The previous code was incorrect (can't getReg a frameindex). We could instead optimize it to reduce tree height, but I'm not sure that's worthwhile yet because we then try to eliminate the frameindex.
This patch also fixes frame index elimination for operations which may load or store: it used to assume the base was operand 2 and immediate offset operand 1. That's not true for stores, where they're 4 and 3.
llvm-svn: 259305
The AMDGPUPromoteAlloca pass was emitting the read.local.size
calls, which with HSA was incorrectly selected to reading from
the offset mesa uses off of the kernarg pointer.
Error on intrinsics which aren't supported by HSA, and start
emitting the correct IR to read the workgroup size
out of the dispatch pointer.
Also initialize the pass so it can be tested with opt, and
start moving towards not depending on the subtarget as an
argument.
Start emitting errors for the intrinsics not handled with HSA.
llvm-svn: 259297
Only the dispatch.ptr intrinsic is supposed to be used now to get
the workgroup size, and the read.local.size intrinsics do not
work correctly.
llvm-svn: 259296
Refine the test for whether an instruction is in an expression tree so that
it detects when one tree ends and another begins, so we can place a block
at that point, rather than continuing to find the first instruction not in
a tree at all.
llvm-svn: 259294
Add an option to llvm-profdata merge for writing out sparse indexed
profiles. These profiles omit InstrProfRecords for functions which are
never executed.
Differential Revision: http://reviews.llvm.org/D16727
llvm-svn: 259258
Loop transformations can sometimes fail because the loop, while in
valid rotated LCSSA form, is not in a canonical CFG form. This is
an extremely simple pass that just merges obviously redundant
blocks, which can be used to fix some known failure cases. In the
future, it may be enhanced with more cases (and have code shared with
SimplifyCFG).
This allows us to run LoopSimplifyCFG -> LoopRotate -> LoopUnroll,
so that SimplifyCFG cleans up the loop before Rotate tries to run.
Not currently used in the pass manager, since this pass doesn't do
anything unless you can hook it up in an LPM with other loop passes.
It'll be added once Chandler cleans up things to allow this.
Tested in a custom pipeline out of tree to confirm it works in
practice (in addition to the included trivial test).
llvm-svn: 259256
The majority of attribute queries checks for the existence of an enum
attribute in the FunctionIndex slot. We only have 48 of those and can
therefore summarize them in an uint64_t bitset which measurably improves
compile time.
Differential Revision: http://reviews.llvm.org/D16618
llvm-svn: 259252
The majority of queries just checks for the existince of an enum
attribute. We only have 48 of those and can summaryiz them in an
uint64_t bitfield so we can avoid searching the list. This improves
"opt" compile time by 1-4% in my measurements.
Differential Revision: http://reviews.llvm.org/D16617
llvm-svn: 259251
We would infinite loop because we created a shufflevector that was wider than
needed and then failed to combine that with the insertelement. When subsequently
visiting the extractelement from that shuffle, we see that it's unnecessary,
delete it, and trigger another visit to the insertelement.
llvm-svn: 259236
This support is _very_ rudimentary, just enough to get some basic data
into the CodeView debug section.
Left to do is:
- Use the combined opcodes to save space.
- Do something about code offsets.
llvm-svn: 259230
The basic optimisation was to convert (mul $LHS, $complex_constant) into
roughly "(shl (mul $LHS, $simple_constant), $simple_amt)" when it was expected
to be cheaper. The original logic checks that the mul only has one use (since
we're mangling $complex_constant), but when used in even more complex
addressing modes there may be an outer addition that can pick up the wrong
value too.
I *think* the ARM addressing-mode problem is actually unreachable at the
moment, but that depends on complex assessments of the profitability of
pre-increment addressing modes so I've put a real check in there instead of an
assertion.
llvm-svn: 259228
Add support for frame pointer use in prolog/epilog.
Supports dynamic allocas but not yet over-aligned locals.
Target-independend CG generates SP updates, but we still need to write
back the SP value to memory when necessary.
llvm-svn: 259220
Summary:
There are three parts to inlined call frames:
1. The inlinee line subsection
2. The inline site symbol record
3. The function ids referenced by both
This change starts by emitting function ids (3) for all subprograms and
emitting the base inline site symbol record (2). The actual line numbers
in (2) use an encoded format that will come next, along with the inlinee
line subsection.
Reviewers: majnemer
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D16333
llvm-svn: 259217
- Locally declare struct, and call it BaseDerivedPair
- Use a lambda to compare, instead of a singleton with uninitialized
fields
- Add a constructor to BaseDerivedPair and use SmallVector::emplace_back
llvm-svn: 259208
The buildSchedGraph() was in need of reworking as the AA features had been
added on top of earlier code. It was very difficult to understand, and buggy.
There had been found cases where scheduling dependencies had actually been
missed (see r228686).
AliasChain, RejectMemNodes, adjustChainDeps() and iterateChainSucc() have
been removed. There are instead now just the four maps from Value to SUs, which
have been renamed to Stores, Loads, NonAliasStores and NonAliasLoads.
An unknown store used to become the AliasChain, but now becomes a store mapped
to 'unknownValue' (in Stores). What used to be PendingLoads is instead the
list of SUs mapped to 'unknownValue' in Loads.
RejectMemNodes and adjustChainDeps() used to be a safety-net for everything.
The SU maps were sometimes cleared and SUs were put in RejectMemNodes, where
adjustChainDeps() would look. Instead of this, a more straight forward approach
is used in maintaining the SU maps without clearing them and simply letting
them grow over time. Instead of the cutt-off in adjustChainDeps() search, a
reduction of maps will be done if needed (see below).
Each SUnit either becomes the BarrierChain, or is put into one of the maps. For
each SUnit encountered, all the information about previous ones are still
available until a new BarrierChain is set, at which point the maps are cleared.
For huge regions, the algorithm becomes slow, therefore the maps will get
reduced at a threshold (current default is 1000 nodes), by a fraction (default 1/2).
These values can be tuned by use of CL options in case some test case shows that
they need to be changed (-dag-maps-huge-region and -dag-maps-reduction-size).
There has not been any considerable change observed in output quality or compile
time. There may now be more DAG edges inserted than before (i.e. if A->B->C,
then A->C is not needed). However, in a comparison run there were fewer total
calls to AA, and a somewhat improved compile time, which means this seems to
be not a problem.
http://reviews.llvm.org/D8705
Reviewers: Hal Finkel, Andy Trick.
llvm-svn: 259201
The trap instruction is emitted as a data-in-text rather
than an instruction. This patch uses the .inst directive
for emitting trap.
Differential Revision: http://reviews.llvm.org/D16684
llvm-svn: 259182
check that the sign extended constant fits into 16-bits if we want a
zero extended value, otherwise go ahead and put it together piecemeal.
Fixes PR26356.
llvm-svn: 259177
This patch enables llvm-bcanalyzer to print the bitcode wrapper header
if the file has one, which is needed to test the changes made in
r258627 (bitcode-wrapper-header-armv7m.ll is the test case for r258627).
Differential Revision: http://reviews.llvm.org/D16642
llvm-svn: 259162
