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
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
ScheduleDAG; they don't correspond to any actual instructions so they
don't need to be scheduled.
This fixes a bug where the EntryToken was being scheduled multiple
times in some cases, though it ended up not causing any trouble because
EntryToken doesn't expand into anything. With this fixed the schedulers
reliably schedule the expected number of units, so we can check this
with an assertion.
This requires a tweak to test/CodeGen/X86/loop-hoist.ll because it
ends up getting scheduled differently in a trivial way, though it was
enough to fool the prcontext+grep that the test does.
llvm-svn: 49701
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
EH info for these functions causes the tests to fail for
random reasons (e.g. looking for 'or' or counting lines
with asm-printer; labels count as lines.)
llvm-svn: 49003
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
This allows us to compile fp-stack-2results.ll into:
_test:
fldz
fld1
ret
which returns 1 in ST(0) and 0 in ST(1). This is needed for x86-64
_Complex long double.
llvm-svn: 48632
1. If part of a register is re-defined, an implicit kill and an implicit def are added to denote read / mod / write. However, this should only be necessary if the register is actually read later. This is a performance issue.
2. If a sub-register is being defined, and it doesn't have a previous use, do not add a implicit kill to the last use of a super-register:
= EAX, AX<imp-use,kill>
...
AX =
In this case, EAX is live but AX is killed, this is wrong and will cause the coalescer to do bad things.
llvm-svn: 48521
If ALR and BLR overlaps and end of BLR extends beyond end of ALR, e.g.
A = or A, B
...
B = A
...
C = A<kill>
...
= B
then do not add kills of A to the newly created B interval.
- Also fix some kill info update bug.
llvm-svn: 48141
an RFP register class.
Teach ScheduleDAG how to handle CopyToReg with different src/dst
reg classes.
This allows us to compile trivial inline asms that expect stuff
on the top of x87-fp stack.
llvm-svn: 48107
in different register classes, e.g. copy of ST(0) to RFP*. This gets
some really trivial inline asm working that plops things on the top of
stack (PR879)
llvm-svn: 48105
into a vector of zeros or undef, and when the top part is obviously
zero, we can just use movd + shuffle. This allows us to compile
vec_set-B.ll into:
_test3:
movl $1234567, %eax
andl 4(%esp), %eax
movd %eax, %xmm0
ret
instead of:
_test3:
subl $28, %esp
movl $1234567, %eax
andl 32(%esp), %eax
movl %eax, (%esp)
movl $0, 4(%esp)
movq (%esp), %xmm0
addl $28, %esp
ret
llvm-svn: 48090
_test3:
movd %rdi, %xmm1
#IMPLICIT_DEF %xmm0
punpcklqdq %xmm1, %xmm0
ret
instead of:
_test3:
#IMPLICIT_DEF %rax
movd %rax, %xmm0
movd %rdi, %xmm1
punpcklqdq %xmm1, %xmm0
ret
This is still not ideal. There is no reason to two xmm regs.
llvm-svn: 48058
except ppc long double. This allows us to shrink constant pool
entries for x86 long double constants, which in turn allows us to
use flds/fldl instead of fldt.
llvm-svn: 47938
For x86, if sse2 is available, it's not a good idea since cvtss2sd is slower than a movsd load and it prevents load folding. On x87, it's important to shrink fp constant since fldt is very expensive.
llvm-svn: 47931
stack slot and store if the SINT_TO_FP is actually legal. This allows
us to compile:
double a(double b) {return (unsigned)b;}
to:
_a:
cvttsd2siq %xmm0, %rax
movl %eax, %eax
cvtsi2sdq %rax, %xmm0
ret
instead of:
_a:
subq $8, %rsp
cvttsd2siq %xmm0, %rax
movl %eax, %eax
cvtsi2sdq %rax, %xmm0
addq $8, %rsp
ret
crazy.
llvm-svn: 47660
_test:
movl %edi, %eax
ret
instead of:
_test:
movl $4294967295, %ecx
movq %rdi, %rax
andq %rcx, %rax
ret
It would be great to write this as a Pat pattern that used subregs
instead of a 'pseudo' instruction, but I don't know how to do that
in td files.
llvm-svn: 47658
instead of with mmx registers. This horribleness is apparently
done by gcc to avoid having to insert emms in places that really
should have it. This is the second half of rdar://5741668.
llvm-svn: 47474
GCC apparently does this, and code depends on not having to do
emms when this happens. This is x86-64 only so far, second half
should handle x86-32.
rdar://5741668
llvm-svn: 47470
inline asms.
