give PATypeHolder an explicit copy ctor which initializes the type pointer,
and make PATypeHolder work with null pointers.

The implicitly generated one didn't work on numerous levels, but was still
accepted, allowing all sorts of bugs with default constructed pa type holders.

Previously, they "sort of" worked if they were default constructed and then
destructed.  Now they really work, and you can even default construct one,
then assign to it, amazing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107195 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the handling of partial redefines in the fast register allocator.

A partial redefine needs to be treated like a tied operand, and the register
must be reloaded while processing use operands.

This fixes a bug where partially redefined registers were processed as normal
defs with a reload added. The reload could clobber another use operand if it was
a kill that allowed register reuse.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107193 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a register scavenger crash when dealing with undefined subregs.
The LowerSubregs pass needs to preserve implicit def operands attached to
EXTRACT_SUBREG instructions when it replaces those instructions with copies.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107189 91177308-0d34-0410-b5e6-96231b3b80d8

Add AVX non-temporal stores

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107178 91177308-0d34-0410-b5e6-96231b3b80d8

Add a few more interesting testcases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107177 91177308-0d34-0410-b5e6-96231b3b80d8

Fix whitespace style.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107175 91177308-0d34-0410-b5e6-96231b3b80d8

Move non-temporal movs to their own section

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107168 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for encoding NEON VMOV (from core register to scalar) instructions.
The encoding is the same as VMOV (from scalar to core register) except that
the operands are in different places.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107167 91177308-0d34-0410-b5e6-96231b3b80d8

Add sqrt, rsqrt and rcp AVX instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107166 91177308-0d34-0410-b5e6-96231b3b80d8

skip dbg_value instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107154 91177308-0d34-0410-b5e6-96231b3b80d8

minor cosmetics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107148 91177308-0d34-0410-b5e6-96231b3b80d8

The t2MOVi16 and t2MOVTi16 instructions do not set CPSR.  Trying to add
a CPSR operand to them causes an assertion failure, so apparently these
instructions haven't been getting a lot of use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107147 91177308-0d34-0410-b5e6-96231b3b80d8

use ArgOffset constant to prepare for operand rotation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107146 91177308-0d34-0410-b5e6-96231b3b80d8

use ArgOperand API

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107145 91177308-0d34-0410-b5e6-96231b3b80d8

use ArgOperand API

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107144 91177308-0d34-0410-b5e6-96231b3b80d8

Return Changed.  This required setting Changed if dbg metadata
is stripped off.  Currently set unconditionally, since the API
does not provide a way of working out if anything was actually
stripped off.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107142 91177308-0d34-0410-b5e6-96231b3b80d8

It seems clear that this should return Changed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107141 91177308-0d34-0410-b5e6-96231b3b80d8

Add a VT argument to getMinimalPhysRegClass and replace the copy related uses
of getPhysicalRegisterRegClass with it.

If we want to make a copy (or estimate its cost), it is better to use the
smallest class as more efficient operations might be possible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107140 91177308-0d34-0410-b5e6-96231b3b80d8

getMachineBasicBlockAddress returns a uintptr_t - don't truncate
to unsigned only to extend back to a pointer sized value on the
next line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107139 91177308-0d34-0410-b5e6-96231b3b80d8

The variable ValueSize is set to 1 on both code paths, and then
ignored!  Remove it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107138 91177308-0d34-0410-b5e6-96231b3b80d8

The variable "Value" is carefully set to Layout.getSymbolAddress,
but then not actually used - maybe a bug?  Remove the variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107137 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused calls to Lexer.getLoc and the pointless variable HasFillExpr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107136 91177308-0d34-0410-b5e6-96231b3b80d8

Remove pointless variable LastDef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107135 91177308-0d34-0410-b5e6-96231b3b80d8

Looks like this test is missing an XFAIL line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107134 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused variable Loc and pointless variables unified_syntax
and thumb_mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107133 91177308-0d34-0410-b5e6-96231b3b80d8

use ArgOperand APIs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107132 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an unused and a pointless variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107131 91177308-0d34-0410-b5e6-96231b3b80d8

