The initial attempt (rG89633320) botched the logic by reversing
the source/dest types. Added x86 tests for additional coverage.
The vector tests show a potential improvement (fold vector load
instead of broadcasting), but that's a known/existing problem.
This fold is done in IR by instcombine, and we have a special
form of it already here in DAGCombiner, but we want the more
general transform too:
https://rise4fun.com/Alive/3jZm
Name: general
Pre: (C1 + zext(C2) < 64)
%s = lshr i64 %x, C1
%t = trunc i64 %s to i16
%r = lshr i16 %t, C2
=>
%s2 = lshr i64 %x, C1 + zext(C2)
%a = and i64 %s2, zext((1 << (16 - C2)) - 1)
%r = trunc %a to i16
Name: special
Pre: C1 == 48
%s = lshr i64 %x, C1
%t = trunc i64 %s to i16
%r = lshr i16 %t, C2
=>
%s2 = lshr i64 %x, C1 + zext(C2)
%r = trunc %s2 to i16
...because D58017 exposes a regression without this fold.
GEP index size can be specified in the DataLayout, introduced in D42123. However, there were still places
in which getIndexSizeInBits was used interchangeably with getPointerSizeInBits. This notably caused issues
with Instcombine's visitPtrToInt; but the unit tests was incorrect, so this remained undiscovered.
This fixes the buildbot failures.
Differential Revision: https://reviews.llvm.org/D68328
Patch by Joseph Faulls!
Summary:
The use of a boolean isInteger flag (generally initialized using
VT.isInteger()) caused errors in our out-of-tree CHERI backend
(https://github.com/CTSRD-CHERI/llvm-project).
In our backend, pointers use a separate ValueType (iFATPTR) and therefore
.isInteger() returns false. This meant that getSetCCInverse() was using the
floating-point variant and generated incorrect code for us:
`(void *)0x12033091e < (void *)0xffffffffffffffff` would return false.
Committing this change will significantly reduce our merge conflicts
for each upstream merge.
Reviewers: spatel, bogner
Reviewed By: bogner
Subscribers: wuzish, arsenm, sdardis, nemanjai, jvesely, nhaehnle, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70917
Updated pred_load patterns added to AArch64SVEInstrInfo.td by this patch
to use reg + imm non-temporal loads to fix previous test failures.
Original commit message:
Adds the following intrinsics:
- llvm.aarch64.sve.ldnt1
- llvm.aarch64.sve.stnt1
This patch creates masked loads and stores with the
MONonTemporal flag set when used with the intrinsics above.
Summary:
This is a quickfix for PR44275. An assertion that checks that the
DIExpression is valid failed due to attempting to create an entry value
for an indirect parameter. This started appearing after D69028, as the
indirect parameter started being represented using an DW_OP_deref,
rather than with the DBG_VALUE's second operand, meaning that the
isIndirectDebugValue() check in LiveDebugValues did not exclude such
parameters. A DIExpression that has an entry value operation can
currently not have any other operation, leading to the failed isValid()
check.
This patch simply makes us stop considering emitting entry values
for such parameters. To support such cases I think we at least need
to do the following changes:
* In DIExpression::isValid(): Remove the limitation that a
DW_OP_LLVM_entry_value operation can be the only operation in a
DIExpression.
* In LiveDebugValues::emitEntryValues(): Create an entry value of size
1, so that it only wraps the register operand, and not the whole
pre-existing expression (the DW_OP_deref).
* In LiveDebugValues::removeEntryValue(): Check that the new debug
value has the same debug expression as the original, rather than
checking that the debug expression is empty.
* In DwarfExpression::addMachineRegExpression(): Modify the logic so
that a DW_OP_reg* expression is emitted for the entry value.
That is how GCC emits entry values for indirect parameters. That will
currently not happen to due the DW_OP_deref causing the
!HasComplexExpression to fail. The LocationKind needs to be changed
also, rather than always emitting a DW_OP_stack_value for entry values.
There are probably more things I have missed, but that could hopefully
be a good starting point for emitting such entry values.
Reviewers: djtodoro, aprantl, jmorse, vsk
Reviewed By: aprantl, vsk
Subscribers: hiraditya, llvm-commits
Tags: #debug-info, #llvm
Differential Revision: https://reviews.llvm.org/D71416
I believe this is a leftover from when fp128 was softened to fp128
on X86-64. In that case type legalization must have been able to
create a load that was the same as N which would make this
replacement fail or assert. Since we no longer do that, this
check should be unneeded.
(except for v4 loclists, which are sufficiently different to not fit
well in this generic implementation)
In subsequent patches I intend to refactor the DebugLoc and ranges data
structures to be more similar so I can common more of the implementation
here.
Summary:
Support alloca-referencing dbg.value in hwasan instrumentation.
Update AsmPrinter to emit DW_AT_LLVM_tag_offset when location is in
loclist format.
Reviewers: pcc
Subscribers: srhines, aprantl, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70753
This fold is done in IR by instcombine, and we have a special
form of it already here in DAGCombiner, but we want the more
general transform too:
https://rise4fun.com/Alive/3jZm
Name: general
Pre: (C1 + zext(C2) < 64)
%s = lshr i64 %x, C1
%t = trunc i64 %s to i16
%r = lshr i16 %t, C2
=>
%s2 = lshr i64 %x, C1 + zext(C2)
%a = and i64 %s2, zext((1 << (16 - C2)) - 1)
%r = trunc %a to i16
Name: special
Pre: C1 == 48
%s = lshr i64 %x, C1
%t = trunc i64 %s to i16
%r = lshr i16 %t, C2
=>
%s2 = lshr i64 %x, C1 + zext(C2)
%r = trunc %s2 to i16
...because D58017 exposes a regression without this fold.
This is not quite NFC because I changed the SDLoc to use the more
standard 'N' (the starting node for the fold).
