Summary: Debug info should *not* affect code generation. This patch properly handles debug info to make sure the generated code are the same with or without debug info.
Reviewers: davidxl, mzolotukhin, jmolloy
Subscribers: aprantl, llvm-commits
Differential Revision: https://reviews.llvm.org/D25286
llvm-svn: 284415
Not all ConstantExprs can be represented by a global variable, for example most
pointer arithmetic other than addition of a constant, so we can't convert these
values from switch statements to lookup tables.
Differential Revision: https://reviews.llvm.org/D25550
llvm-svn: 284379
Summary:
Constant bundle operands may need to retain their constant-ness for
correctness. I'll admit that this is slightly odd, but it looks like
SimplifyCFG already does this for things like @llvm.frameaddress and
@llvm.stackmap, so I suppose adding one more case is not a big deal.
It is possible to add a mechanism to denote bundle operands that need to
remain constants, but that's probably too complicated for the time
being.
Reviewers: jmolloy
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D25502
llvm-svn: 284028
With the ROPI and RWPI relocation models we can't always have pointers
to global data or functions in constant data, so don't try to convert switches
into lookup tables if any value in the lookup table would require a relocation.
We can still safely emit lookup tables of other values, such as simple
constants.
Differential Revision: https://reviews.llvm.org/D24462
llvm-svn: 283530
GetCaseResults assumed that a terminator with one successor was an
unconditional branch. This is not necessarily the case, it could be a
cleanupret.
Strengthen the check by querying whether or not the terminator is
exceptional.
llvm-svn: 283517
We were updating metadata but not IR flags. Because we pick an arbitrary instruction to be the CSE candidate, it comes down to luck (50% or less chance) if this results in broken codegen or not, which is why PR30373 which is actually not the fault of the commit it was bisected down to.
Fixes PR30373.
llvm-svn: 281889
This patch reverses the edge from DIGlobalVariable to GlobalVariable.
This will allow us to more easily preserve debug info metadata when
manipulating global variables.
Fixes PR30362. A program for upgrading test cases is attached to that
bug.
Differential Revision: http://reviews.llvm.org/D20147
llvm-svn: 281284
This should *actually* fix PR30244. This cranks up the workaround for PR30188 so that we never sink loads or stores of allocas.
The idea is that these should be removed by SROA/Mem2Reg, and any movement of them may well confuse SROA or just cause unwanted code churn. It's not ideal that the midend should be crippled like this, but that unwanted churn can really cause significant regressions in important workloads (tsan).
llvm-svn: 281162
Exposed by PR30244, we will split a block currently if we think we can sink at least one instruction. However this isn't right - the reason we split predecessors is so that we can sink instructions that otherwise couldn't be sunk because it isn't safe to do so - stores, for example.
So, change the heuristic to only split if it thinks it can sink at least one non-speculatable instruction.
Should fix PR30244.
llvm-svn: 281160
Summary: The hoisted instruction is executed speculatively. It could affect the debugging experience as user would see gdb go into code that may not be expected to execute. It will also affect sample profile accuracy by assigning incorrect frequency to source within then/else branch.
Reviewers: davidxl, dblaikie, chandlerc, kcc, echristo
Subscribers: mehdi_amini, probinson, eric_niebler, andreadb, llvm-commits
Differential Revision: https://reviews.llvm.org/D24164
llvm-svn: 280995
We can't create metadata-valued PHIs; don't try to do so when sinking.
I created a test case for this using the @llvm.type.test intrinsic, because it
takes a metadata parameter and does not have severe side effects (thus
SimplifyCFG is willing to otherwise sink it).
Previously, running the test case would crash with:
Invalid use of metadata!
%.sink = select i1 %flag, metadata <...>, metadata <0x4e45dc0>
LLVM ERROR: Broken function found, compilation aborted!
llvm-svn: 280866
I should have realised this the first time around, but if we're avoiding sinking stores where the operands come from allocas so they don't create selects, we also have to do the same for loads because SROA will be just as defective looking at loads of selected addresses as stores.
Fixes PR30188 (again).
llvm-svn: 280792
PR30292 showed a case where our PHI checking wasn't correct. We were checking that all values were used by the same PHI before deciding to sink, but we weren't checking that the incoming values for that PHI were what we expected. As a result, we had to bail out after block splitting which caused us to never reach a steady state in SimplifyCFG.
Fixes PR30292.
llvm-svn: 280790
This test code previously caused a failure in the module verifier,
because SimplifyCFG created this invalid instruction, which tries to
take the address of inline asm:
%.sink = select i1 %1, i64 ()* asm "mov $0, #1", "=r", i64 ()* asm %"mov $0, #2", "=r"
This has been fixed recently, presumably by James Molloy's patches that
re-wrote and changed parts of SimplifyCFG, so this patch just adds a
regression test for it.
Differential Revision: https://reviews.llvm.org/D24231
llvm-svn: 280660
We're sinking stores, which is a good thing, but in the process creating selects for the store address operand, which SROA/Mem2Reg can't look through, which caused serious regressions.
The real fix is in SROA, which I'll be looking into.
llvm-svn: 280470
This was a real restriction in the original version of SinkIfThenCodeToEnd. Now it's been rewritten, the restriction can be lifted.
As part of this, we handle a very common and useful case where one of the incoming branches is actually conditional. Consider:
if (a)
x(1);
else if (b)
x(2);
This produces the following CFG:
[if]
/ \
[x(1)] [if]
| | \
| | \
| [x(2)] |
\ | /
[ end ]
[end] has two unconditional predecessor arcs and one conditional. The conditional refers to the implicit empty 'else' arc. This same pattern can also be caused by an empty default block in a switch.
