Commit Graph

182 Commits

Author SHA1 Message Date
Evan Cheng c0af9c6478 Don't forget to promote xform function to an explicit node for def : Pat<>
patterns.

llvm-svn: 26929
2006-03-21 20:44:17 +00:00
Chris Lattner 02ad00ad93 minor code simplification
llvm-svn: 26918
2006-03-21 06:42:58 +00:00
Evan Cheng af7de1fba8 The node wrapped in PatLeaf<> should be treated as a leaf even if it isn't
one, i.e. don't select it.

llvm-svn: 26909
2006-03-20 22:53:06 +00:00
Evan Cheng 5ece6fa3e0 It should be ok for a xform output type to be different from input type.
llvm-svn: 26899
2006-03-20 08:09:17 +00:00
Evan Cheng a84bdebfd2 Copy matching pattern's output type info to instruction result pattern.
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
2006-03-20 06:04:09 +00:00
Chris Lattner c1b31d8a83 Add a new SDTCisIntVectorOfSameSize type constraint
llvm-svn: 26890
2006-03-20 05:39:48 +00:00
Evan Cheng c47620b5d8 Temporary hack to enable more (store (op (load ...))) folding. This makes
it possible when a TokenFactor is between the load and store. But is still
missing some cases due to ordering issue.

llvm-svn: 26638
2006-03-09 08:19:11 +00:00
Evan Cheng 0fab08eae3 Don't generate silly matching code like this:
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
2006-03-07 08:31:27 +00:00
Chris Lattner dc445eadc0 Select inline asm memory operands.
llvm-svn: 26349
2006-02-24 02:13:31 +00:00
Evan Cheng 33f4156663 Bump up pattern cost if the resulting instruction is marked
usesCustomDAGSchedInserter.

llvm-svn: 26282
2006-02-18 02:33:09 +00:00
Evan Cheng 5940bc210e Call InsertISelMapEntry rather than map insertion operator to prevent overly
aggrssive inlining. This reduces Select_store frame size from 24k to 10k.

llvm-svn: 26095
2006-02-09 22:12:27 +00:00
Evan Cheng 9ce762d51f Match getTargetNode() changes (now returns SDNode* instead of SDOperand).
llvm-svn: 26084
2006-02-09 07:16:09 +00:00
Evan Cheng 6dc90ca172 Change Select() from
SDOperand Select(SDOperand N);
to
void Select(SDOperand &Result, SDOperand N);

llvm-svn: 26067
2006-02-09 00:37:58 +00:00
Evan Cheng 1630c2848f Hoist all SDOperand declarations within a Select_{opcode}() to the top level
to reduce stack memory usage. This is intended to work around the gcc bug.

llvm-svn: 26026
2006-02-07 00:37:41 +00:00
Evan Cheng ebf05bea1b At the end of isel, select a replacement node for each handle that does not
have one. This can happen if a load's real uses are dead (i.e. they do not
have uses themselves).

llvm-svn: 26014
2006-02-06 08:12:55 +00:00
Evan Cheng 308d0090e6 Name change.
llvm-svn: 26013
2006-02-06 06:03:35 +00:00
Evan Cheng aa26555b25 Handle HANDLENODE: just return itself.
llvm-svn: 26011
2006-02-05 08:46:14 +00:00
Evan Cheng 368f20e53d Allow more loads to be folded which were previously prevented from happening
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
2006-02-05 06:43:12 +00:00
Evan Cheng 7ab4579c68 Re-committing the last bit of change. It shouldn't break PPC this time.
llvm-svn: 25982
2006-02-05 05:22:18 +00:00
Chris Lattner 586b06281c Temporarily revert the last change, which breaks PPC and other targets that
DO select things.

llvm-svn: 25970
2006-02-04 09:23:06 +00:00
Evan Cheng ce87cac555 Complex pattern's custom matcher should not call Select() on any operands.
Select them afterwards if it returns true.

llvm-svn: 25968
2006-02-04 08:50:49 +00:00
Chris Lattner 2bcfc52af6 node predicates add to the complexity of a pattern. This ensures that the
X86 backend attempts to match small-immediate versions of instructions before
the full size immediate versions.