This transform is a special-case of a more general fold that we
do in IR, but it seems like the general fold is needed here too
to avoid a potential regression seen in D58017.
https://rise4fun.com/Alive/3jZm
During SelectionDAG, if a value which is associated with a DBG_VALUE
needs to be split across multiple registers, the DBG_VALUE will be split
into a set of fragment expressions to recreate the original value.
If one or more of these fragments cannot be created, they would
previously be silently dropped, causing the old debug value to live past
its expiry date. This patch fixes this issue by keeping invalid
fragments while setting their value as Undef.
Differential revision: https://reviews.llvm.org/D70248
GEP index size can be specified in the DataLayout, introduced in D42123. However, there were still places
in which getIndexSizeInBits was used interchangeably with getPointerSizeInBits. This notably caused issues
with Instcombine's visitPtrToInt; but the unit tests was incorrect, so this remained undiscovered.
Differential Revision: https://reviews.llvm.org/D68328
Patch by Joseph Faulls!
No more hash collisions for memoperands. Now the MIRCanonicalization
pass shouldn't hit hash collisions when dealing with nearly identical
memory accessing instructions when their memoperands are in fact different.
Differential Revision: https://reviews.llvm.org/D71328
This has two main effects:
- Optimizes debug info size by saving 221.86 MB of obj file size in a
Windows optimized+debug build of 'all'. This is 3.03% of 7,332.7MB of
object file size.
- Incremental step towards decoupling target intrinsics.
The enums are still compact, so adding and removing a single
target-specific intrinsic will trigger a rebuild of all of LLVM.
Assigning distinct target id spaces is potential future work.
Part of PR34259
Reviewers: efriedma, echristo, MaskRay
Reviewed By: echristo, MaskRay
Differential Revision: https://reviews.llvm.org/D71320
This reverts commit 30038da15b. It causes
the stage2 thinLTO bot to fail with:
Assertion failed: (CU.getDIE(CalleeSP) && "Expected declaration subprogram DIE for callee")
rdar://57840415
The fp16 to larger than fp32 inserts an extend that need to
re-legalized if fp16 is promoted. But if we check for fp16
promotion first, then we can avoid emiting the fp_extend all
together.
This is an alternate fix for the bug discussed in D70595.
This also includes minimal tests for other in-tree targets
to show the problem more generally.
We check the number of uses as a predicate for whether some
value is free to negate, but that use count can change as we
rewrite the expression in getNegatedExpression(). So something
that was marked free to negate during the cost evaluation
phase becomes not free to negate during the rewrite phase (or
the inverse - something that was not free becomes free).
This can lead to a crash/assert because we expect that
everything in an expression that is negatible to be handled
in the corresponding code within getNegatedExpression().
This patch skips the use check during the rewrite phase.
So we determine that some expression isNegatibleForFree
(identically to without this patch), but during the rewrite,
don't rely on use counts to decide how to create the optimal
expression.
Differential Revision: https://reviews.llvm.org/D70975
Summary:
Adds the following intrinsics:
- llvm.aarch64.sve.ldnt1
- llvm.aarch64.sve.stnt1
This patch creates masked loads and stores with the
MONonTemporal flag set when used with the intrinsics above.
Reviewers: sdesmalen, paulwalker-arm, dancgr, mgudim, efriedma, rengolin
Reviewed By: efriedma
Subscribers: tschuett, kristof.beyls, hiraditya, rkruppe, psnobl, cfe-commits, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71000
After creating a low-overhead loop, the loop update instruction was still
lingering around hurting performance. This removes dead loop update
instructions, which in our case are mostly SUBS instructions.
To support this, some helper functions were added to MachineLoopUtils and
ReachingDefAnalysis to analyse live-ins of loop exit blocks and find uses
before a particular loop instruction, respectively.
This is a first version that removes a SUBS instruction when there are no other
uses inside and outside the loop block, but there are some more interesting
cases in test/CodeGen/Thumb2/LowOverheadLoops/mve-tail-data-types.ll which
shows that there is room for improvement. For example, we can't handle this
case yet:
..
dlstp.32 lr, r2
.LBB0_1:
mov r3, r2
subs r2, #4
vldrh.u32 q2, [r1], #8
vmov q1, q0
vmla.u32 q0, q2, r0
letp lr, .LBB0_1
@ %bb.2:
vctp.32 r3
..
which is a lot more tricky because r2 is not only used by the subs, but also by
the mov to r3, which is used outside the low-overhead loop by the vctp
instruction, and that requires a bit of a different approach, and I will follow
up on this.
Differential Revision: https://reviews.llvm.org/D71007
Summary:
In the function `EarlyIfPredicator::shouldConvertIf()`, we call
`TII->isProfitableToIfCvt()` with `BranchProbability::getUnknown()`, it may
cause the potential assertion error for those hook which use `BranchProbability`
in `isProfitableToIfCvt()`, for example `SystemZ`.
`SystemZ` use `Probability < BranchProbability(1, 8))` in the function
`SystemZInstrInfo::isProfitableToIfCvt()`, if we call this function with
`BranchProbability::getUnknown()`, it will cause assertion error.
This patch is to fix the potential bug.
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D71273
This iterator range just includes physical registers and register masks,
which are interesting when dealing with register liveness.
Reviewers: evandro, t.p.northover, paquette, MatzeB, arsenm
Reviewed By: paquette
Differential Revision: https://reviews.llvm.org/D70562
I think this is no longer needed. The system should take care
of legalizing any new nodes that are added. I think this might
have been needed prior to r371709 or r307053.
Now, flags will result in differing hashes for a given MI. In effect, if
you have two instructions with everything identical except for their
flags then you should get two different hashes and fewer collisions.
Differential Revision: https://reviews.llvm.org/D70479
Summary: This is a follow up of D69281, it enables the X86 backend support for the FP comparision.