Fix PR2078 by marking aliases of registers used when a register is
marked used. This prevents EAX from being allocated when AX is listed
in the clobber set for the asm.
llvm-svn: 47426
has plain one-result scalar integer multiplication instructions.
This avoids expanding such instructions into MUL_LOHI sequences that
must be special-cased at isel time, and avoids the problem with that
code that provented memory operands from being folded.
This fixes PR1874, addressesing the most common case. The uncommon
cases of optimizing multiply-high operations will require work
in DAGCombiner.
llvm-svn: 47277
node as soon as we create it in SDISel. Previously we would lower it in
legalize. The problem with this is that it only exposes the argument
loads implied by FORMAL_ARGUMENTs after legalize, so that only dag combine 2
can hack on them. This causes us to miss some optimizations because
datatype expansion also happens here.
Exposing the loads early allows us to do optimizations on them. For example
we now compile arg-cast.ll to:
_foo:
movl $2147483647, %eax
andl 8(%esp), %eax
ret
where we previously produced:
_foo:
subl $12, %esp
movsd 16(%esp), %xmm0
movsd %xmm0, (%esp)
movl $2147483647, %eax
andl 4(%esp), %eax
addl $12, %esp
ret
It might also make sense to do this for ISD::CALL nodes, which have implicit
stores on many targets.
llvm-svn: 47054
registers if used by a bitconvert or using a bitconvert. This allows us to
avoid constant pool loads and use cheaper integer instructions when the
values come from or end up in integer regs anyway. For example, we now
compile CodeGen/X86/fp-in-intregs.ll to:
_test1:
movl $2147483648, %eax
xorl 4(%esp), %eax
ret
_test2:
movl $1065353216, %eax
orl 4(%esp), %eax
andl $3212836864, %eax
ret
Instead of:
_test1:
movss 4(%esp), %xmm0
xorps LCPI2_0, %xmm0
movd %xmm0, %eax
ret
_test2:
movss 4(%esp), %xmm0
andps LCPI3_0, %xmm0
movss LCPI3_1, %xmm1
andps LCPI3_2, %xmm1
orps %xmm0, %xmm1
movd %xmm1, %eax
ret
bitconverts can happen due to various calling conventions that require
fp values to passed in integer regs in some cases, e.g. when returning
a complex.
llvm-svn: 46414
This case returns the value in ST(0) and then has to convert it to an SSE
register. This causes significant codegen ugliness in some cases. For
example in the trivial fp-stack-direct-ret.ll testcase we used to generate:
_bar:
subl $28, %esp
call L_foo$stub
fstpl 16(%esp)
movsd 16(%esp), %xmm0
movsd %xmm0, 8(%esp)
fldl 8(%esp)
addl $28, %esp
ret
because we move the result of foo() into an XMM register, then have to
move it back for the return of bar.
Instead of hacking ever-more special cases into the call result lowering code
we take a much simpler approach: on x86-32, fp return is modeled as always
returning into an f80 register which is then truncated to f32 or f64 as needed.
Similarly for a result, we model it as an extension to f80 + return.
This exposes the truncate and extensions to the dag combiner, allowing target
independent code to hack on them, eliminating them in this case. This gives
us this code for the example above:
_bar:
subl $12, %esp
call L_foo$stub
addl $12, %esp
ret
The nasty aspect of this is that these conversions are not legal, but we want
the second pass of dag combiner (post-legalize) to be able to hack on them.
To handle this, we lie to legalize and say they are legal, then custom expand
them on entry to the isel pass (PreprocessForFPConvert). This is gross, but
less gross than the code it is replacing :)
This also allows us to generate better code in several other cases. For
example on fp-stack-ret-conv.ll, we now generate:
_test:
subl $12, %esp
call L_foo$stub
fstps 8(%esp)
movl 16(%esp), %eax
cvtss2sd 8(%esp), %xmm0
movsd %xmm0, (%eax)
addl $12, %esp
ret
where before we produced (incidentally, the old bad code is identical to what
gcc produces):
_test:
subl $12, %esp
call L_foo$stub
fstpl (%esp)
cvtsd2ss (%esp), %xmm0
cvtss2sd %xmm0, %xmm0
movl 16(%esp), %eax
movsd %xmm0, (%eax)
addl $12, %esp
ret
Note that we generate slightly worse code on pr1505b.ll due to a scheduling
deficiency that is unrelated to this patch.
llvm-svn: 46307
1. Legalize now always promotes truncstore of i1 to i8.
2. Remove patterns and gunk related to truncstore i1 from targets.
3. Rename the StoreXAction stuff to TruncStoreAction in TLI.
4. Make the TLI TruncStoreAction table a 2d table to handle from/to conversions.
5. Mark a wide variety of invalid truncstores as such in various targets, e.g.
X86 currently doesn't support truncstore of any of its integer types.