llvm-svn: 25937
2006-02-03 18:06:02 +00:00
Evan Cheng 717db484a5 (store (op (load ...))) folding problem. In the generated matching code,
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
2006-02-03 06:22:41 +00:00
Evan Cheng 88e616d803 If a pattern's root node is a constant, its size should be 3 rather than 2.
llvm-svn: 25870
2006-02-01 06:06:31 +00:00
Chris Lattner a8821624d4 simplify some code
llvm-svn: 25791
2006-01-29 20:01:35 +00:00
Andrew Lenharth 1af077c0c7 it is nice not to chop off bits for those blessed with lots of bits
llvm-svn: 25766
2006-01-29 05:22:37 +00:00
Andrew Lenharth 49f323a90e make the casts actually cast to the variable type
llvm-svn: 25765
2006-01-29 05:17:22 +00:00
Andrew Lenharth c438c51e82 start of the 64bit safety cleanup
llvm-svn: 25764
2006-01-29 05:07:04 +00:00
Chris Lattner 0e352963fd Emit series of conditionals with &&, emitting stuff like this:
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
2006-01-29 04:41:05 +00:00
Chris Lattner 7c0149d1d6 Factor matching code that is common between patterns. This works around
GCC not jump-threading across this common code, and produces far nicer
output.

llvm-svn: 25762
2006-01-29 04:25:26 +00:00
Chris Lattner d3a63de91e Split out code generation from analysis from emission
llvm-svn: 25759
2006-01-29 02:57:39 +00:00
Chris Lattner afe9baeb69 move some code around, no change in the generated code
llvm-svn: 25758
2006-01-29 02:43:35 +00:00
Chris Lattner 5255c6c627 now that we have control over emission of the code, emit the code using nested
"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
2006-01-28 20:43:52 +00:00
Chris Lattner 4c4da201bf Change PatternCodeEmitter to emit code into a buffer instead of emitting it
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
2006-01-28 20:31:24 +00:00
Chris Lattner a7d6bbff00 Fit to 80 columns, no functionality change
llvm-svn: 25752
2006-01-28 19:06:51 +00:00
Jeff Cohen e251e92c23 Teach tablegen to generate code that is VC++ warning-free.
llvm-svn: 25709
2006-01-27 22:22:28 +00:00
Evan Cheng e555a92c8d (store (op (load ...)), ...)
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
2006-01-27 22:13:45 +00:00
Chris Lattner 5815669251 Teach the dag selectors to select InlineAsm nodes.
Aren't we happy the pattern selectors are almost all gone?

llvm-svn: 25666
2006-01-26 23:08:55 +00:00
Evan Cheng e9025f1f3b Another folding problem: if a node r/w chain or flag, don't fold it if it
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
2006-01-26 19:13:45 +00:00
Evan Cheng dc8365d4f3 Add a FIXME comment.
llvm-svn: 25635
2006-01-26 02:13:31 +00:00
Evan Cheng ecfaa0a1c9 Incoming (and optional) flag bugs. They may be embedded inside a inner node of
a pattern. Also, nodes which take incoming flag should not be folded if it has
more than one use.

llvm-svn: 25627
2006-01-26 00:22:25 +00:00
Evan Cheng c5c228fa59 Fix an optional in flag bug.
llvm-svn: 25590
2006-01-24 20:46:50 +00:00
Evan Cheng 295e196558 Optional InFlag was not being included in node.
llvm-svn: 25588
2006-01-24 20:07:38 +00:00
Evan Cheng e272b4ec17 Prevent folding of a node with multiple uses if the node already folds a load!
Here is an example where the load ended up being done twice:

%A = global uint 0

uint %test(uint %B, ubyte %C) {
	%tmp = load uint *%A;
	%X = shl uint %tmp, ubyte %C
	%Cv = sub ubyte 32, %C
	%Y = shr uint %B, ubyte %Cv
	%Z = or uint %Y, %X
	store uint %Z, uint* %A
	ret uint %Z
}

==>

	subl $4, %esp
	movl %ebx, (%esp)
	movl 8(%esp), %edx
	movl A, %eax
	movb 12(%esp), %bl
	movb %bl, %cl
	shldl %cl, %edx, %eax
	movb %bl, %cl
	shldl %cl, %edx, A
	movl (%esp), %ebx
	addl $4, %esp
	ret

llvm-svn: 25471
2006-01-20 01:11:03 +00:00
Evan Cheng 9e7fb7b2fc Bug fix. Flag operand number may be calculated incorrectly.
llvm-svn: 25465
2006-01-19 21:57:10 +00:00
Evan Cheng a15731cd50 Use pattern information to determine whether the use expects this
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
2006-01-19 10:12:58 +00:00
Evan Cheng acec02ebf5 Prevent unnecessary CopyToReg when the same HW register appears in two spots
in the pattern.

llvm-svn: 25437
2006-01-19 01:55:45 +00:00
Evan Cheng 61864ec3fe Emit a type matching check for ComplexPatterns.
llvm-svn: 25392
2006-01-17 07:36:41 +00:00
Evan Cheng a039d439dc Type inferencing bug
llvm-svn: 25337
2006-01-15 10:04:45 +00:00
Evan Cheng bd1de84121 Allow transformation from GlobalAddress to TargetGlobalAddress and
ExternalSymbol to TargetExternalSymbol.

llvm-svn: 25252
2006-01-12 19:35:54 +00:00