Reviewers: uweigand, kpn, craig.topper, RKSimon, cameron.mcinally, andrew.w.kaylor
Subscribers: hiraditya, llvm-commits, annita.zhang, LuoYuanke, LiuChen3
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70582
This allows a call site tag in CU A to reference a callee DIE in CU B
without resorting to creating an incomplete duplicate DIE for the callee
inside of CU A.
We already allow cross-CU references of subprogram declarations, so it
doesn't seem like definitions ought to be special.
This improves entry value evaluation and tail call frame synthesis in
the LTO setting. During LTO, it's common for cross-module inlining to
produce a call in some CU A where the callee resides in a different CU,
and there is no declaration subprogram for the callee anywhere. In this
case llvm would (unnecessarily, I think) emit an empty DW_TAG_subprogram
in order to fill in the call site tag. That empty 'definition' defeats
entry value evaluation etc., because the debugger can't figure out what
it means.
As a follow-up, maybe we could add a DWARF verifier check that a
DW_TAG_subprogram at least has a DW_AT_name attribute.
rdar://46577651
Differential Revision: https://reviews.llvm.org/D70350
This caused non-determinism in the compiler, see command on the Phabricator
code review.
> This patch addresses a performance problem reported in PR43855, and
> present in the reapplication in in 001574938e5. It turns out that
> MachineSink will (often) move instructions to the first block that
> post-dominates the current block, and then try to sink further. This
> means if we have a lot of conditionals, we can needlessly create large
> numbers of DBG_VALUEs, one in each block the sunk instruction passes
> through.
>
> To fix this, rather than immediately sinking DBG_VALUEs, record them in
> a pass structure. When sinking is complete and instructions won't be
> sunk any further, new DBG_VALUEs are added, avoiding lots of
> intermediate DBG_VALUE $noregs being created.
>
> Differential revision: https://reviews.llvm.org/D70676
Summary:
This patch fixes a few issues when large arrays are allocated on the
stack. Currently, clang has inconsistent behaviour, for debug builds
there is an assertion failure when the array size on stack is around 2GB
but there is no assertion when the stack is around 8GB. For release
builds there is no assertion, the compilation succeeds but generates
incorrect code. The incorrect code generated is due to using
int/unsigned int instead of their 64-bit counterparts. This patch,
1) Removes the assertion in frame legality check.
2) Converts int/unsigned int in some places to the 64-bit variants. This
helps in generating correct code and removes the inconsistent behaviour.
3) Adds a test which runs without optimisations.
Reviewers: sdesmalen, efriedma, fhahn, aemerson
Reviewed By: efriedma
Subscribers: eli.friedman, fpetrogalli, kristof.beyls, hiraditya,
llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70496
Summary:
This fixes PR44135.
The special case when we promote a bitcast from a vector to an int
needs special handling when we are on a big-endian target.
Prior to this fix, for the added vec_to_int we see the following in the
SelectionDAG printouts
Type-legalized selection DAG: %bb.1 'foo:bb.1'
SelectionDAG has 9 nodes:
t0: ch = EntryToken
t2: v8i16,ch = CopyFromReg t0, Register:v8i16 %0
t17: v4i32 = bitcast t2
t23: i32 = extract_vector_elt t17, Constant:i32<3>
t8: ch,glue = CopyToReg t0, Register:i32 $r0, t23
t9: ch = ARMISD::RET_FLAG t8, Register:i32 $r0, t8:1
and I think here the extract_vector_elt is wrong and extracts the value
from the wrong index.
The program program should return the 32 bits made up of the elements at
index 4 and 5 in the vec6 array, but with
t23: i32 = extract_vector_elt t17, Constant:i32<3>
as far as I can tell, we will extract values that originally didn't even
exist in the vec6 vectore.
If we would instead extract the element at index 2 we would get the wanted
values.
With this fix we insert a right shift after the bitcast in
DAGTypeLegalizer::PromoteIntRes_BITCAST which then gives us
Type-legalized selection DAG: %bb.1 'vec_to_int:bb.1'
SelectionDAG has 9 nodes:
t0: ch = EntryToken
t2: v8i16,ch = CopyFromReg t0, Register:v8i16 %0
t23: v4i32 = bitcast t2
t27: i32 = extract_vector_elt t23, Constant:i32<2>
t8: ch,glue = CopyToReg t0, Register:i32 $r0, t27
t9: ch = ARMISD::RET_FLAG t8, Register:i32 $r0, t8:1
So now we get
t27: i32 = extract_vector_elt t23, Constant:i32<2>
which is what we want.
Similarly, the new int_to_vec testcase exposes a bug where we cast the other
direction. Then we instead need to add a left shift before the bitcast on
big-endian targets for the bits in the input integer to end up at the exptected
place in the vector.
Reviewers: bogner, spatel, craig.topper, t.p.northover, dmgreen, efriedma, SjoerdMeijer, samparker
Reviewed By: efriedma
Subscribers: eli.friedman, bjope, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70942
Making some changes to MIRVRegNamerUtils.cpp to use some more modern c++
features as well as some changes to generally make the code more concise
and more understandable.
I make this an NFCi because in one case I drop the whole
"if (!MO->isDef()) MO->setIsKill(false);" thing that was added in the
original implementation, generally because I don't think this is really
semantically sound. I also changed up the implementation of
VRegRenamer::createVirtualRegisterWithLowerName somewhat because I am
now lower-casing the name unconditionally because I confirmed that that
was in fact aditya_nandakumar@apple.com's intent.
In all other cases, behavior should not be changed.
Differential Revision: https://reviews.llvm.org/D71182
D34393 added MCCodePadder as an infrastructure for padding code with
NOP instructions. It lacked tests and was not being worked on since
then.
