callee-saved registers at the end of the lists. Also prefer to avoid using
the low registers that are in register subclasses required by certain
instructions, so that those registers will more likely be available when needed.
This change makes a huge improvement in spilling in some cases. Thanks to
Jakob for helping me realize the problem.
Most of this patch is fixing the testsuite. There are quite a few places
where we're checking for specific registers. I changed those to wildcards
in places where that doesn't weaken the tests. The spill-q.ll and
thumb2-spill-q.ll tests stopped spilling with this change, so I added a bunch
of live values to force spills on those tests.
llvm-svn: 116055
reapply: reimplement the second half of the or/add optimization. We should now
with no changes. Turns out that one missing "Defs = [EFLAGS]" can upset things
a bit.
llvm-svn: 116040
only end up emitting LEA instead of OR. If we aren't able to promote
something into an LEA, we should never be emitting it as an ADD.
Add some testcases that we emit "or" in cases where we used to produce
an "add".
llvm-svn: 116026
allow target to correctly compute latency for cases where static scheduling
itineraries isn't sufficient. e.g. variable_ops instructions such as
ARM::ldm.
This also allows target without scheduling itineraries to compute operand
latencies. e.g. X86 can return (approximated) latencies for high latency
instructions such as division.
- Compute operand latencies for those defined by load multiple instructions,
e.g. ldm and those used by store multiple instructions, e.g. stm.
llvm-svn: 115755
having to do a double cast (uint64_t --> double --> float). This is based on the algorithm from compiler_rt's __floatundisf
for X86-64.
llvm-svn: 115634
The x86_mmx type is used for MMX intrinsics, parameters and
return values where these use MMX registers, and is also
supported in load, store, and bitcast.
Only the above operations generate MMX instructions, and optimizations
do not operate on or produce MMX intrinsics.
MMX-sized vectors <2 x i32> etc. are lowered to XMM or split into
smaller pieces. Optimizations may occur on these forms and the
result casted back to x86_mmx, provided the result feeds into a
previous existing x86_mmx operation.
The point of all this is prevent optimizations from introducing
MMX operations, which is unsafe due to the EMMS problem.
llvm-svn: 115243
edited during emission.
If the basic block ends in a switch that gets lowered to a jump table, any
phis at the default edge were getting updated wrong. The jump table data
structure keeps a pointer to the header blocks that wasn't getting updated
after the MBB is split.
This bug was exposed on 32-bit Linux when disabling critical edge splitting in
codegen prepare.
The fix is to uipdate stale MBB pointers whenever a block is split during
emission.
llvm-svn: 115191
cost modeling for if-conversion. Now if only we had a way to estimate the misprediction probability.
Adjsut CodeGen/ARM/ifcvt10.ll. The pipeline on Cortex-A8 is long enough that it is still profitable
to predicate an ldm, but the shorter pipeline on Cortex-A9 makes it unprofitable.
llvm-svn: 114995
Rather than having arbitrary cutoffs, actually try to cost model the conversion.
For now, the constants are tuned to more or less match our existing behavior, but these will be
changed to reflect realistic values as this work proceeds.
llvm-svn: 114973
This reverts revision 114633. It was breaking llvm-gcc-i386-linux-selfhost.
It seems there is a downstream bug that is exposed by
-cgp-critical-edge-splitting=0. When that bug is fixed, this patch can go back
in.
Note that the changes to tailcallfp2.ll are not reverted. They were good are
required.
llvm-svn: 114859
x86-32: 32-bit calls were named "call" not "calll". 64-bit calls were correctly
named "callq", so this only impacted x86-32.
This fixes rdar://8456370 - llvm-mc rejects 'calll'
This also exposes that mingw/64 is generating a 32-bit call instead of a 64-bit call,
I will file a bugzilla.
llvm-svn: 114534
(sbbl x, x) sets the registers to 0 or ~0. Combined with two's complement arithmetic, we can fold
the intermediate AND and the ADD into a single SUB.
This fixes <rdar://problem/8449754>.
llvm-svn: 114460
CombinerAA cannot assume that different FrameIndex's never alias, but can instead use
MachineFrameInfo to get the actual offsets of these slots and check for actual aliasing.