Remove pointless and unused variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107130 91177308-0d34-0410-b5e6-96231b3b80d8

encode operand initializations (at fixed index)
in terms of Op<> and ArgOffset. This works for
values of {0, 1} for ArgOffset.
Please note that ArgOffset will become 0 soon and
will go away eventually.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107129 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a pointless variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107128 91177308-0d34-0410-b5e6-96231b3b80d8

Remove initialized but otherwise unused variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107127 91177308-0d34-0410-b5e6-96231b3b80d8

Remove variables that are written by not read.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107126 91177308-0d34-0410-b5e6-96231b3b80d8

Use a more obvious way to avoid compiling functions which are only used when XDEBUG is enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107125 91177308-0d34-0410-b5e6-96231b3b80d8

Jump through some silly hoops to make GCC accept that a function may not always
be called.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107124 91177308-0d34-0410-b5e6-96231b3b80d8

PR7503: uxtb16 is not available for ARMv7-M. Patch by Brian G. Lucas.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107122 91177308-0d34-0410-b5e6-96231b3b80d8

Change if-cvt options to something that actually as useable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107121 91177308-0d34-0410-b5e6-96231b3b80d8

When processing loops for scheduling latencies (used for live outs on loop
back-edges), make sure not to include dbg_value instructions in the count.
Closing in on the end of rdar://7797940

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107119 91177308-0d34-0410-b5e6-96231b3b80d8

Just as its not safe to blindly transfer the nsw bit from an add
instruction to an add scev, it's not safe to blindly transfer the
inbounds flag from a gep instruction to an nsw on the scev for the
gep.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107117 91177308-0d34-0410-b5e6-96231b3b80d8

Refactoring of arithmetic instruction classes with unary operator

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107116 91177308-0d34-0410-b5e6-96231b3b80d8

When no memoperands are present, assume unaligned, volatile.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107114 91177308-0d34-0410-b5e6-96231b3b80d8

Strip resulting binaries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107112 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply my if-conversion cleanup from svn r106939 with fixes.
There are 2 changes relative to the previous version of the patch:

1) For the "simple" if-conversion case, there's no need to worry about
RemoveExtraEdges not handling an unanalyzable branch.  Predicated terminators
are ignored in this context, so RemoveExtraEdges does the right thing.
This might break someday if we ever treat indirect branches (BRIND) as
predicable, but for now, I just removed this part of the patch, because
in the case where we do not add an unconditional branch, we rely on keeping
the fall-through edge to CvtBBI (which is empty after this transformation).

The change relative to the previous patch is:

@@ -1036,10 +1036,6 @@
     IterIfcvt = false;
   }

-  // RemoveExtraEdges won't work if the block has an unanalyzable branch,
-  // which is typically the case for IfConvertSimple, so explicitly remove
-  // CvtBBI as a successor.
-  BBI.BB->removeSuccessor(CvtBBI->BB);
   RemoveExtraEdges(BBI);

   // Update block info. BB can be iteratively if-converted.

2) My patch exposed a bug in the code for merging the tail of a "diamond",
which had previously never been exercised.  The code was simply checking that
the tail had a single predecessor, but there was a case in
MultiSource/Benchmarks/VersaBench/dbms where that single predecessor was
neither edge of the diamond.  I added the following change to check for
that:

@@ -1276,7 +1276,18 @@
   // tail, add a unconditional branch to it.
   if (TailBB) {
     BBInfo TailBBI = BBAnalysis[TailBB->getNumber()];
-    if (TailBB->pred_size() == 1 && !TailBBI.HasFallThrough) {
+    bool CanMergeTail = !TailBBI.HasFallThrough;
+    // There may still be a fall-through edge from BBI1 or BBI2 to TailBB;
+    // check if there are any other predecessors besides those.
+    unsigned NumPreds = TailBB->pred_size();
+    if (NumPreds > 1)
+      CanMergeTail = false;
+    else if (NumPreds == 1 && CanMergeTail) {
+      MachineBasicBlock::pred_iterator PI = TailBB->pred_begin();
+      if (*PI != BBI1->BB && *PI != BBI2->BB)
+        CanMergeTail = false;
+    }
+    if (CanMergeTail) {
       MergeBlocks(BBI, TailBBI);
       TailBBI.IsDone = true;
     } else {