Intel has now worked on an assembler patch to mitigate performance loss
after applying microcode update for the Jump Conditional Code Erratum.
https://www.intel.com/content/www/us/en/support/articles/000055650/processors.html
This new patch shares similarity with MCCodePadder, but has a concrete
use case in mind and is being actively developed. The infrastructure it
introduces can potentially be used for general performance improvement
via alignment. Delete the unused MCCodePadder so that people can develop
the new feature from a clean state.
Reviewed By: jyknight, skan
Differential Revision: https://reviews.llvm.org/D71106
As discussed in https://reviews.llvm.org/D69998, we miss to create some dependency edges
if chained more than 2 instructions. Adding an assertion here if someone want to chain
more than 2 instructions.
Differential Revision: https://reviews.llvm.org/D71180
Summary:
Split off of D67120.
Add the profile guided size optimization instrumentation / queries in the code
gen or target passes. This doesn't enable the size optimizations in those passes
yet as they are currently disabled in shouldOptimizeForSize (for non-IR pass
queries).
A second try after reverted D71072.
Reviewers: davidxl
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71149
The cycle values in modulo scheduling results can be negative.
The result of ModuloSchedule::getCycle() must be received as an int type.
Patch by Masaki Arai!
Differential Revision: https://reviews.llvm.org/D71122
CodeGenPrepare::placeDebugValues moves variable location intrinsics to be
immediately after the Value they refer to. This makes tracking of locations
very easy; but it changes the order in which assignments appear to the
debugger, from the source programs order to the order in which the
optimised program computes values. This then leads to PR43986 and PR38754,
where variable locations that were in a conditional block are made
unconditional, which is highly misleading.
This patch adjusts placeDbgValues to only re-order variable location
intrinsics if they use a Value before it is defined, significantly reducing
the damage that it does. This is still not 100% safe, but the rest of
CodeGenPrepare needs polishing to correctly update debug info when
optimisations are performed to fully fix this.
This will probably break downstream debuginfo tests -- if the
instruction-stream position of variable location changes isn't the focus of
the test, an easy fix should be to manually apply placeDbgValues' behaviour
to the failing tests, moving dbg.value intrinsics next to SSA variable
definitions thus:
%foo = inst1
%bar = ...
%baz = ...
void call @llvm.dbg.value(metadata i32 %foo, ...
to
%foo = inst1
void call @llvm.dbg.value(metadata i32 %foo, ...
%bar = ...
%baz = ...
This should return your test to exercising whatever it was testing before.
Differential Revision: https://reviews.llvm.org/D58453
Summary:
Currently the describeLoadedValue() hook is assumed to describe the
value of the instruction's first explicit define. The hook will not be
called for instructions with more than one explicit define.
This commit adds a register parameter to the describeLoadedValue() hook,
and invokes the hook for all registers in the worklist.
This will allow us to for example describe instructions which produce
more than two parameters' values; e.g. Hexagon's various combine
instructions.
This also fixes situations in our downstream target where we may pass
smaller parameters in the high part of a register. If such a parameter's
value is produced by a larger copy instruction, we can't describe the
call site value using the super-register, and we instead need to know
which sub-register that should be used.
This also allows us to handle cases like this:
$ebx = [...]
$rdi = MOVSX64rr32 $ebx
$esi = MOV32rr $edi
CALL64pcrel32 @call
The hook will first be invoked for the MOV32rr instruction, which will
say that @call's second parameter (passed in $esi) is described by $edi.
As $edi is not preserved it will be added to the worklist. When we get
to the MOVSX64rr32 instruction, we need to describe two values; the
sign-extended value of $ebx -> $rdi for the first parameter, and $ebx ->
$edi for the second parameter, which is now possible.
This commit modifies the dbgcall-site-lea-interpretation.mir test case.
In the test case, the values of some 32-bit parameters were produced
with LEA64r. Perhaps we can in general cases handle such by emitting
expressions that AND out the lower 32-bits, but I have not been able to
land in a case where a LEA64r is used for a 32-bit parameter instead of
LEA64_32 from C code.
I have not found a case where it would be useful to describe parameters
using implicit defines, so in this patch the hook is still only invoked
for explicit defines of forwarding registers.
Reviewers: djtodoro, NikolaPrica, aprantl, vsk
Reviewed By: djtodoro, vsk
Subscribers: ormris, hiraditya, llvm-commits
Tags: #debug-info, #llvm
Differential Revision: https://reviews.llvm.org/D70431
Currently the describeLoadedValue() hook is assumed to describe the
value of the instruction's first explicit define. The hook will not be
called for instructions with more than one explicit define.
This commit adds a register parameter to the describeLoadedValue() hook,
and invokes the hook for all registers in the worklist.
This will allow us to for example describe instructions which produce
more than two parameters' values; e.g. Hexagon's various combine
instructions.
This also fixes a case in our downstream target where we may pass
smaller parameters in the high part of a register. If such a parameter's
value is produced by a larger copy instruction, we can't describe the
call site value using the super-register, and we instead need to know
which sub-register that should be used.
This also allows us to handle cases like this:
$ebx = [...]
$rdi = MOVSX64rr32 $ebx
$esi = MOV32rr $edi
CALL64pcrel32 @call
The hook will first be invoked for the MOV32rr instruction, which will
say that @call's second parameter (passed in $esi) is described by $edi.
As $edi is not preserved it will be added to the worklist. When we get
to the MOVSX64rr32 instruction, we need to describe two values; the
sign-extended value of $ebx -> $rdi for the first parameter, and $ebx ->
$edi for the second parameter, which is now possible.
This commit modifies the dbgcall-site-lea-interpretation.mir test case.
In the test case, the values of some 32-bit parameters were produced
with LEA64r. Perhaps we can in general cases handle such by emitting
expressions that AND out the lower 32-bits, but I have not been able to
land in a case where a LEA64r is used for a 32-bit parameter instead of
LEA64_32 from C code.