This fixes CodeGen/X86/2010-02-19-TailCallRetAddrBug.ll and CodeGen/X86/tailcallstack64.ll
when CombinerAA is enabled, modulo a different register allocation sequence.
llvm-svn: 114348
between the high and low registers for prologue/epilogue code. This was
a Darwin-only thing that wasn't providing a realistic benefit anymore.
Combining the save areas simplifies the compiler code and results in better
ARM/Thumb2 codegen.
For example, previously we would generate code like:
push {r4, r5, r6, r7, lr}
add r7, sp, #12
stmdb sp!, {r8, r10, r11}
With this change, we combine the register saves and generate:
push {r4, r5, r6, r7, r8, r10, r11, lr}
add r7, sp, #12
rdar://8445635
llvm-svn: 114340
NO path to the destination containing side effects, not that SOME path contains no side effects.
In practice, this only manifests with CombinerAA enabled, because otherwise the chain has little
to no branching, so "any" is effectively equivalent to "all".
llvm-svn: 114268
value should be in GPRs when it's going to be used as a scalar, and we use
VMOVRRD to make that happen, but if the value is converted back to a vector
we need to fold to a simple bit_convert. Radar 8407927.
llvm-svn: 114233
1) Do forward copy propagation. This makes it easier to estimate the cost of the
instruction being sunk.
2) Break critical edges on demand, including cases where the value is used by
PHI nodes.
Critical edge splitting is not yet enabled by default.
llvm-svn: 114227
legacy asm printer uses instructions of the form, "mov r0, r0, lsl #3", while
the MC-instruction printer uses the form "lsl r0, r0, #3". The latter mnemonic
is correct and preferred according the ARM documentation (A8.6.98). The former
are pseudo-instructions for the latter.
llvm-svn: 114221
walking the asm arguments once and stashing their Values. This is
wrong because the same memory location can be in the list twice, and
if the first one has a sunkaddr substituted, the stashed value for the
second one will be wrong (use-after-free). PR 8154.
llvm-svn: 114104
This cleans up after the mess r108567 left in the CellSPU backend.
ORCvt-instruction were used to reinterpret registers, and the ORs were then
removed by isMoveInstr(). This patch now removes 350 instrucions of format:
or $3, $3, $3
(from the 52 testcases in CodeGen/CellSPU). One case of a nonexistant or is
checked for.
Some moves of the form 'ori $., $., 0' and 'ai $., $., 0' still remain.
llvm-svn: 114074
encountered while building llvm-gcc for arm. This is probably the same issue
that the ppc buildbot hit. llvm::prior works on a MachineBasicBlock::iterator,
not a plain MachineInstr.
llvm-svn: 113983
backing out following to get it back to green,
so I can investigate in peace:
svn merge -c -113840 llvm/test/CodeGen/ARM/arm-and-tst-peephole.ll
svn merge -c -113876 -c -113839 llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp
llvm-svn: 113980
to expose greater opportunities for store narrowing in codegen. This patch fixes a potential
infinite loop in instcombine caused by one of the introduced transforms being overly aggressive.
llvm-svn: 113763
to use AddrMode4, there was a count of the registers stored in one of the
operands. I changed that to just count the operands but forgot to adjust for
the size of D registers. This was noticed by Evan as a performance problem
but it is a potential correctness bug as well, since it is possible that this
could merge a base update with a non-matching immediate.
llvm-svn: 113576
take multiple cycles to decode.
For the current if-converter clients (actually only ARM), the instructions that
are predicated on false are not nops. They would still take machine cycles to
decode. Micro-coded instructions such as LDM / STM can potentially take multiple
cycles to decode. If-converter should take treat them as non-micro-coded
simple instructions.
