LowerSubregs, and fix an x86-64 isel bug that this exposed.
SUBREG_TO_REG for x86-64 implicit zero extension is only safe for
isel to generate when the source is known to always have zeros in
the high 32 bits. The EXTRACT_SUBREG instruction does not clear
the high 32 bits.
llvm-svn: 54444
subreg form on x86-64, to avoid the problem with x86-32
having GPRs that don't have 8-bit subregs.
Also, change several 16-bit instructions to use
equivalent 32-bit instructions. These have a smaller
encoding and avoid partial-register updates.
llvm-svn: 54223
to different address spaces. This alters the naming scheme for those
intrinsics, e.g., atomic.load.add.i32 => atomic.load.add.i32.p0i32
llvm-svn: 54195
which is represented in codegen as an 'and' operation. This matches them
with movz instructions, instead of leaving them to be matched by and
instructions with an immediate field.
llvm-svn: 54147
and knowledge of PseudoSourceValues. This unfortunately isn't sufficient to allow
constants to be rematerialized in PIC mode -- the extra indirection is a
complication.
llvm-svn: 54000
replacement of multiple values. This is slightly more efficient
than doing multiple ReplaceAllUsesOfValueWith calls, and theoretically
could be optimized even further. However, an important property of this
new function is that it handles the case where the source value set and
destination value set overlap. This makes it feasible for isel to use
SelectNodeTo in many very common cases, which is advantageous because
SelectNodeTo avoids a temporary node and it doesn't require CSEMap
updates for users of values that don't change position.
Revamp MorphNodeTo, which is what does all the work of SelectNodeTo, to
handle operand lists more efficiently, and to correctly handle a number
of corner cases to which its new wider use exposes it.
This commit also includes a change to the encoding of post-isel opcodes
in SDNodes; now instead of being sandwiched between the target-independent
pre-isel opcodes and the target-dependent pre-isel opcodes, post-isel
opcodes are now represented as negative values. This makes it possible
to test if an opcode is pre-isel or post-isel without having to know
the size of the current target's post-isel instruction set.
These changes speed up llc overall by 3% and reduce memory usage by 10%
on the InstructionCombining.cpp testcase with -fast and -regalloc=local.
llvm-svn: 53728
of all sizes from i1 to i256. The code is not
always that great, for example (x86)
movw %di, %ax
movw %ax, i17_s
where the store could be directly from %di.
llvm-svn: 53677
simply does the atomic.cmp.swap on the larger type,
which means it blows away whatever is sitting in
the bytes just after the memory location, i.e.
causes a buffer overflow. This really requires
target specific code, which is why LegalizeTypes
doesn't try to handle this case generically. The
existing (wrong) code in LegalizeDAG will go away
automatically once the type legalization code is
removed from LegalizeDAG so I'm leaving it there
for the moment. Meanwhile, don't test for this
feature.
llvm-svn: 53669
In LegalizeDAG the value is zero-extended to
the new type before byte swapping. It doesn't
matter how the extension is done since the new
bits are shifted off anyway after the swap, so
extend by any old rubbish bits. This results
in the final assembler for the testcase being
one line shorter.
llvm-svn: 53604
8 %reg1024<def> = IMPLICIT_DEF
12 %reg1024<def> = INSERT_SUBREG %reg1024<kill>, %reg1025, 2
The live range [12, 14) are not part of the r1024 live interval since it's defined by an implicit def. It will not conflicts with live interval of r1025. Now suppose both registers are spilled, you can easily see a situation where both registers are reloaded before the INSERT_SUBREG and both target registers that would overlap.
llvm-svn: 53503
getTargetNode and SelectNodeTo to reduce duplication, and to
make some of the getTargetNode code available to SelectNodeTo.
Use SelectNodeTo instead of getTargetNode in several new
interesting cases, as it mutates nodes in place instead of
creating new ones.
