<16 x float> is 64-byte aligned (for some reason),
which gets us into the stack realignment code. The
computation changing FP-relative offsets to SP-relative
was broken, assiging a spill temp to a location
also used for parameter passing. This
fixes it by rounding up the stack frame to a multiple
of the largest alignment (I concluded it wasn't fixable
without doing this, but I'm not very sure.)
llvm-svn: 52750
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
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
of apint codegen failure is the DAG combiner doing
the wrong thing because it was comparing MVT's using
< rather than comparing the number of bits. Removing
the < method makes this mistake impossible to commit.
Instead, add helper methods for comparing bits and use
them.
llvm-svn: 52098
and better control the abstraction. Rename the type
to MVT. To update out-of-tree patches, the main
thing to do is to rename MVT::ValueType to MVT, and
rewrite expressions like MVT::getSizeInBits(VT) in
the form VT.getSizeInBits(). Use VT.getSimpleVT()
to extract a MVT::SimpleValueType for use in switch
statements (you will get an assert failure if VT is
an extended value type - these shouldn't exist after
type legalization).
This results in a small speedup of codegen and no
new testsuite failures (x86-64 linux).
llvm-svn: 52044
are the same as in unpacked structs, only field
positions differ. This only matters for structs
containing x86 long double or an apint; it may
cause backwards compatibility problems if someone
has bitcode containing a packed struct with a
field of one of those types.
The issue is that only 10 bytes are needed to
hold an x86 long double: the store size is 10
bytes, but the ABI size is 12 or 16 bytes (linux/
darwin) which comes from rounding the store size
up by the alignment. Because it seemed silly not
to pack an x86 long double into 10 bytes in a
packed struct, this is what was done. I now
think this was a mistake. Reserving the ABI size
for an x86 long double field even in a packed
struct makes things more uniform: the ABI size is
now always used when reserving space for a type.
This means that developers are less likely to
make mistakes. It also makes life easier for the
CBE which otherwise could not represent all LLVM
packed structs (PR2402).
Front-end people might need to adjust the way
they create LLVM structs - see following change
to llvm-gcc.
llvm-svn: 51928
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
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
index for the input pattern in terms of the output pattern. Instead
keep track of how many fixed operands the input pattern actually
has, and have the input matching code pass the output-emitting
function that index value. This simplifies the code, disentangles
variables_ops from the support for predication operations, and
makes variable_ops more robust.
llvm-svn: 51808
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
definitions. This adds a new construct, "discard", for indicating
that a named node in the input matching pattern is to be discarded,
instead of corresponding to a node in the output pattern. This
allows tblgen to know where the arguments for the varaible_ops are
supposed to begin.
This fixes "rdar://5791600", whatever that is ;-).
llvm-svn: 51699
instruction to execute. This can be used for transformations (like two-address
conversion) to remat an instruction instead of generating a "move"
instruction. The idea is to decrease the live ranges and register pressure and
all that jazz.
llvm-svn: 51660
code generator would do something like this:
f64 = load f32 <anyext>, f32mem
v2f64 = insertelt undef, %0, 0
v2f64 = insertelt %1, 0.0, 1
into
v2f64 = vzext_load f32mem
which on x86 is movsd, when you really wanted a cvtss2sd/movsd pair.
llvm-svn: 51624
are represented as "weak", but there are subtle differences
in some cases on Darwin, so we need both. The intent
is that "common" will behave identically to "weak" unless
somebody changes their target to do something else.
No functional change as yet.
llvm-svn: 51118
A brief description about PIC16:
===============================
PIC16 is an 8-bit microcontroller with only one 8-bit register which is the
accumulator. All arithmetic/load/store operations are 8-bit only.
The architecture has two address spaces: program and data. The program memory
is divided into 2K pages and the data memory is divided into banks of 128 byte, with only 80 usable bytes, resulting in an non-contiguous data memory.
It supports direct data memory access (by specifying the address as part of the instruction) and indirect data and program memory access (in an unorthodox fashion which utilize a 16 bit pointer register).