llvm-svn: 113570
Since mem2reg isn't run at -O0, we get a ton of reloads from the stack,
for example, before, this code:
int foo(int x, int y, int z) {
return x+y+z;
}
used to compile into:
_foo: ## @foo
subq $12, %rsp
movl %edi, 8(%rsp)
movl %esi, 4(%rsp)
movl %edx, (%rsp)
movl 8(%rsp), %edx
movl 4(%rsp), %esi
addl %edx, %esi
movl (%rsp), %edx
addl %esi, %edx
movl %edx, %eax
addq $12, %rsp
ret
Now we produce:
_foo: ## @foo
subq $12, %rsp
movl %edi, 8(%rsp)
movl %esi, 4(%rsp)
movl %edx, (%rsp)
movl 8(%rsp), %edx
addl 4(%rsp), %edx ## Folded load
addl (%rsp), %edx ## Folded load
movl %edx, %eax
addq $12, %rsp
ret
Fewer instructions and less register use = faster compiles.
llvm-svn: 113102
"For ARM stack frames that utilize variable sized objects and have either
large local stack areas or require dynamic stack realignment, allocate a
base register via which to access the local frame. This allows efficient
access to frame indices not accessible via the FP (either due to being out
of range or due to dynamic realignment) or the SP (due to variable sized
object allocation). In particular, this greatly improves efficiency of access
to spill slots in Thumb functions which contain VLAs."
r112986 fixed a latent bug exposed by the above.
llvm-svn: 112989
vabd intrinsic and add and/or zext operations. In the case of vaba, this
also avoids the need for a DAG combine pattern to combine vabd with add.
Update tests. Auto-upgrade the old intrinsics.
llvm-svn: 112941
large local stack areas or require dynamic stack realignment, allocate a
base register via which to access the local frame. This allows efficient
access to frame indices not accessible via the FP (either due to being out
of range or due to dynamic realignment) or the SP (due to variable sized
object allocation). In particular, this greatly improves efficiency of access
to spill slots in Thumb functions which contain VLAs.
rdar://7352504
rdar://8374540
rdar://8355680
llvm-svn: 112883
there are clearly no stores between the load and the store. This fixes
this miscompile reported as PR7833.
This breaks the test/CodeGen/X86/narrow_op-2.ll optimization, which is
safe, but awkward to prove safe. Move it to X86's README.txt.
llvm-svn: 112861
add, and subtract operations with zero-extended or sign-extended vectors.
Update tests. Add auto-upgrade support for the old intrinsics.
llvm-svn: 112773
check more strict, breaking some cases not checked in the
testsuite, but also exposes some foldings not done before,
as this example:
movaps (%rdi), %xmm0
movaps (%rax), %xmm1
movaps %xmm0, %xmm2
movss %xmm1, %xmm2
shufps $36, %xmm2, %xmm0
now is generated as:
movaps (%rdi), %xmm0
movaps %xmm0, %xmm1
movlps (%rax), %xmm1
shufps $36, %xmm1, %xmm0
llvm-svn: 112753
int x(int t) {
if (t & 256)
return -26;
return 0;
}
We generate this:
tst.w r0, #256
mvn r0, #25
it eq
moveq r0, #0
while gcc generates this:
ands r0, r0, #256
it ne
mvnne r0, #25
bx lr
Scandalous really!
During ISel time, we can look for this particular pattern. One where we have a
"MOVCC" that uses the flag off of a CMPZ that itself is comparing an AND
instruction to 0. Something like this (greatly simplified):
%r0 = ISD::AND ...
ARMISD::CMPZ %r0, 0 @ sets [CPSR]
%r0 = ARMISD::MOVCC 0, -26 @ reads [CPSR]
All we have to do is convert the "ISD::AND" into an "ARM::ANDS" that sets [CPSR]
when it's zero. The zero value will all ready be in the %r0 register and we only
need to change it if the AND wasn't zero. Easy!
llvm-svn: 112664
1) nuke ConstDataCoalSection, which is dead.