I have not found a case where it would be useful to describe parameters
using implicit defines, so in this patch the hook is still only invoked
for explicit defines of forwarding registers.
This caused "Too many bits for uint64_t" asserts when building Chromium. See
https://crbug.com/1031978#c2 for a reproducer. I'll follow up on the
llvm-commits thread with a creduced version.
> ARMCodeGenPrepare has already been generalized and renamed to
> TypePromotion. We've had it enabled and tested downstream for a
> while, so enable it by default.
>
> Differential Revision: https://reviews.llvm.org/D70998
This adds support for constrained floating-point comparison intrinsics.
Specifically, we add:
declare <ty2>
@llvm.experimental.constrained.fcmp(<type> <op1>, <type> <op2>,
metadata <condition code>,
metadata <exception behavior>)
declare <ty2>
@llvm.experimental.constrained.fcmps(<type> <op1>, <type> <op2>,
metadata <condition code>,
metadata <exception behavior>)
The first variant implements an IEEE "quiet" comparison (i.e. we only
get an invalid FP exception if either argument is a SNaN), while the
second variant implements an IEEE "signaling" comparison (i.e. we get
an invalid FP exception if either argument is any NaN).
The condition code is implemented as a metadata string. The same set
of predicates as for the fcmp instruction is supported (except for the
"true" and "false" predicates).
These new intrinsics are mapped by SelectionDAG codegen onto two new
ISD opcodes, ISD::STRICT_FSETCC and ISD::STRICT_FSETCCS, again
representing quiet vs. signaling comparison operations. Otherwise
those nodes look like SETCC nodes, with an additional chain argument
and result as usual for strict FP nodes. The patch includes support
for the common legalization operations for those nodes.
The patch also includes full SystemZ back-end support for the new
ISD nodes, mapping them to all available SystemZ instruction to
fully implement strict semantics (scalar and vector).
Differential Revision: https://reviews.llvm.org/D69281
D53794 introduced code to perform the FP_TO_UINT expansion via FP_TO_SINT in a way that would never expose floating-point exceptions in the intermediate steps. Unfortunately, I just noticed there is still a way this can happen. As discussed in D53794, the compiler now generates this sequence:
// Sel = Src < 0x8000000000000000
// Val = select Sel, Src, Src - 0x8000000000000000
// Ofs = select Sel, 0, 0x8000000000000000
// Result = fp_to_sint(Val) ^ Ofs
The problem is with the Src - 0x8000000000000000 expression. As I mentioned in the original review, that expression can never overflow or underflow if the original value is in range for FP_TO_UINT. But I missed that we can get an Inexact exception in the case where Src is a very small positive value. (In this case the result of the sub is ignored, but that doesn't help.)
Instead, I'd suggest to use the following sequence:
// Sel = Src < 0x8000000000000000
// FltOfs = select Sel, 0, 0x8000000000000000
// IntOfs = select Sel, 0, 0x8000000000000000
// Result = fp_to_sint(Val - FltOfs) ^ IntOfs
In the case where the value is already in range of FP_TO_SINT, we now simply compute Val - 0, which now definitely cannot trap (unless Val is a NaN in which case we'd want to trap anyway).
In the case where the value is not in range of FP_TO_SINT, but still in range of FP_TO_UINT, the sub can never be inexact, as Val is between 2^(n-1) and (2^n)-1, i.e. always has the 2^(n-1) bit set, and the sub is always simply clearing that bit.
There is a slight complication in the case where Val is a constant, so we know at compile time whether Sel is true or false. In that scenario, the old code would automatically optimize the sub away, while this no longer happens with the new code. Instead, I've added extra code to check for this case and then just fall back to FP_TO_SINT directly. (This seems to catch even slightly more cases.)
Original version of the patch by Ulrich Weigand. X86 changes added by Craig Topper
Differential Revision: https://reviews.llvm.org/D67105
Summary:
Split off of D67120.
Add the profile guided size optimization instrumentation / queries in the code
gen or target passes. This doesn't enable the size optimizations in those passes
yet as they are currently disabled in shouldOptimizeForSize (for non-IR pass
queries).
Reviewers: davidxl
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71072
Current tail duplication integrated in bb layout is designed to increase the fallthrough from a BB's predecessor to its successor, but we have observed cases that duplication doesn't increase fallthrough, or it brings too much size overhead.
To overcome these two issues in function canTailDuplicateUnplacedPreds I add two checks:
make sure there is at least one duplication in current work set.
the number of duplication should not exceed the number of successors.
The modification in hasBetterLayoutPredecessor fixes a bug that potential predecessor must be at the bottom of a chain.
Differential Revision: https://reviews.llvm.org/D64376
One of CodeGenPrepare's optimizations is to duplicate address calculations
into basic blocks, so that as much information as possible can be folded
into memory addressing operands. This is great -- but the dbg.value
variable location intrinsics are not updated in the same way. This can lead
to dbg.values referring to address computations in other blocks that will
never be encoded into the DAG, while duplicate address computations are
performed locally that could be used by the dbg.value. Some of these (such
as non-constant-offset GEPs) can't be salvaged past.
Fix this by, whenever we duplicate an address computation into a block,
looking for dbg.value users of the original memory address in the same
block, and redirecting those to the local computation.
Differential Revision: https://reviews.llvm.org/D58403
This patch implements the following changes:
1) SelectionDAGBuilder::visitConstrainedFPIntrinsic currently treats
each constrained intrinsic like a global barrier (e.g. a function call)
and fully serializes all pending chains. This is actually not required;
it is allowed for constrained intrinsics to be reordered w.r.t one
another or (nonvolatile) memory accesses. The MI-level scheduler already
allows for that flexibility, so it makes sense to allow it at the DAG
level as well.
