As noted in the code comment, transforming this in the other direction might require
a separate transform here in CGP given the block-at-a-time DAG constraint.
Besides that theoretical motivation, there are 2 practical motivations for the
subtract-of-cmps form:
1. The codegen for both x86 and PPC is better for this IR (though PPC could be better still).
There is discussion about canonicalizing IR to the select form
( http://lists.llvm.org/pipermail/llvm-dev/2017-July/114885.html ),
so we probably need to add DAG transforms for those patterns anyway, but this improves the
memcmp output without waiting for that step.
2. If we allow vector-sized chunks for the load and compare, x86 is better prepared to convert
that to optimal code when using subtract-of-cmps, so another prerequisite patch is avoided
if we choose to enable that.
Differential Revision: https://reviews.llvm.org/D34904
llvm-svn: 309597
D35067/rL308322 attempted to support up to 4 load pairs for memcmp inlining which resulted in regressions for some optimized libc memcmp implementations (PR33914).
Until we can match these more optimal cases, this patch reduces the memcmp expansion to a maximum of 2 load pairs (which matches what we do for -Os).
This patch should be considered for the 5.0.0 release branch as well
Differential Revision: https://reviews.llvm.org/D35830
llvm-svn: 308986
Allowing cycles in Phi traversal increases the scope of optimize memory instruction
in case we are in loop.
The added test shows an example of enabling optimization inside a loop.
Reviewers: loladiro, spatel, efriedma
Reviewed By: efriedma
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D35294
llvm-svn: 308419
It should be a win to avoid going out to the system lib for all small memcmp() calls using scalar ops. For x86 32-bit, this means most everything up to 16 bytes. For 64-bit, that doubles because we can do 8-byte loads.
Notes:
Reduced from 4 to 2 loads for -Os behavior, which might not be optimal in all cases. It's effectively a question of how much do we trust the system implementation. Linux and macOS (and Windows I assume, but did not test) have optimized memcmp() code for x86, so it's probably not bad either way? PPC is using 8/4 for defaults on these. We do not expand at all for -Oz.
There are still potential improvements to make for the CGP expansion IR and/or lowering such as avoiding select-of-constants (D34904) and not doing zexts to the max load type before doing a compare.
We have special-case SSE/AVX codegen for (memcmp(x, y, 16/32) == 0) that will no longer be produced after this patch. I've shown the experimental justification for that change in PR33329:
https://bugs.llvm.org/show_bug.cgi?id=33329#c12
TLDR: While the vector code is a likely winner, we can't guarantee that it's a winner in all cases on all CPUs, so I'm willing to sacrifice it for the greater good of expanding all small memcmp(). If we want to resurrect that codegen, it can be done by adjusting the CGP params or poking a hole to let those fall-through the CGP expansion.
Committed on behalf of Sanjay Patel
Differential Revision: https://reviews.llvm.org/D35067
llvm-svn: 308322
When we fail to sink an instruction, we must make sure not to modify
the function; otherwise, we end up in an infinite loop because
CodeGenPrepare iterates until it doesn't make any changes.
Fixes https://bugs.llvm.org/show_bug.cgi?id=33608 .
llvm-svn: 307866
As promised in D35003.
Uses -codegenprepare instead of -instcombine since we hit the same
buggy path anyway, and CGP lets us keep this test really simple
(instcombine likes turning the alloca T, N into alloca [N x T], which
hides the bug this is testing for).
llvm-svn: 307811
This was auto-generated using an older version of the script,
and that version does not work with phis, so if we enable
expansion it will go bad.
llvm-svn: 307267
Summary:
Arguably non-integral pointers probably shouldn't show up here at all,
but since the backend doesn't complain and this takes valid (according
to the Verifier) IR and makes it invalid, make sure not to introduce
any inttoptr instructions if we're dealing with non-integral pointers.
Reviewed By: sanjoy
Differential Revision: https://reviews.llvm.org/D33110
llvm-svn: 306737
As noted in D34071, there are some IR optimization opportunities that could be
handled by normal IR passes if this expansion wasn't happening so late in CGP.
Regardless of that, it seems wasteful to knowingly produce suboptimal IR here,
so I'm proposing this change:
%s = sub i32 %x, %y
%r = icmp ne %s, 0
=>
%r = icmp ne %x, %y
Changing the predicate to 'eq' mimics what InstCombine would do, so that's just
an efficiency improvement if we decide this expansion should happen sooner.
The fact that the PowerPC backend doesn't eliminate the 'subf.' might be
something for PPC folks to investigate separately.
Differential Revision: https://reviews.llvm.org/D34416
llvm-svn: 306471
I don't think there's any visible difference from having the wrong layout
for the 32-bit case at this point, but that could change in the future.
llvm-svn: 305931
There are a couple of potential improvements as seen in the IR and asm:
1. We're unnecessarily extending to a larger type to compare values.
2. The codegen for (select cond, 1, -1) could avoid a cmov.
(or we could change the order of the compares, so we have a select with 0 operand)
llvm-svn: 305802
No IR tests were added with rL304313 ( https://reviews.llvm.org/D28637 ),
so I want these for extra coverage if we enable memcmp expansion for x86.