2) revise my previous patch for rdar://8018335,
which was completely wrong. Specifically, it doesn't
make sense to mark __TEXT,__const_coal as PURE_INSTRUCTIONS,
because it is for readonly data. templates (it turns out)
go to const_coal_nt. The real fix for rdar://8018335 was
to give ConstTextCoalSection a section kind of ReadOnly
instead of Text.
llvm-svn: 112496
when the top elements of a vector are undefined. This happens all
the time for X86-64 ABI stuff because only the low 2 elements of
a 4 element vector are defined. For example, on:
_Complex float f32(_Complex float A, _Complex float B) {
return A+B;
}
We used to produce (with SSE2, SSE4.1+ uses insertps):
_f32: ## @f32
movdqa %xmm0, %xmm2
addss %xmm1, %xmm2
pshufd $16, %xmm2, %xmm2
pshufd $1, %xmm1, %xmm1
pshufd $1, %xmm0, %xmm0
addss %xmm1, %xmm0
pshufd $16, %xmm0, %xmm1
movdqa %xmm2, %xmm0
unpcklps %xmm1, %xmm0
ret
We now produce:
_f32: ## @f32
movdqa %xmm0, %xmm2
addss %xmm1, %xmm2
pshufd $1, %xmm1, %xmm1
pshufd $1, %xmm0, %xmm3
addss %xmm1, %xmm3
movaps %xmm2, %xmm0
unpcklps %xmm3, %xmm0
ret
This implements rdar://8368414
llvm-svn: 112378
all the other LDM/STM instructions. This fixes asm printer crashes when
compiling with -O0. I've changed one of the NEON tests (vst3.ll) to run
with -O0 to check this in the future.
Prior to this change VLDM/VSTM used addressing mode #5, but not really.
The offset field was used to hold a count of the number of registers being
loaded or stored, and the AM5 opcode field was expanded to specify the IA
or DB mode, instead of the standard ADD/SUB specifier. Much of the backend
was not aware of these special cases. The crashes occured when rewriting
a frameindex caused the AM5 offset field to be changed so that it did not
have a valid submode. I don't know exactly what changed to expose this now.
Maybe we've never done much with -O0 and NEON. Regardless, there's no longer
any reason to keep a count of the VLDM/VSTM registers, so we can use
addressing mode #4 and clean things up in a lot of places.
llvm-svn: 112322
expanding: e.g. <2 x float> -> <4 x float> instead of -> 2 floats. This
affects two places in the code: handling cross block values and handling
function return and arguments. Since vectors are already widened by
legalizetypes, this gives us much better code and unblocks x86-64 abi
and SPU abi work.
For example, this (which is a silly example of a cross-block value):
define <4 x float> @test2(<4 x float> %A) nounwind {
%B = shufflevector <4 x float> %A, <4 x float> undef, <2 x i32> <i32 0, i32 1>
%C = fadd <2 x float> %B, %B
br label %BB
BB:
%D = fadd <2 x float> %C, %C
%E = shufflevector <2 x float> %D, <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
ret <4 x float> %E
}
Now compiles into:
_test2: ## @test2
## BB#0:
addps %xmm0, %xmm0
addps %xmm0, %xmm0
ret
previously it compiled into:
_test2: ## @test2
## BB#0:
addps %xmm0, %xmm0
pshufd $1, %xmm0, %xmm1
## kill: XMM0<def> XMM0<kill> XMM0<def>
insertps $0, %xmm0, %xmm0
insertps $16, %xmm1, %xmm0
addps %xmm0, %xmm0
ret
This implements rdar://8230384
llvm-svn: 112101
comparison that would overflow.
- The other under/overflow cases can't actually happen because the immediates
which would trigger them are legal (so we don't enter this code), but
adjusted the style to make it clear the transform is always valid.
llvm-svn: 112053
comparison is in a different basic block from the branch. In such
cases, the comparison's operands may not have initialized virtual
registers available.
llvm-svn: 111709
The "half vectors" are now widened to full size by the legalizer.
The only exception is in parameter passing, where half vectors are
expanded. This causes changes to some dejagnu tests.
llvm-svn: 111360
where the step value is an induction variable from an outer loop, to
avoid trouble trying to re-expand such expressions. This effectively
hides such expressions from indvars and lsr, which prevents them
from getting into trouble.
llvm-svn: 111317
- Make foldMemoryOperandImpl aware of 256-bit zero vectors folding and support the 128-bit counterparts of AVX too.
- Make sure MOV[AU]PS instructions are only selected when SSE1 is enabled, and duplicate the patterns to match AVX.