With these fixes, I was able to run all the SingleSource and MultiSource
tests successfully.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107110 91177308-0d34-0410-b5e6-96231b3b80d8

Add an Intraprocedural form of BasicAliasAnalysis, which aims to
properly handles instructions and arguments defined in different
functions, or across recursive function iterations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107109 91177308-0d34-0410-b5e6-96231b3b80d8

Described the missing AVX forms of SSE2 convert instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107108 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Thumb encoding of VMOV (scalar to ARM core register).  The encoding is
the same as ARM except that the condition code field is always set to ARMCC::AL.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107107 91177308-0d34-0410-b5e6-96231b3b80d8

The comment string does not match for all targets. PowerPC uses ;.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107103 91177308-0d34-0410-b5e6-96231b3b80d8

Unlike other targets, ARM now uses BUILD_VECTORs post-legalization so they
can't be changed arbitrarily by the DAGCombiner without checking if it is
running after legalization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107097 91177308-0d34-0410-b5e6-96231b3b80d8

Refix XTARGET.  Previous attempt matches on powerpc-apple-darwin,
although I don't see why.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107090 91177308-0d34-0410-b5e6-96231b3b80d8

Attempt to fix XTARGET.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107088 91177308-0d34-0410-b5e6-96231b3b80d8

Make the ARMCodeEmitter identify Thumb functions via ARMFunctionInfo instead
of the Subtarget.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107086 91177308-0d34-0410-b5e6-96231b3b80d8

Use DW_FORM_addr for DW_AT_entry_pc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107085 91177308-0d34-0410-b5e6-96231b3b80d8

Add a blurb about -scev-aa.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107080 91177308-0d34-0410-b5e6-96231b3b80d8

In asm's, output operands with matching input constraints
have to be registers, per gcc documentation.  This affects
the logic for determining what "g" should lower to.  PR 7393.
A couple of existing testcases are affected.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107079 91177308-0d34-0410-b5e6-96231b3b80d8

Added the darwin .secure_log_unique and .secure_log_reset directives.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107077 91177308-0d34-0410-b5e6-96231b3b80d8

Constant fold x == undef to undef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107074 91177308-0d34-0410-b5e6-96231b3b80d8

tidy up style. no functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107073 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Value::stripPointerCasts and BasicAA to avoid trouble on
code in unreachable blocks, which have have use-def cycles.
This fixes PR7514.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107071 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor encoding function for NEON 1-register with modified immediate format.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107070 91177308-0d34-0410-b5e6-96231b3b80d8

Support Thumb mode encoding of NEON instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107068 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce indentation via early exit. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107067 91177308-0d34-0410-b5e6-96231b3b80d8

Include inlined function in list of processed subprograms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107065 91177308-0d34-0410-b5e6-96231b3b80d8

new, no longer brain-dead, r106907

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107060 91177308-0d34-0410-b5e6-96231b3b80d8

Remove this weak test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107059 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for llvm-gcc fix 107051.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107052 91177308-0d34-0410-b5e6-96231b3b80d8

Don't write temporary files in test directory

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107049 91177308-0d34-0410-b5e6-96231b3b80d8

After physreg coalescing, physical registers might not have live ranges where
you would expect.

Don't assert on that case, just give up.

This fixes PR7513.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107046 91177308-0d34-0410-b5e6-96231b3b80d8

Add a triple so test runs on Linux as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107045 91177308-0d34-0410-b5e6-96231b3b80d8

Add more special treatment for inline asm in RegAllocFast.

When an instruction has tied operands and physreg defines, we must take extra
care that the tied operands conflict with neither physreg defs nor uses.

The special treatment is given to inline asm and instructions with tied operands
/ early clobbers and physreg defines.

This fixes PR7509.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107043 91177308-0d34-0410-b5e6-96231b3b80d8

Fix thinko.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107042 91177308-0d34-0410-b5e6-96231b3b80d8

Pull in the libCrashReporterClient.a information with a warning comment.