As shown, nothing is expanded for x86 in CGP yet.
Also fundamentally, we're doing an IR transform, so we should have IR tests
for just that part. If something goes wrong, we need to know if the bug is
in CGP or later lowering.
llvm-svn: 305011
Summary:
Don't use the metadata on call instructions for determining hotness
unless we are in sample PGO mode, where it is needed because profile
counts are not accurate. In instrumentation mode this is not necessary
and does more harm than good when calls have VP metadata that hasn't
been properly scaled after transformations or dropped after constant
prop based devirtualization (both should be fixed, but we don't need
to do this in the first place for instrumentation PGO).
This required adjusting a number of tests to distinguish between sample
and instrumentation PGO handling, and to add in profile summary metadata
so that getProfileCount can get the summary.
Reviewers: davidxl, danielcdh
Subscribers: aemerson, rengolin, mehdi_amini, Prazek, llvm-commits
Differential Revision: https://reviews.llvm.org/D32877
llvm-svn: 302844
Summary:
r284533 added hot and cold section prefixes based on profile
information, to enable grouping of hot/cold functions at link time.
However, it used "cold" as the prefix for cold sections, but gold only
recognizes "unlikely" (which is used by gcc for cold sections).
Therefore, cold sections were not properly being grouped. Switch to
using "unlikely"
Reviewers: danielcdh, davidxl
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D32983
llvm-svn: 302502
The splitIndirectCriticalEdges function generates and invalid CFG when the
'Target' basic block is a loop to itself. When this occurs, the code that
updates the predecessor terminator needs to update the terminator in the split
basic block.
This occurs when there is an edge from block D back to D. Since D is split in
to D0 and D1, the code needs to update the terminator in D1. But D1 is not in
the OtherPreds vector, so it was not getting updated.
Differential Revision: https://reviews.llvm.org/D32126
llvm-svn: 300480
The new codepath has been in the tree for years, and there isn't any
reason to use two codepaths here.
Differential Revision: https://reviews.llvm.org/D30596
llvm-svn: 299723
Disable bypassing if one of the operands looks like a hash value. Slow
division often occurs in hashtable implementations and fast division is
never taken there because a hash value is extremely unlikely to have
enough upper bits set to zero.
A value is considered to be hash-like if it is produced by
1) XOR operation
2) Multiplication by a constant wider than the shorter type
3) PHI node with all incoming values being hash-like
Differential Revision: https://reviews.llvm.org/D28200
llvm-svn: 299329
Summary: The current prefix based function layout algorithm only looks at function's entry count, which is not sufficient. A function should be grouped together if its entry count or any call edge count is hot.
Reviewers: davidxl, eraman
Reviewed By: eraman
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D31225
llvm-svn: 298656
Currently the default C calling convention functions are treated
the same as compute kernels. Make this explicit so the default
calling convention can be changed to a non-kernel.
Converted with perl -pi -e 's/define void/define amdgpu_kernel void/'
on the relevant test directories (and undoing in one place that actually
wanted a non-kernel).
llvm-svn: 298444
This adds a parameter to @llvm.objectsize that makes it return
conservative values if it's given null.
This fixes PR23277.
Differential Revision: https://reviews.llvm.org/D28494
llvm-svn: 298430
ValueTracking is used for more thorough analysis of operands. Based on the
analysis, either run-time checks can be simplified (e.g. check only one operand
instead of two) or the transformation can be avoided. For example, it is quite
often the case that a divisor is promoted from a shorter type and run-time
checks for it are redundant.
With additional compile-time analysis of values, two special cases naturally
arise and are addressed by the patch:
1) Both operands are known to be short enough. Then, the long division can be
simply replaced with a short one without CFG modification.
2) If a division is unsigned and the dividend is known to be short then the
long division is not needed at all. Because if the divisor is too big for
short division then the quotient is obviously zero (and the remainder is
equal to the dividend). Actually, the division is not needed when
(divisor > dividend).
Differential Revision: https://reviews.llvm.org/D29897
llvm-svn: 296832
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
llvm-svn: 296416
When we construct addressing modes, we use isNoopAddrSpaceCast to ignore
addrspacecast instructions. Make sure we insert the correct addrspacecast
when we reconstruct the addressing mode.
Differential Revision: https://reviews.llvm.org/D30114
llvm-svn: 296167
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
llvm-svn: 296149
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
llvm-svn: 296060
We're currently doing nearly the same thing for @llvm.objectsize in
three different places: two of them are missing checks for overflow,
and one of them could subtly break if InstCombine gets much smarter
about removing alloc sites. Seems like a good idea to not do that.
llvm-svn: 290214
This is recommit of r287553 after fixing the invalid loop info after eliminating an empty block and unit test failures in AVR and WebAssembly :
Summary: Merging an empty case block into the header block of switch could cause ISel to add COPY instructions in the header of switch, instead of the case block, if the case block is used as an incoming block of a PHI. This could potentially increase dynamic instructions, especially when the switch is in a loop. I added a test case which was reduced from the benchmark I was targetting.