This patch therefore changes the way chains for constrained intrisincs
are created, and handles them basically like load operations are handled.
This has the effect that constrained intrinsics are no longer serialized
against one another or (nonvolatile) loads. They are still serialized
against stores, but that seems hard to change with the current DAG chain
setup, and it also doesn't seem to be a big problem preventing DAG
2) The OPC_CheckFoldableChainNode check requires that each of the
intermediate nodes in a multi-node pattern match only has a single use.
This check tends to fail if those intermediate nodes are strict operations
as those have a chain output that typically indeed has another use.
However, we don't really need to consider chains here at all, since they
will all be rewritten anyway by UpdateChains later. Other parts of the
matcher therefore already ignore chains, but this hasOneUse check doesn't.
This patch replaces hasOneUse by a custom test that verifies there is no
more than one use of any non-chain output value.
In theory, this change could affect code unrelated to strict FP nodes,
but at least on SystemZ I could not find any single instance of that
happening
3) The SystemZ back-end currently does not allow matching multiply-and-
extend operations (32x32 -> 64bit or 64x64 -> 128bit FP multiply) for
strict FP operations. This was not possible in the past due to the
problems described under 1) and 2) above.
With those issues fixed, it is now possible to fully support those
instructions in strict mode as well, and this patch does so.
Differential Revision: https://reviews.llvm.org/D70913
That refactoring moves NonRelocatableStringpool into common CodeGen folder.
So that NonRelocatableStringpool could be used not only inside dsymutil.
Differential Revision: https://reviews.llvm.org/D71068
This is for the case where -gmlt -gsplit-dwarf -fsplit-dwarf-inlining
are used together in some but not all units during LTO (or, in the
reduced case, even without LTO) - ensuring that no split dwarf is used
(because split-dwarf-inlining puts the same data in the .o file, so
there's no need to duplicate it into the .dwo file)
* Context *
During register coalescing, we use rematerialization when coalescing is not
possible. That means we may rematerialize a super register when only a smaller
register is actually used.
E.g.,
0B v1 = ldimm 0xFF
1B v2 = COPY v1.low8bits
2B = v2
=>
0B v1 = ldimm 0xFF
1B v2 = ldimm 0xFF
2B = v2.low8bits
Where xB are the slot indexes.
Here v2 grew from a 8-bit register to a 16-bit register.
When that happens and subregister liveness is enabled, we create subranges for
the newly created value.
E.g., before remat, the live range of v2 looked like:
main range: [1r, 2r)
(Reads v2 is defined at index 1 slot register and used before the slot register
of index 2)
After remat, it should look like:
main range: [1r, 2r)
low 8 bits: [1r, 2r)
high 8 bits: [1r, 1d) <-- dead def
I.e., the unsused lanes of v2 should be marked as dead definition.
* The Problem *
Prior to this patch, the live-ranges from the previous exampel, would have the
full live-range for all subranges:
main range: [1r, 2r)
low 8 bits: [1r, 2r)
high 8 bits: [1r, 2r) <-- too long
* The Fix *
Technically, the code that this patch changes is not wrong:
When we create the subranges for the newly rematerialized value, we create only
one subrange for the whole bit mask.
In other words, at this point v2 live-range looks like this:
main range: [1r, 2r)
low & high: [1r, 2r)
Then, it gets wrong when we call LiveInterval::refineSubRanges on low 8 bits:
main range: [1r, 2r)
low 8 bits: [1r, 2r)
high 8 bits: [1r, 2r) <-- too long
Ideally, we would like LiveInterval::refineSubRanges to be able to do the right
thing and mark the dead lanes as such. However, this is not possible, because by
the time we update / refine the live ranges, the IR hasn't been updated yet,
therefore we actually don't have enough information to do the right thing.
Another option to fix the problem would have been to call
LiveIntervals::shrinkToUses after the IR is updated. This is not desirable as
this may have a noticeable impact on compile time.
Instead, what this patch does is when we create the subranges for the
rematerialized value, we explicitly create one subrange for the lanes that were
used before rematerialization and one for the lanes that were not used. The used
one inherits the live range of the main range and the unused one is just created
empty. The existing rematerialization code then detects that the unused one are
not live and it correctly sets dead def intervals for them.
https://llvm.org/PR41372
The loclists_table_base was being overwritten for each CU even though
only one loclists contribution is made so everything but the last CU
would have a label that was never defined and fail to assemble.
Summary:
Previously, it was not possible to skip running the localizer pass
conditionally. This patch adds an input function to the pass which
decides if the pass should run on the given MachineFunction or not.
No test case as there is no upstream target needs this functionality.
Reviewers: qcolombet
Reviewed By: qcolombet
Subscribers: rovka, hiraditya, Petar.Avramovic, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71038
This patch addresses a performance problem reported in PR43855, and
present in the reapplication in in 001574938e5. It turns out that
MachineSink will (often) move instructions to the first block that
post-dominates the current block, and then try to sink further. This
means if we have a lot of conditionals, we can needlessly create large
numbers of DBG_VALUEs, one in each block the sunk instruction passes
through.
To fix this, rather than immediately sinking DBG_VALUEs, record them in
a pass structure. When sinking is complete and instructions won't be
sunk any further, new DBG_VALUEs are added, avoiding lots of
intermediate DBG_VALUE $noregs being created.
Differential revision: https://reviews.llvm.org/D70676
Fix part of PR43855, resolving a problem that comes from the reapplication
in 001574938e5. If we have two DBG_VALUE insts in a block that specify
the location of the same variable, for example:
%0 = someinst
DBG_VALUE %0, !123, !DIExpression()
%1 = anotherinst
DBG_VALUE %1, !123, !DIExpression()
if %0 were to sink, the corresponding DBG_VALUE would sink too, past the
next DBG_VALUE, effectively re-ordering assignments. To fix this, I've
added a SeenDbgVars set recording what variable locations have been seen in
a block already (working bottom up), and now flag DBG_VALUEs that would
pass a later DBG_VALUE for the same variable.