Remove library check and regenerate configure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107028 91177308-0d34-0410-b5e6-96231b3b80d8

Preserve deleted function's local variables' debug info.
Radar 8122864.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107027 91177308-0d34-0410-b5e6-96231b3b80d8

Make this test darwin specific.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107025 91177308-0d34-0410-b5e6-96231b3b80d8

use ArgOperand API

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107017 91177308-0d34-0410-b5e6-96231b3b80d8

use ArgOperand API

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107016 91177308-0d34-0410-b5e6-96231b3b80d8

employ CallInst::ArgOffset (for now)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107015 91177308-0d34-0410-b5e6-96231b3b80d8

simplify: we have solid argument iterator range

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107014 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize AAEval so that it can be used both per-function and
interprocedurally. Note that as of this writing, existing alias
analysis passes are not prepared to be used interprocedurally.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107013 91177308-0d34-0410-b5e6-96231b3b80d8

Fix this build message so that it displays the correct library
name, specifically the "lib" prefix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107011 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r106907, "make sure to handle dbg_value instructions in the middle of the
block, not...", it caused a bunch of nightly test regressions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107009 91177308-0d34-0410-b5e6-96231b3b80d8

use setArgOperand

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107004 91177308-0d34-0410-b5e6-96231b3b80d8

use CallInst::ArgOffset

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107003 91177308-0d34-0410-b5e6-96231b3b80d8

use ArgOperand API and CallInst::ArgOffset

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107002 91177308-0d34-0410-b5e6-96231b3b80d8

extend ArgOperand interface: setArgOperand
(in both CallInst and InvokeInst)

also add a (short-lived) constant to CallInst, that names
the operand index of the first call argument. This is
strictly transitional and should not be used for new code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107001 91177308-0d34-0410-b5e6-96231b3b80d8

use cached value

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107000 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106990 91177308-0d34-0410-b5e6-96231b3b80d8

Use named MDNode, llvm.dbg.sp, to collect subprogram info. This will be used to emit  local variable's debug info of deleted functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106989 91177308-0d34-0410-b5e6-96231b3b80d8

minor housekeeping cleanup: 80-column, trailing whitespace, spelling, etc.. No functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106988 91177308-0d34-0410-b5e6-96231b3b80d8

Do not forget last element, function, while creating Subprogram definition MDNode from subprogram declare MDNode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106985 91177308-0d34-0410-b5e6-96231b3b80d8

minor cleanup to SROA: when lowering type unsafe accesses to
large integers, the first inserted value would always create
an 'or X, 0'.  Even though this is trivially zapped by
instcombine, don't bother creating this pointless instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106979 91177308-0d34-0410-b5e6-96231b3b80d8

add some named accessors for StoreInst

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106969 91177308-0d34-0410-b5e6-96231b3b80d8

fit in 80 cols

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106968 91177308-0d34-0410-b5e6-96231b3b80d8

this test is failing nondeterministically and blaming me, just disable
it for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106960 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test weirdness.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106959 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some tests that didn't test anything.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106954 91177308-0d34-0410-b5e6-96231b3b80d8

Partial specialization test should not depend on the order of specialization operations or the names assigned to the specialized functions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106953 91177308-0d34-0410-b5e6-96231b3b80d8

When splitting a VAARG, remember its alignment.
This produces terrible but correct code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106952 91177308-0d34-0410-b5e6-96231b3b80d8

Revert my if-conversion cleanup since it caused a bunch of nightly test
regressions.

--- Reverse-merging r106939 into '.':
U    test/CodeGen/Thumb2/thumb2-ifcvt3.ll
U    lib/CodeGen/IfConversion.cpp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106951 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR7328: when turning a tail recursion into a loop, need to preserve
the returned value after the tail call if it differs from other return
values.  The optimal thing to do would be to introduce a phi node for
the return value, but for the moment just fix the miscompile.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106947 91177308-0d34-0410-b5e6-96231b3b80d8

use ArgOperand API

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106946 91177308-0d34-0410-b5e6-96231b3b80d8

use ArgOperand API

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106945 91177308-0d34-0410-b5e6-96231b3b80d8