Reviewers: t.p.northover, mcrosier, manmanren, wmi, joerg, davidxl
Subscribers: joerg, qcolombet, danielcdh, hfinkel, mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D22696
llvm-svn: 289988
`dropUnknownNonDebugMetadata` takes a list of "known" metadata IDs. The
only reason it worked at all is that `getMetadataID` returns something
unrelated -- it returns the subclass ID of the receiver (which is used
in `dyn_cast` etc.). That does not numerically match
`LLVMContext::MD_invariant_group` and ends up dropping `invariant_group`
along with every other metadata that does not numerically match
`LLVMContext::MD_invariant_group`.
llvm-svn: 289973
This is recommit of r287553 after fixing the invalid loop info after eliminating an empty block:
Summary: Merging an empty case block into the header block of switch could cause ISel to add COPY instructions in the header of switch, instead of the case block, if the case block is used as an incoming block of a PHI. This could potentially increase dynamic instructions, especially when the switch is in a loop. I added a test case which was reduced from the benchmark I was targetting.
Reviewers: t.p.northover, mcrosier, manmanren, wmi, joerg, davidxl
Subscribers: joerg, qcolombet, danielcdh, hfinkel, mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D22696
llvm-svn: 289951
Summary:
Previously, CGP would unconditionally sink addrspacecast instructions,
even going so far as to sink them into a loop.
Now we check that the cast is "cheap", as defined by TLI.
We introduce a new "is-cheap" function to TLI rather than using
isNopAddrSpaceCast because some GPU platforms want the ability to ask
for non-nop casts to be sunk.
Reviewers: arsenm, tra
Subscribers: jholewinski, wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26923
llvm-svn: 287591
Summary: Merging an empty case block into the header block of switch could cause
ISel to add COPY instructions in the header of switch, instead of the case
block, if the case block is used as an incoming block of a PHI. This could
potentially increase dynamic instructions, especially when the switch is in a
loop. I added a test case which was reduced from the benchmark I was targetting.
Reviewers: t.p.northover, mcrosier, manmanren, wmi, davidxl
Subscribers: qcolombet, danielcdh, hfinkel, mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D22696
llvm-svn: 287553
Summary:
We don't do BypassSlowDivision when the denominator is a constant, but
we do do it when the numerator is a constant.
This patch makes two related changes to BypassSlowDivision when the
numerator is a constant:
* If the numerator is too large to fit into the bypass width, don't
bypass slow division (because we'll never run the smaller-width
code).
* If we bypass slow division where the numerator is a constant, don't
OR together the numerator and denominator when determining whether
both operands fit within the bypass width. We need to check only the
denominator.
Reviewers: tra
Subscribers: llvm-commits, jholewinski
Differential Revision: https://reviews.llvm.org/D26699
llvm-svn: 287062
Summary:
This "pass" eagerly creates div and rem instructions even when only one
is needed -- it relies on a later pass (machine DCE?) to clean them up.
This is problematic not just from a cleanliness perspective (this pass
is running during CodeGenPrepare, so should leave the IR in a better
state), but it also creates a problem for instruction selection. If we
always have a div+rem, isel will always select a divrem instruction (if
possible), even when a single div or rem would do.
Specifically, in NVPTX, we want to compute rem from the output of div,
if available. But if a div is not available, we want to leave the rem
alone. This transformation is overeager if div is always available.
Because this code runs as part of CodeGenPrepare, it's nontrivial to
write a test for this change. But this will effectively be tested by
a later patch which adds the aforementioned change to NVPTX isel.
Reviewers: tra
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D26088
llvm-svn: 285460
Summary:
In BypassSlowDivision's short-dividend path, we would create e.g.
udiv exact i32 %a, %b
"exact" here means that we are asserting that %a is a multiple of %b.
But we have no reason to believe this must be true -- this is just a
bug, as far as I can tell.
Reviewers: tra
Subscribers: jholewinski, llvm-commits
Differential Revision: https://reviews.llvm.org/D26097
llvm-svn: 285459
Summary:
The original implementation is in r261607, which was reverted in r269726 to accomendate the ProfileSummaryInfo analysis pass. The new implementation:
1. add a new metadata for function section prefix
2. query against ProfileSummaryInfo in CGP to set the correct section prefix for each function
3. output the section prefix set by CGP
Reviewers: davidxl, eraman
Subscribers: vsk, llvm-commits
Differential Revision: https://reviews.llvm.org/D24989
llvm-svn: 284533
The sink cast machinery is supposed to sink casts as close to their user
as possible. However, an EH pad is the first instruction in it's basic
block. Don't sink if the user is an EH pad.
This fixes PR27536.
llvm-svn: 267767
Sanjay Patel