Two classes of registers exist: (8-bit class which is only one
accumulator) (16-bit class, which contains one or more 16 bit
pointer(s))
llvm-svn: 51027
on x86-64 linux. This causes no regressions on
32 bit linux and 32 bit ppc. More tests pass
on 64 bit ppc with no regressions. I didn't
turn on eh on 64 bit linux because the intrinsics
needed to compile the eh runtime aren't done
yet. But if you turn it on and link with the
mainline runtime then eh seems to work fine
on x86-64 linux with this patch. Thanks to
Dale for testing. The main point of the patch
is that if you output that some object is
encoded using 4 bytes you had better not output
8 bytes for it: the patch makes everything
consistent.
llvm-svn: 50825
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
fixes are target-specific lowering of frame indices, fix constants generated
for the FSMBI instruction, and fixing SPUTargetLowering::computeMaskedBitsFor-
TargetNode().
llvm-svn: 50462
- Make targetlowering.h fit in 80 cols.
- Make LowerAsmOperandForConstraint const.
- Make lowerXConstraint -> LowerXConstraint
- Make LowerXConstraint return a const char* instead of taking a string byref.
llvm-svn: 50312
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
Rename SDOperandImpl back to SDOperand.
Introduce the SDUse class that represents a use of the SDNode referred by
an SDOperand. Now it is more similar to Use/Value classes.
Patch is approved by Dan Gohman.
llvm-svn: 49795
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
optimized x86-64 (and x86) calls so that they work (... at least for
my test cases).
Should fix the following problems:
Problem 1: When i introduced the optimized handling of arguments for
tail called functions (using a sequence of copyto/copyfrom virtual
registers instead of always lowering to top of the stack) i did not
handle byval arguments correctly e.g they did not work at all :).
Problem 2: On x86-64 after the arguments of the tail called function
are moved to their registers (which include ESI/RSI etc), tail call
optimization performs byval lowering which causes xSI,xDI, xCX
registers to be overwritten. This is handled in this patch by moving
the arguments to virtual registers first and after the byval lowering
the arguments are moved from those virtual registers back to
RSI/RDI/RCX.
llvm-svn: 49584
on any current target and aren't optimized in DAGCombiner. Instead
of using intermediate nodes, expand the operations, choosing between
simple loads/stores, target-specific code, and library calls,
immediately.
Previously, the code to emit optimized code for these operations
was only used at initial SelectionDAG construction time; now it is
used at all times. This fixes some cases where rep;movs was being
used for small copies where simple loads/stores would be better.
This also cleans up code that checks for alignments less than 4;
let the targets make that decision instead of doing it in
target-independent code. This allows x86 to use rep;movs in
low-alignment cases.
Also, this fixes a bug that resulted in the use of rep;stos for
memsets of 0 with non-constant memory size when the alignment was
at least 4. It's better to use the library in this case, which
can be significantly faster when the size is large.
This also preserves more SourceValue information when memory
intrinsics are lowered into simple loads/stores.
llvm-svn: 49572
MOVZQI2PQIrr. This would be better handled as a dag combine
(with the goal of eliminating the bitconvert) but I don't know
how to do that safely. Thoughts welcome.
llvm-svn: 49463
review feedback.
-enable-eh is still accepted but doesn't do anything.
EH intrinsics use Dwarf EH if the target supports that,
and are handled by LowerInvoke otherwise.
The separation of the EH table and frame move data is,
I think, logically figured out, but either one still
causes full EH info to be generated (not sure how to
split the metadata correctly).
MachineModuleInfo::needsFrameInfo is no longer used and
is removed.
llvm-svn: 49064
not marked nounwind, or for all functions when -enable-eh
is set, provided the target supports Dwarf EH.
llvm-gcc generates nounwind in the right places; other FEs
will need to do so also. Given such a FE, -enable-eh should
no longer be needed.
llvm-svn: 49006
LLVM Value/Use does and MachineRegisterInfo/MachineOperand does.
This allows constant time for all uses list maintenance operations.
The idea was suggested by Chris. Reviewed by Evan and Dan.
Patch is tested and approved by Dan.
On normal use-cases compilation speed is not affected. On very big basic
blocks there are compilation speedups in the range of 15-20% or even better.
llvm-svn: 48822
other things, this allows the scheduler to unfold a load operand
in the 2008-01-08-SchedulerCrash.ll testcase, so it now successfully
clones the comparison to avoid a pushf+popf.
llvm-svn: 48777
Anton Korobeynikov