This triggers some scheduling behavior differences due to nodes
being presented to the scheduler in a different order. Some of the
arbitrary scheduling decisions it makes are now arbitrarily made
differently. This is visible in CodeGen/PowerPC/LargeAbsoluteAddr.ll,
where a trivial scheduling difference led to a trivial register
allocation difference.
llvm-svn: 53203
1. LSR runOnLoop is always returning false regardless if any transformation is made.
2. AddUsersIfInteresting can create new instructions that are added to DeadInsts. But there is a later early exit which prevents them from being freed.
llvm-svn: 53193
Added abstract class MemSDNode for any Node that have an associated MemOperand
Changed atomic.lcs => atomic.cmp.swap, atomic.las => atomic.load.add, and
atomic.lss => atomic.load.sub
llvm-svn: 52706
test (doesn't work for any MMX vector types, it's
not me). Rewritten to use v2i16 which is generic
and going to stay that way; I think that preserves
the point of the test.
llvm-svn: 52692
,------.
| |
| v
| t2 = phi ... t1 ...
| |
| v
| t1 = ...
| ... = ... t1 ...
| |
`------'
where there is a use in a PHI node that's a predecessor to the defining
block. We don't want to mark all predecessors as having the value "alive" in
this case. Also, the assert was too restrictive and didn't handle this case.
llvm-svn: 52655
shuffle could be skipped. The check is invalid because the loop index i
doesn't correspond to the element actually inserted. The correct check is
already done a few lines earlier, for whether the element is already in
the right spot, so this shouldn't have any effect on the codegen for
code that was already correct.
llvm-svn: 52486
wrong for volatile loads and stores. In fact this
is almost all of them! There are three types of
problems: (1) it is wrong to change the width of
a volatile memory access. These may be used to
do memory mapped i/o, in which case a load can have
an effect even if the result is not used. Consider
loading an i32 but only using the lower 8 bits. It
is wrong to change this into a load of an i8, because
you are no longer tickling the other three bytes. It
is also unwise to make a load/store wider. For
example, changing an i16 load into an i32 load is
wrong no matter how aligned things are, since the
fact of loading an additional 2 bytes can have
i/o side-effects. (2) it is wrong to change the
number of volatile load/stores: they may be counted
by the hardware. (3) it is wrong to change a volatile
load/store that requires one memory access into one
that requires several. For example on x86-32, you
can store a double in one processor operation, but to
store an i64 requires two (two i32 stores). In a
multi-threaded program you may want to bitcast an i64
to a double and store as a double because that will
occur atomically, and be indivisible to other threads.
So it would be wrong to convert the store-of-double
into a store of an i64, because this will become two
i32 stores - no longer atomic. My policy here is
to say that the number of processor operations for
an illegal operation is undefined. So it is alright
to change a store of an i64 (requires at least two
stores; but could be validly lowered to memcpy for
example) into a store of double (one processor op).
In short, if the new store is legal and has the same
size then I say that the transform is ok. It would
also be possible to say that transforms are always
ok if before they were illegal, whether after they
are illegal or not, but that's more awkward to do
and I doubt it buys us anything much.
However this exposed an interesting thing - on x86-32
a store of i64 is considered legal! That is because
operations are marked legal by default, regardless of
whether the type is legal or not. In some ways this
is clever: before type legalization this means that
operations on illegal types are considered legal;
after type legalization there are no illegal types
so now operations are only legal if they really are.
But I consider this to be too cunning for mere mortals.
Better to do things explicitly by testing AfterLegalize.
So I have changed things so that operations with illegal
types are considered illegal - indeed they can never
map to a machine operation. However this means that
the DAG combiner is more conservative because before
it was "accidentally" performing transforms where the
type was illegal because the operation was nonetheless
marked legal. So in a few such places I added a check
on AfterLegalize, which I suppose was actually just
forgotten before. This causes the DAG combiner to do
slightly more than it used to, which resulted in the X86
backend blowing up because it got a slightly surprising
node it wasn't expecting, so I tweaked it.
llvm-svn: 52254
don't fail when (expected) error output is produced. This fixes 17 tests.