6. Add legalize support for truncstores with invalid value input types.
7. Add a dag combine transform to turn store(truncate) into truncstore when
safe.
The later allows us to compile CodeGen/X86/storetrunc-fp.ll to:
_foo:
fldt 20(%esp)
fldt 4(%esp)
faddp %st(1)
movl 36(%esp), %eax
fstps (%eax)
ret
instead of:
_foo:
subl $4, %esp
fldt 24(%esp)
fldt 8(%esp)
faddp %st(1)
fstps (%esp)
movl 40(%esp), %eax
movss (%esp), %xmm0
movss %xmm0, (%eax)
addl $4, %esp
ret
llvm-svn: 46140
and the spill is its kill. However, if the local allocator has determined the
register has not been modified (possible when its value was reloaded), it would
not issue a restore. In that case, mark the last use of the virtual register as
kill.
llvm-svn: 46111
It's not safe to use the two value CombineTo variant to combine away a dead load.
e.g.
v1, chain2 = load chain1, loc
v2, chain3 = load chain2, loc
v3 = add v2, c
Now we replace use of v1 with undef, use of chain2 with chain1.
ReplaceAllUsesWith() will iterate through uses of the first load and update operands:
v1, chain2 = load chain1, loc
v2, chain3 = load chain1, loc
v3 = add v2, c
Now the second load is the same as the first load, SelectionDAG cse will ensure
the use of second load is replaced with the first load.
v1, chain2 = load chain1, loc
v3 = add v1, c
Then v1 is replaced with undef and bad things happen.
llvm-svn: 46099
it should work, but I have no machine to test
it on. Committed because it will at least
cause no harm, and maybe someone can test it
for me!
llvm-svn: 46098
make the 'fp return in ST(0)' optimization smart enough to
look through token factor nodes. THis allows us to compile
testcases like CodeGen/X86/fp-stack-retcopy.ll into:
_carg:
subl $12, %esp
call L_foo$stub
fstpl (%esp)
fldl (%esp)
addl $12, %esp
ret
instead of:
_carg:
subl $28, %esp
call L_foo$stub
fstpl 16(%esp)
movsd 16(%esp), %xmm0
movsd %xmm0, 8(%esp)
fldl 8(%esp)
addl $28, %esp
ret
Still not optimal, but much better and this is a trivial patch. Fixing
the rest requires invasive surgery that is is not llvm 2.2 material.
llvm-svn: 46054
Likewise fix up a bunch of other libcalls. While
there I remove NEG_F32 and NEG_F64 since they are
not used anywhere. This fixes 9 Ada ACATS failures.
llvm-svn: 45833
eliminating the llvm.x86.sse2.loadl.pd intrinsic?), one shuffle optzn
may be done (if shufps is better than pinsw, Evan, please review), and
we already know about LICM of simple instructions.
llvm-svn: 45407
SelectionDAG::getConstant, in the same way as vector floating-point
constants. This allows the legalize expansion code for @llvm.ctpop and
friends to be usable with vector types.
llvm-svn: 44954
possible before resorting to pextrw and pinsrw.
- Better codegen for v4i32 shuffles masquerading as v8i16 or v16i8 shuffles.
- Improves (i16 extract_vector_element 0) codegen by recognizing
(i32 extract_vector_element 0) does not require a pextrw.
llvm-svn: 44836
_foo:
movl $12, %eax
andl 4(%esp), %eax
movl _array(%eax), %eax
ret
instead of:
_foo:
movl 4(%esp), %eax
shrl $2, %eax
andl $3, %eax
movl _array(,%eax,4), %eax
ret
As it turns out, this triggers all the time, in a wide variety of
situations, for example, I see diffs like this in various programs:
- movl 8(%eax), %eax
- shll $2, %eax
- andl $1020, %eax
- movl (%esi,%eax), %eax
+ movzbl 8(%eax), %eax
+ movl (%esi,%eax,4), %eax
- shll $2, %edx
- andl $1020, %edx
- movl (%edi,%edx), %edx
+ andl $255, %edx
+ movl (%edi,%edx,4), %edx
Unfortunately, I also see stuff like this, which can be fixed in the
X86 backend:
- andl $85, %ebx
- addl _bit_count(,%ebx,4), %ebp
+ shll $2, %ebx
+ andl $340, %ebx
+ addl _bit_count(%ebx), %ebp
llvm-svn: 44656
optimized. This avoids creating illegal divisions when the combiner is
running after legalize; this fixes PR1815. Also, it produces better
code in the included testcase by avoiding the subtract and multiply
when the division isn't optimized.