- Add a testcase for a simple 128-bit zero vector creation.
llvm-svn: 110946
term goal here is to be able to match enough of vector_shuffle and build_vector
so all avx intrinsics which aren't mapped to their own built-ins but to
shufflevector calls can be codegen'd. This is the first (baby) step, support
building zeroed vectors.
llvm-svn: 110897
platform. It's apparently "bl __muldf3" on linux, for example. Since that's
not what we're checking here, it's more robust to just force a triple. We
just wwant to check that the inline FP instructions are only generated
on cpus that have them."
llvm-svn: 110830
float t1(int argc) {
return (argc == 1123) ? 1.234f : 2.38213f;
}
We would generate truly awful code on ARM (those with a weak stomach should look
away):
_t1:
movw r1, #1123
movs r2, #1
movs r3, #0
cmp r0, r1
mov.w r0, #0
it eq
moveq r0, r2
movs r1, #4
cmp r0, #0
it ne
movne r3, r1
adr r0, #LCPI1_0
ldr r0, [r0, r3]
bx lr
The problem was that legalization was creating a cascade of SELECT_CC nodes, for
for the comparison of "argc == 1123" which was fed into a SELECT node for the ?:
statement which was itself converted to a SELECT_CC node. This is because the
ARM back-end doesn't have custom lowering for SELECT nodes, so it used the
default "Expand".
I added a fairly simple "LowerSELECT" to the ARM back-end. It takes care of this
testcase, but can obviously be expanded to include more cases.
Now we generate this, which looks optimal to me:
_t1:
movw r1, #1123
movs r2, #0
cmp r0, r1
adr r0, #LCPI0_0
it eq
moveq r2, #4
ldr r0, [r0, r2]
bx lr
.align 2
LCPI0_0:
.long 1075344593 @ float 2.382130e+00
.long 1067316150 @ float 1.234000e+00
llvm-svn: 110799
memory and synchronization barrier dmb and dsb instructions.
- Change instruction names to something more sensible (matching name of actual
instructions).
- Added tests for memory barrier codegen.
llvm-svn: 110785
Also added a test case to check for the added benefit of this patch: it's optimizing away the unnecessary restore of sp from fp for some non-leaf functions.
llvm-svn: 110707
reserved, not available for general allocation. This eliminates all the
extra checks for Darwin.
This change also fixes the use of FP to access frame indices in leaf
functions and cleaned up some confusing code in epilogue emission.
llvm-svn: 110655
Without this what was happening was:
* R3 is not marked as "used"
* ARM backend thinks it has to save it to the stack because of vaarg
* Offset computation correctly ignores it
* Offsets are wrong
llvm-svn: 110446
nice farm in the country where it can play with other tests. And bunnies.
It is not clear what is being tested, and the revision history shows a bunch of
random changes to the expected instruction count. Clearly, we are just fudging
it to pass whenever it fails.
llvm-svn: 110118
such registers in SPU, this support boils down to "emulating"
them by duplicating instructions on the general purpose registers.
This adds the most basic operations on v2i32: passing parameters,
addition, subtraction, multiplication and a few others.
llvm-svn: 110035
away from a computer now.
--- Reverse-merging r109881 into '.':
D test/CodeGen/X86/avx-intrinsics-x86.ll
D test/CodeGen/X86/avx-intrinsics-x86_64.ll
llvm-svn: 109959
have 4 bits per register in the operand encoding), but have undefined
behavior when the operand value is 13 or 15 (SP and PC, respectively).
The trivial coalescer in linear scan sometimes will merge a copy from
SP into a subsequent instruction which uses the copy, and if that
instruction cannot legally reference SP, we get bad code such as:
mls r0,r9,r0,sp
instead of:
mov r2, sp
mls r0, r9, r0, r2
This patch adds a new register class for use by Thumb2 that excludes
the problematic registers (SP and PC) and is used instead of GPR
for those operands which cannot legally reference PC or SP. The
trivial coalescer explicitly requires that the register class
of the destination for the COPY instruction contain the source
register for the COPY to be considered for coalescing. This prevents
errant instructions like that above.