NB, this only works for repeated DBG_VALUEs in the same basic block, the
general case involving control flow is much harder, which I've written
up in PR44117.
Differential revision: https://reviews.llvm.org/D70672
These were:
* D58386 / f5e1b718a6 / reverted in d382a8a768
* D58238 / ee50590e16 / reverted in a8db456b53
Of which the latter has a performance regression tracked in PR43855,
fixed by D70672 / D70676, which will be committed atomically with this
reapplication.
Contains a minor difference to account for a change in the IsCopyInstr
signature.
ARMCodeGenPrepare has already been generalized and renamed to
TypePromotion. We've had it enabled and tested downstream for a
while, so enable it by default.
Differential Revision: https://reviews.llvm.org/D70998
Summary:
If a call is bundled then the code that looks for instructions that
produce parameter values would break when reaching the call's bundle
header, due to the `ifCall(/*AnyInBundle*/)` invocation returning true.
It is not enough to simply ignore bundle headers in the `isCall()`
invocation, as the bundle header may have defines of parameter registers
due to the call, meaning that such registers would incorrectly be
removed from the worklist. Therefore, do not look at bundle headers at
all.
Reviewers: djtodoro, NikolaPrica, aprantl, vsk
Reviewed By: aprantl, vsk
Subscribers: hiraditya, llvm-commits
Tags: #debug-info, #llvm
Differential Revision: https://reviews.llvm.org/D71024
This patch adds forward iterators mc_difflist_iterator,
mc_subreg_iterator and mc_superreg_iterator, based on the existing
DiffListIterator. Those are used to provide iterator ranges over
sub- and super-register from TRI, which are slightly more convenient
than the existing MCSubRegIterator/MCSuperRegIterator. Unfortunately,
it duplicates a bit of functionality, but the new iterators are a bit
more convenient (and can be used with various existing iterator
utilities) and should probably replace the old iterators in the future.
This patch updates some existing users.
Reviewers: evandro, qcolombet, paquette, MatzeB, arsenm
Reviewed By: qcolombet
Differential Revision: https://reviews.llvm.org/D70565
This patch turns MachineOperandIteratorBase into a regular forward
iterator, which can be used with iterator_range.
It also adds mi_bundle_ops and const_mi_bundle_ops that return iterator
ranges over all operands in a bundle and updates a use of the old
iterator.
Reviewers: evandro, t.p.northover, paquette, MatzeB, arsenm
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D70561
Fix assertion error
```
bool llvm::MachineOperand::isRenamable() const: Assertion `Register::isPhysicalRegister(getReg()) && "isRenamable should only be checked on physical registers"' failed.
```
by checking if the register is 0 before invoking `isRenamable`.
Summary:
This patch mainly do such transformation
```
$R0 = OP ...
... // No read/clobber of $R0 and $R1
$R1 = COPY $R0 // $R0 is killed
```
Replace $R0 with $R1 and remove the COPY, we have
```
$R1 = OP ...
```
This transformation can also expose more opportunities for existing
copy elimination in MCP.
Differential Revision: https://reviews.llvm.org/D67794
An interplay of code from D70210, along with code from the
Value-Numbering-esque hash-based namer from D70210, as well as some
crusty code from the original MIR-Canon code lead to multiple causes of
failure when canonicalizing or renaming vregs for MIR with multiple
basic blocks. This patch fixes those issues while deleting some no
longer needed code and adding a nice diamond test case to boot.
Differential Revision: https://reviews.llvm.org/D70478
That patch fixes incompatible compilation unit type (DW_UT_skeleton) and root DIE (DW_TAG_compile_unit) error.
cat split-dwarf.cpp
int main()
{
int a = 1;
return 0;
}
clang++ -O -g -gsplit-dwarf -gdwarf-5 split-dwarf.cpp; llvm-dwarfdump --verify ./a.out | grep skeleton
error: Compilation unit type (DW_UT_skeleton) and root DIE (DW_TAG_compile_unit) do not match.
The fix is to change DW_TAG_compile_unit into DW_TAG_skeleton_unit when skeleton file is generated.
Differential Revision: https://reviews.llvm.org/D70880
Summary:
This follows a previous patch that changes the X86 datalayout to represent
mixed size pointers (32-bit sext, 32-bit zext, and 64-bit) with address spaces
(https://reviews.llvm.org/D64931)
This patch implements the address space cast lowering to the corresponding
sign extension, zero extension, or truncate instructions.
Related to https://bugs.llvm.org/show_bug.cgi?id=42359
Reviewers: rnk, craig.topper, RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D69639
Revise the coverage mapping format to reduce binary size by:
1. Naming function records and marking them `linkonce_odr`, and
2. Compressing filenames.
This shrinks the size of llc's coverage segment by 82% (334MB -> 62MB)
and speeds up end-to-end single-threaded report generation by 10%. For
reference the compressed name data in llc is 81MB (__llvm_prf_names).
Rationale for changes to the format:
- With the current format, most coverage function records are discarded.
E.g., more than 97% of the records in llc are *duplicate* placeholders
for functions visible-but-not-used in TUs. Placeholders *are* used to
show under-covered functions, but duplicate placeholders waste space.
- We reached general consensus about giving (1) a try at the 2017 code
coverage BoF [1]. The thinking was that using `linkonce_odr` to merge
duplicates is simpler than alternatives like teaching build systems
about a coverage-aware database/module/etc on the side.
- Revising the format is expensive due to the backwards compatibility
requirement, so we might as well compress filenames while we're at it.
This shrinks the encoded filenames in llc by 86% (12MB -> 1.6MB).
