the first literal string) do not run any different code at all. On the X86,
this turns this code:
switch (MI->getOpcode()) {
case X86::ADC32mi: printOperand(MI, 4, MVT::i32); break;
case X86::ADC32mi8: printOperand(MI, 4, MVT::i8); break;
case X86::ADC32mr: printOperand(MI, 4, MVT::i32); break;
case X86::ADD32mi: printOperand(MI, 4, MVT::i32); break;
case X86::ADD32mi8: printOperand(MI, 4, MVT::i8); break;
case X86::ADD32mr: printOperand(MI, 4, MVT::i32); break;
case X86::AND32mi: printOperand(MI, 4, MVT::i32); break;
case X86::AND32mi8: printOperand(MI, 4, MVT::i8); break;
case X86::AND32mr: printOperand(MI, 4, MVT::i32); break;
case X86::CMP32mi: printOperand(MI, 4, MVT::i32); break;
case X86::CMP32mr: printOperand(MI, 4, MVT::i32); break;
case X86::MOV32mi: printOperand(MI, 4, MVT::i32); break;
case X86::MOV32mr: printOperand(MI, 4, MVT::i32); break;
case X86::OR32mi: printOperand(MI, 4, MVT::i32); break;
case X86::OR32mi8: printOperand(MI, 4, MVT::i8); break;
case X86::OR32mr: printOperand(MI, 4, MVT::i32); break;
case X86::ROL32mi: printOperand(MI, 4, MVT::i8); break;
case X86::ROR32mi: printOperand(MI, 4, MVT::i8); break;
case X86::SAR32mi: printOperand(MI, 4, MVT::i8); break;
case X86::SBB32mi: printOperand(MI, 4, MVT::i32); break;
case X86::SBB32mi8: printOperand(MI, 4, MVT::i8); break;
case X86::SBB32mr: printOperand(MI, 4, MVT::i32); break;
case X86::SHL32mi: printOperand(MI, 4, MVT::i8); break;
case X86::SHLD32mrCL: printOperand(MI, 4, MVT::i32); break;
case X86::SHR32mi: printOperand(MI, 4, MVT::i8); break;
case X86::SHRD32mrCL: printOperand(MI, 4, MVT::i32); break;
case X86::SUB32mi: printOperand(MI, 4, MVT::i32); break;
case X86::SUB32mi8: printOperand(MI, 4, MVT::i8); break;
case X86::SUB32mr: printOperand(MI, 4, MVT::i32); break;
case X86::TEST32mi: printOperand(MI, 4, MVT::i32); break;
case X86::TEST32mr: printOperand(MI, 4, MVT::i32); break;
case X86::TEST8mi: printOperand(MI, 4, MVT::i8); break;
case X86::XCHG32mr: printOperand(MI, 4, MVT::i32); break;
case X86::XOR32mi: printOperand(MI, 4, MVT::i32); break;
case X86::XOR32mi8: printOperand(MI, 4, MVT::i8); break;
case X86::XOR32mr: printOperand(MI, 4, MVT::i32); break;
}
into this:
switch (MI->getOpcode()) {
case X86::ADC32mi:
case X86::ADC32mr:
case X86::ADD32mi:
case X86::ADD32mr:
case X86::AND32mi:
case X86::AND32mr:
case X86::CMP32mi:
case X86::CMP32mr:
case X86::MOV32mi:
case X86::MOV32mr:
case X86::OR32mi:
case X86::OR32mr:
case X86::SBB32mi:
case X86::SBB32mr:
case X86::SHLD32mrCL:
case X86::SHRD32mrCL:
case X86::SUB32mi:
case X86::SUB32mr:
case X86::TEST32mi:
case X86::TEST32mr:
case X86::XCHG32mr:
case X86::XOR32mi:
case X86::XOR32mr: printOperand(MI, 4, MVT::i32); break;
case X86::ADC32mi8:
case X86::ADD32mi8:
case X86::AND32mi8:
case X86::OR32mi8:
case X86::ROL32mi:
case X86::ROR32mi:
case X86::SAR32mi:
case X86::SBB32mi8:
case X86::SHL32mi:
case X86::SHR32mi:
case X86::SUB32mi8:
case X86::TEST8mi:
case X86::XOR32mi8: printOperand(MI, 4, MVT::i8); break;
}
After this, the generated asmwriters look pretty much as though they were
generated by hand. This shrinks the X86 asmwriter.inc files from 55101->39669
and 55429->39551 bytes each, and PPC from 16766->12859 bytes.
llvm-svn: 19760
strings starts out with a constant string, we emit the string first, using
a table lookup (instead of a switch statement).
Because this is usually the opcode portion of the asm string, the differences
between the instructions have now been greatly reduced. This allows many
more case statements to be grouped together.
This patch also allows instruction cases to be grouped together when the
instruction patterns are exactly identical (common after the opcode string
has been ripped off), and when the differing operand is a MachineInstr
operand that needs to be formatted.
The end result of this is a mean and lean generated AsmPrinter!
llvm-svn: 19759
emitting code like this:
case PPC::ADD: O << "add "; printOperand(MI, 0, MVT::i64); O << ", "; prin
tOperand(MI, 1, MVT::i64); O << ", "; printOperand(MI, 2, MVT::i64); O << '\n
'; break;
case PPC::ADDC: O << "addc "; printOperand(MI, 0, MVT::i64); O << ", "; pr
intOperand(MI, 1, MVT::i64); O << ", "; printOperand(MI, 2, MVT::i64); O << '
\n'; break;
case PPC::ADDE: O << "adde "; printOperand(MI, 0, MVT::i64); O << ", "; pr
intOperand(MI, 1, MVT::i64); O << ", "; printOperand(MI, 2, MVT::i64); O << '
\n'; break;
...
Emit code like this:
case PPC::ADD:
case PPC::ADDC:
case PPC::ADDE:
...
switch (MI->getOpcode()) {
case PPC::ADD: O << "add "; break;
case PPC::ADDC: O << "addc "; break;
case PPC::ADDE: O << "adde "; break;
...
}
printOperand(MI, 0, MVT::i64);
O << ", ";
printOperand(MI, 1, MVT::i64);
O << ", ";
printOperand(MI, 2, MVT::i64);
O << "\n";
break;
This shrinks the PPC asm writer from 24785->15205 bytes (even though the new
asmwriter has much more whitespace than the old one), and the X86 printers shrink
quite a bit too. The important implication of this is that GCC no longer hits swap
when building the PPC backend in optimized mode. Thus this fixes PR448.
-Chris
llvm-svn: 19755
and more understandable. It also allows us to do simple things like fold
consequtive literal strings together. For example, instead of emitting this
for the X86 backend:
O << "adc" << "l" << " ";
we now generate this:
O << "adcl ";
*whoa* :)
This shrinks the X86 asmwriters from 62729->58267 and 65176->58644 bytes
for the intel/att asm writers respectively.
llvm-svn: 19749