While I was there, I also made all RUN lines of the form "not llvm-as..." a bit
more consistent, they now all redirect stderr and stdout to /dev/null and use
input redirect to read their input.
llvm-svn: 52174
and insertvalue and extractvalue instructions.
First-class array values are not trivial because C doesn't
support them. The approach I took here is to wrap all arrays
in structs. Feedback is welcome.
The 2007-01-15-NamedArrayType.ll test needed to be modified
because it has a "not grep" for a string that now exists,
because array types now have associated struct types, and
those struct types have names.
llvm-svn: 51881
in DAGISelEmitter output. This bug was recently uncovered by the
addition of patterns for CALL32m and CALL64m, which are nodes
that now have both MemOperands and variadic_ops.
This bug was especially visible with PIC in various configurations,
because the new patterns are matching the indirect call code used
in many PIC configurations.
llvm-svn: 51877
we did not truncate the value down to i1 with (x&1). This caused a problem
when the computation of x was nontrivial, for example, "add i1 1, 1" would
return 2 instead of 0.
This makes the testcase compile into:
...
llvm_cbe_t = (((llvm_cbe_r == 0u) + (llvm_cbe_r == 0u))&1);
llvm_cbe_u = (((unsigned int )(bool )llvm_cbe_t));
...
instead of:
...
llvm_cbe_t = ((llvm_cbe_r == 0u) + (llvm_cbe_r == 0u));
llvm_cbe_u = (((unsigned int )(bool )llvm_cbe_t));
...
This fixes a miscompilation of mediabench/adpcm/rawdaudio/rawdaudio and
403.gcc with the CBE, regressions from LLVM 2.2. Tanya, please pull
this into the release branch.
llvm-svn: 51813
cases due to an isel deficiency already noted in
lib/Target/X86/README.txt, but they can be matched in this fold-call.ll
testcase, for example.
This is interesting mainly because it exposes a tricky tblgen bug;
tblgen was incorrectly computing the starting index for variable_ops
in the case of a complex pattern.
llvm-svn: 51706
sometimes a "mov %ebp, %esp" in the epilogue.
Force these tests that rely on counting 'mov' to use i686-apple-darwin8.8.0
where they were written.
llvm-svn: 51568
BB1:
vr1025 = copy vr1024
..
BB2:
vr1024 = op
= op vr1025
<loop eventually branch back to BB1>
Even though vr1025 is copied from vr1024, it's not safe to coalesced them since live range of vr1025 intersects the def of vr1024. This happens when vr1025 is assigned the value of the previous iteration of vr1024 in the loop.
llvm-svn: 51394
If local spiller optimization turns some instruction into an identity copy, it will be removed. If the output register happens to be dead (and source is obviously killed), transfer the kill / dead information to last use / def in the same MBB.
llvm-svn: 51306
use-before-def. The problem comes up in code with multiple PHIs where
one PHI is being rewritten in terms of the other, but the other needs
to be casted first. LLVM rules requre the cast instruction to be
inserted after any PHI instructions, but when instructions were
inserted to replace the second PHI value with a function of the first,
they were ended up going before the cast instruction. Avoid this
problem by remembering the location of the cast instruction, when one
is needed, and inserting the expansion of the new value after it.
This fixes a bug that surfaced in 255.vortex on x86-64 when
instcombine was removed from the middle of the loop optimization
passes.
llvm-svn: 51169
%ecx = op
store %cl<kill>, (addr)
(addr) = op %al
It's not safe to unfold the last operand and eliminate store even though %cl is marked kill. It's a sub-register use which means one of its super-register(s) may be used below.
llvm-svn: 50794
Move platform independent code (lowering of possibly overwritten
arguments, check for tail call optimization eligibility) from
target X86ISelectionLowering.cpp to TargetLowering.h and
SelectionDAGISel.cpp.