See CoverageMappingFormat.rst for the details on what exactly has
changed.
Fixes PR34533 [2], hopefully.
[1] http://lists.llvm.org/pipermail/llvm-dev/2017-October/118428.html
[2] https://bugs.llvm.org/show_bug.cgi?id=34533
Differential Revision: https://reviews.llvm.org/D69471
Summary:
The default case handles the majority of MVTs so most of the individual
cases can be removed. Also added a case for floating point types.
Reviewed By: sdesmalen
Differential Revision: https://reviews.llvm.org/D70955
InstCombine may synthesize FMINNUM/FMAXNUM nodes from fcmp+select
sequences (where the fcmp is marked nnan). Currently, if the
target does not otherwise handle these nodes, they'll get expanded
to libcalls to fmin/fmax. However, these functions may reside in
libm, which may introduce a library dependency that was not originally
present in the source code, potentially resulting in link failures.
To fix this problem, add code to TargetLowering::expandFMINNUM_FMAXNUM
to expand FMINNUM/FMAXNUM to a compare+select sequence instead of the
libcall. This is done only if the node is marked as "nnan"; in this case,
the expansion to compare+select is always correct. This also suffices to
catch all cases where FMINNUM/FMAXNUM was synthesized as above.
Differential Revision: https://reviews.llvm.org/D70965
This is the example:
int foo(int a, int b, int c, int d) {
return a + b + c + d;
}
And this is the Dependency Graph:
+------+ +------+ +------+ +------+
| A | | B | | C | | D |
+--+--++ +---+--+ +--+---+ +--+---+
^ ^ ^ ^ ^ ^
| | | | | |
| | | |New1 +--------------+
| | | | |
| | | | +--+---+
| |New2 | +-------+ ADD1 |
| | | +--+---+
| | | Fuse ^
| | +-------------+
| +------------+
| |
| Fuse +--+---+
+----------->+ ADD2 |
| +------+
+--+---+
| ADD3 |
+------+
We need also create an artificial edge from ADD1 to A if
https://reviews.llvm.org/D69998 is landed. That will force the Node A scheduled
before the ADD1 and ADD2. But in fact, it is ok to schedule the Node A
in-between ADD3 and ADD2, as ADD3 and ADD2 are NOT a fusion pair because
ADD2 has been matched to ADD1. We are creating these unnecessary dependency
edges that override the heuristics.
Differential Revision: https://reviews.llvm.org/D70066
This is an alternative to D64662 that shares more code between
strict and non-strict nodes. It's modeled after the implementation
that I did for softening.
Differential Revision: https://reviews.llvm.org/D70867
https://reviews.llvm.org/D70922
This adds a hook to allow targets to define exactly what extension
operation should be performed for widening constants. This handles cases
like widening i1 true which would end up becoming -1 which affects code
quality during combines.
Additionally, in order to stay consistent with how DAG is promoting
constants, we now signextend for byte sized types and zero extend
otherwise (by default). Targets can of course override this if
necessary.
As it can be seen from accompanying cleanup, it is not unheard of
to write `~Known.Zero` meaning "what maximal value can this KnownBits
produce". But i think `~Known.Zero` isn't *that* self-explanatory,
as compared to a method with a name.
Note that not all `~Known.Zero` places were cleaned up,
only those where this arguably improves things.
The DebugVariable class is a class declared in LiveDebugValues.cpp which
is used to uniquely identify a single variable, using its source
variable, inline location, and fragment info to do so. This patch moves
this class into DebugInfoMetadata.h, making it available in a much
broader scope.
Convert ARMCodeGenPrepare into a generic type promotion pass by:
- Removing the insertion of arm specific intrinsics to handle narrow
types as we weren't using this.
- Removing ARMSubtarget references.
- Now query a generic TLI object to know which types should be
promoted and what they should be promoted to.
- Move all codegen tests into Transforms folder and testing using opt
and not llc, which is how they should have been written in the
first place...
The pass searches up from icmp operands in an attempt to safely
promote types so we can avoid generating unnecessary unsigned extends
during DAG ISel.
Differential Revision: https://reviews.llvm.org/D69556
The idea is to remove front-end analysis for the parameter's value
modification and leave it to the value tracking system. Front-end in some
cases marks a parameter as modified even the line of code that modifies the
parameter gets optimized, that implies that this will cover more entry
values even. In addition, extending the support for modified parameters
will be easier with this approach.
Since the goal is to recognize if a parameter’s value has changed, the idea
at very high level is: If we encounter a DBG_VALUE other than the entry
value one describing the same variable (parameter), we can assume that the
variable’s value has changed and we should not track its entry value any
more. That would be ideal scenario, but due to various LLVM optimizations,
a variable’s value could be just moved around from one register to another
(and there will be additional DBG_VALUEs describing the same variable), so
we have to recognize such situation (otherwise, we will lose a lot of entry
values) and salvage the debug entry value.
Differential Revision: https://reviews.llvm.org/D68209
While working with a patch for instruction selection, the splitting of a
large immediate ended up begin treated incorrectly by the backend. Where a
register operand should have been created, it instead became an immediate. To
my surprise the machine verifier failed to report this, which at the time
would have been helpful.
This patch improves the verifier so that it will report this type of error.
This patch XFAILs CodeGen/SPARC/fp128.ll, which has been reported at
https://bugs.llvm.org/show_bug.cgi?id=44091
Review: thegameg, arsenm, fhahn
https://reviews.llvm.org/D63973
These nodes have a FIXME that they only get here because a Custom
handler returned SDValue() instead of the original Op.
Even though we aren't expanding them, we should return true here to
prevent ConvertNodeToLibcall from also trying to process them until
the FIXME has been addressed.
I'm hoping to add checking to ConvertNodeToLibcall to make sure
we don't give it nodes it doesn't have support for.