llvm-svn: 44341
sometimes emit "zero" and "all one" vectors multiple times,
for example:
_test2:
pcmpeqd %mm0, %mm0
movq %mm0, _M1
pcmpeqd %mm0, %mm0
movq %mm0, _M2
ret
instead of:
_test2:
pcmpeqd %mm0, %mm0
movq %mm0, _M1
movq %mm0, _M2
ret
This patch fixes this by always arranging for zero/one vectors
to be defined as v4i32 or v2i32 (SSE/MMX) instead of letting them be
any random type. This ensures they get trivially CSE'd on the dag.
This fix is also important for LegalizeDAGTypes, as it gets unhappy
when the x86 backend wants BUILD_VECTOR(i64 0) to be legal even when
'i64' isn't legal.
This patch makes the following changes:
1) X86TargetLowering::LowerBUILD_VECTOR now lowers 0/1 vectors into
their canonical types.
2) The now-dead patterns are removed from the SSE/MMX .td files.
3) All the patterns in the .td file that referred to immAllOnesV or
immAllZerosV in the wrong form now use *_bc to match them with a
bitcast wrapped around them.
4) X86DAGToDAGISel::SelectScalarSSELoad is generalized to handle
bitcast'd zero vectors, which simplifies the code actually.
5) getShuffleVectorZeroOrUndef is updated to generate a shuffle that
is legal, instead of generating one that is illegal and expecting
a later legalize pass to clean it up.
6) isZeroShuffle is generalized to handle bitcast of zeros.
7) several other minor tweaks.
This patch is definite goodness, but has the potential to cause random
code quality regressions. Please be on the lookout for these and let
me know if they happen.
llvm-svn: 44310
can be eliminated by the allocator is the destination and source targets the
same register. The most common case is when the source and destination registers
are in different class. For example, on x86 mov32to32_ targets GR32_ which
contains a subset of the registers in GR32.
The allocator can do 2 things:
1. Set the preferred allocation for the destination of a copy to that of its source.
2. After allocation is done, change the allocation of a copy destination (if
legal) so the copy can be eliminated.
This eliminates 443 extra moves from 403.gcc.
llvm-svn: 43662
transformation. Previously, it's restricted by ensuring the number of load uses
is one. Now the restriction is loosened up by allowing setcc uses to be
"extended" (e.g. setcc x, c, eq -> setcc sext(x), sext(c), eq).
llvm-svn: 43465
and the compaison is against a constant value, try eliminate the stride
by moving the compare instruction to another stride and change its
constant operand accordingly. e.g.
loop:
...
v1 = v1 + 3
v2 = v2 + 1
if (v2 < 10) goto loop
=>
loop:
...
v1 = v1 + 3
if (v1 < 30) goto loop
llvm-svn: 43336
- Avoid attempting stride-reuse in the case that there are users that
aren't addresses. In that case, there will be places where the
multiplications won't be folded away, so it's better to try to
strength-reduce them.
- Several SSE intrinsics have operands that strength-reduction can
treat as addresses. The previous item makes this more visible, as
any non-address use of an IV can inhibit stride-reuse.
- Make ValidStride aware of whether there's likely to be a base
register in the address computation. This prevents it from thinking
that things like stride 9 are valid on x86 when the base register is
already occupied.
Also, XFAIL the 2007-08-10-LEA16Use32.ll test; the new logic to avoid
stride-reuse elimintes the LEA in the loop, so the test is no longer
testing what it was intended to test.
llvm-svn: 43231
To do this it is necessary to add a "always inline" argument to the
memcpy node. For completeness I have also added this node to memmove
and memset. I have also added getMem* functions, because the extra
argument makes it cumbersome to use getNode and because I get confused
by it :-)
llvm-svn: 43172
enabled by passing -tailcallopt to llc. The optimization is
performed if the following conditions are satisfied:
* caller/callee are fastcc
* elf/pic is disabled OR
elf/pic enabled + callee is in module + callee has
visibility protected or hidden
llvm-svn: 42870
The only generated code difference is that now we call memcpy when
the size of the array is unknown. This matches GCC behavior and is
better since the run time value can be arbitrarily large.
llvm-svn: 42433
both results with a single div or idiv instruction. This uses new X86ISD
nodes for DIV and IDIV which are introduced during the legalize phase
so that the SelectionDAG's CSE can automatically eliminate redundant
computations.
llvm-svn: 42308