PR7499
llvm-svn: 109842
integers with mov + vdup. 8003375. This is
currently disabled by default because LICM will
not hoist a VDUP, so it pessimizes the code if
the construct occurs inside a loop (8248029).
llvm-svn: 109799
This assumption is not satisfied due to global mergeing.
Workaround the issue by temporary disablinge mergeing of const globals.
Also, ignore LLVM "special" globals. This fixes PR7716
llvm-svn: 109423
it's too late to start backing off aggressive latency scheduling when most
of the registers are in use so the threshold should be a bit tighter.
- Correctly handle live out's and extract_subreg etc.
- Enable register pressure aware scheduling by default for hybrid scheduler.
For ARM, this is almost always a win on # of instructions. It's runtime
neutral for most of the tests. But for some kernels with high register
pressure it can be a huge win. e.g. 464.h264ref reduced number of spills by
54 and sped up by 20%.
llvm-svn: 109279
- Fix a typo for PIC check during jmp table lowering
- Also fix the "first jump table basic block is not
considered only reachable by fall through" problem, use this
ad-hoc solution until I come up with something better.
Patch by stetorvs@gmail.com
llvm-svn: 108820
void foo() { __builtin_unreachable(); }
It will output the following on Darwin X86:
_func1:
Leh_func_begin0:
pushq %rbp
Ltmp0:
movq %rsp, %rbp
Ltmp1:
Leh_func_end0:
This prolog adds a new Call Frame Information (CFI) row to the FDE with an
address that is not within the address range of the code it describes -- part is
equal to the end of the function -- and therefore results in an invalid EH
frame. If we emit a nop in this situation, then the CFI row is now within the
address range.
llvm-svn: 108568
occasions, caused code to be generated in a different order.
All cases I've seen involved float softening in the type
legalizer, and this could be perhaps be fixed there, but
it's better not to generate things differently in the first
place. 7797940 (6/29/2010..7/15/2010).
llvm-svn: 108484
the function. We'll just turn it into a "trap" instruction instead.
The problem with not handling this is that it might generate a prologue without
the equivalent epilogue to go with it:
$ cat t.ll
define void @foo() {
entry:
unreachable
}
$ llc -o - t.ll -relocation-model=pic -disable-fp-elim -unwind-tables
.section __TEXT,__text,regular,pure_instructions
.globl _foo
.align 4, 0x90
_foo: ## @foo
Leh_func_begin0:
## BB#0: ## %entry
pushq %rbp
Ltmp0:
movq %rsp, %rbp
Ltmp1:
Leh_func_end0:
...
The unwind tables then have bad data in them causing all sorts of problems.
Fixes <rdar://problem/8096481>.
llvm-svn: 108473
-enable-no-nans-fp-math and -enable-no-infs-fp-math. All of the current codegen fp math optimizations only care whether the fp arithmetics arguments and results can never be NaN.
llvm-svn: 108465
to keep "Text" in sync with the "pure instructions" section attribute.
Lack of this attribute was preventing the assembler from emitting
multibyte noops instructions for templates (and inlines, and other
coalesced stuff) and was causing the assembler to mismatch .o files.
This fixes rdar://8018335
llvm-svn: 108461
address cannot be allocated a register is in 32-bit mode where the first
three arguments are marked inreg. In that case EAX, EDX, and ECX will be
used for argument passing.
This fixes PR7610.
llvm-svn: 108327
Targets must now implement TargetInstrInfo::copyPhysReg instead. There is no
longer a default implementation forwarding to copyRegToReg.
llvm-svn: 108095
correct alignment information, which simplifies ExpandRes_VAARG a bit.
The patch introduces a new alignment information to TargetLoweringInfo. This is
needed since the two natural candidates cannot be used:
* The 's' in target data: If this is set to the minimal alignment of any
argument, getCallFrameTypeAlignment would return 4 for doubles on ARM for
example.
* The getTransientStackAlignment method. It is possible for an architecture to
have argument less aligned than what we maintain the stack pointer.
llvm-svn: 108072
- Check getBytesToPopOnReturn().
- Eschew ST0 and ST1 for return values.
- Fix the PIC base register initialization so that it doesn't ever
fail to end up the top of the entry block.
llvm-svn: 108039
Bob Wilson