Initial PowerPC tail call implementation:
Support ppc32 implemented and tested (passes my tests and
test-suite llvm-test).
Support ppc64 implemented and half tested (passes my tests).
On ppc tail call optimization is performed if
caller and callee are fastcc
call is a tail call (in tail call position, call followed by ret)
no variable argument lists or byval arguments
option -tailcallopt is enabled
Supported:
* non pic tail calls on linux/darwin
* module-local tail calls on linux(PIC/GOT)/darwin(PIC)
* inter-module tail calls on darwin(PIC)
If constraints are not met a normal call will be emitted.
A test checking the argument lowering behaviour on x86-64 was added.
llvm-svn: 50477
We now compile test2/test3 to:
_test2:
## InlineAsm Start
set %xmm0, %xmm1
## InlineAsm End
addps %xmm1, %xmm0
ret
_test3:
## InlineAsm Start
set %xmm0, %xmm1
## InlineAsm End
paddd %xmm1, %xmm0
ret
as expected.
llvm-svn: 50389
towards PR2094. It now compiles the attached .ll file to:
_sad16_sse2:
movslq %ecx, %rax
## InlineAsm Start
%ecx %rdx %rax %rax %r8d %rdx %rsi
## InlineAsm End
## InlineAsm Start
set %eax
## InlineAsm End
ret
which is pretty decent for a 3 output, 4 input asm.
llvm-svn: 50386
e.g.
vr1024<2> extract_subreg vr1025, 2
If vr1024 do not have the same register class as vr1025, it's not safe to coalesce this away. For example, vr1024 might be a GPR32 while vr1025 might be a GPR64.
llvm-svn: 50385
When choosing between constraints with multiple options,
like "ir", test to see if we can use the 'i' constraint and
go with that if possible. This produces more optimal ASM in
all cases (sparing a register and an instruction to load it),
and fixes inline asm like this:
void test () {
asm volatile (" %c0 %1 " : : "imr" (42), "imr"(14));
}
Previously we would dump "42" into a memory location (which
is ok for the 'm' constraint) which would cause a problem
because the 'c' modifier is not valid on memory operands.
Isn't it great how inline asm turns 'missed optimization'
into 'compile failed'??
Incidentally, this was the todo in
PowerPC/2007-04-24-InlineAsm-I-Modifier.ll
Please do NOT pull this into Tak.
llvm-svn: 50315
to the block that defines their operands. This doesn't work in the
case that the operand is an invoke, because invoke is a terminator
and must be the last instruction in a block.
Replace it with support in SelectionDAGISel for copying struct values
into sequences of virtual registers.
llvm-svn: 50279
On Darwin / Linux x86-32, v8i8, v4i16, v2i32 values are passed in MM[0-2].
On Darwin / Linux x86-32, v1i64 values are passed in memory.
On Darwin x86-64, v8i8, v4i16, v2i32 values are passed in XMM[0-7].
On Darwin x86-64, v1i64 values are passed in 64-bit GPRs.
llvm-svn: 50257
idea what this code (findNonImmUse) does, so I'm only guessing
that this is the right thing. It would be really really nice
if this had comments and perhaps switched to SmallPtrSet
(hint hint) :)
This fixes rdar://5886601, a crash on gcc.target/i386/sse4_1-pblendw.c
llvm-svn: 50252
argument. The x86-64 ABI requires the incoming value of %rdi to
be copied to %rax on exit from a function that is returning a
large C struct.
Also, add a README-X86-64 entry detailing the missed optimization
opportunity and proposing an alternative approach.
llvm-svn: 50075
memcpy lowering code; this ensures that the size node has the desired
result type. This fixes a regression from r49572 with @llvm.memcpy.i64
on x86-32.
llvm-svn: 49761