We can't sink the call to x() down to end because no call to x() happens on the third incoming arc (assume that x() has sideeffects for the sake of argument; if something is safe to speculate we could indeed sink nevertheless but this cannot happen in the general case and causes many extra selects).
We are now able to detect this case and split off the unconditional arcs to a common successor:
[if]
/ \
[x(1)] [if]
| | \
| | \
| [x(2)] |
\ / |
[sink.split] |
\ /
[ end ]
Now we can sink the call to x() into %sink.split. This can cause significant code simplification in many testcases.
llvm-svn: 280364
r279460 rewrote this function to be able to handle more than two incoming edges and took pains to ensure this didn't regress anything.
This time we change the logic for determining if an instruction should be sunk. Previously we used a single pass greedy algorithm - sink instructions until one requires more than one PHI node or we run out of instructions to sink.
This had the problem that sinking instructions that had non-identical but trivially the same operands needed extra logic so we sunk them aggressively. For example:
%a = load i32* %b %d = load i32* %b
%c = gep i32* %a, i32 0 %e = gep i32* %d, i32 1
Sinking %c and %e would naively require two PHI merges as %a != %d. But the loads are obviously equivalent (and maybe can't be hoisted because there is no common predecessor).
This is why we implemented the fairly complex function areValuesTriviallySame(), to look through trivial differences like this. However it's just not clever enough.
Instead, throw areValuesTriviallySame away, use pointer equality to check equivalence of operands and switch to a two-stage algorithm.
In the "scan" stage, we look at every sinkable instruction in isolation from end of block to front. If it's sinkable, we keep track of all operands that required PHI merging.
In the "sink" stage, we iteratively sink the last non-terminator in the source blocks. But when calculating how many PHIs are actually required to be inserted (to work out if we should stop or not) we remove any values that have already been sunk from the set of PHI-merges required, which allows us to be more aggressive.
This turns an algorithm with potentially recursive lookahead (looking through GEPs, casts, loads and any other instruction potentially not CSE'd) to two linear scans.
llvm-svn: 280351
We check that a sinking candidate is used by only one PHI node during our legality checks. However for instructions that are used by other sinking candidates our heuristic is less conservative. This can result in a candidate actually being illegal when we come to sink it because of how we sunk a predecessor. Do the used-by-only-one-PHI checks again during sinking to ensure we don't crash.
llvm-svn: 280228
We're sinking stores, which is a good thing, but in the process creating selects for the store address operand, which SROA/Mem2Reg can't look through, which caused serious regressions.
The real fix is in SROA, which I'll be looking into.
llvm-svn: 280219
This was a real restriction in the original version of SinkIfThenCodeToEnd. Now it's been rewritten, the restriction can be lifted.
As part of this, we handle a very common and useful case where one of the incoming branches is actually conditional. Consider:
if (a)
x(1);
else if (b)
x(2);
This produces the following CFG:
[if]
/ \
[x(1)] [if]
| | \
| | \
| [x(2)] |
\ | /
[ end ]
[end] has two unconditional predecessor arcs and one conditional. The conditional refers to the implicit empty 'else' arc. This same pattern can also be caused by an empty default block in a switch.
We can't sink the call to x() down to end because no call to x() happens on the third incoming arc (assume that x() has sideeffects for the sake of argument; if something is safe to speculate we could indeed sink nevertheless but this cannot happen in the general case and causes many extra selects).
We are now able to detect this case and split off the unconditional arcs to a common successor:
[if]
/ \
[x(1)] [if]
| | \
| | \
| [x(2)] |
\ / |
[sink.split] |
\ /
[ end ]
Now we can sink the call to x() into %sink.split. This can cause significant code simplification in many testcases.
llvm-svn: 280217
r279460 rewrote this function to be able to handle more than two incoming edges and took pains to ensure this didn't regress anything.
This time we change the logic for determining if an instruction should be sunk. Previously we used a single pass greedy algorithm - sink instructions until one requires more than one PHI node or we run out of instructions to sink.
This had the problem that sinking instructions that had non-identical but trivially the same operands needed extra logic so we sunk them aggressively. For example:
%a = load i32* %b %d = load i32* %b
%c = gep i32* %a, i32 0 %e = gep i32* %d, i32 1
Sinking %c and %e would naively require two PHI merges as %a != %d. But the loads are obviously equivalent (and maybe can't be hoisted because there is no common predecessor).
This is why we implemented the fairly complex function areValuesTriviallySame(), to look through trivial differences like this. However it's just not clever enough.
Instead, throw areValuesTriviallySame away, use pointer equality to check equivalence of operands and switch to a two-stage algorithm.
In the "scan" stage, we look at every sinkable instruction in isolation from end of block to front. If it's sinkable, we keep track of all operands that required PHI merging.
In the "sink" stage, we iteratively sink the last non-terminator in the source blocks. But when calculating how many PHIs are actually required to be inserted (to work out if we should stop or not) we remove any values that have already been sunk from the set of PHI-merges required, which allows us to be more aggressive.
This turns an algorithm with potentially recursive lookahead (looking through GEPs, casts, loads and any other instruction potentially not CSE'd) to two linear scans.
llvm-svn: 280216
This was deliberately disabled during my rewrite of SinkIfThenToEnd to keep behaviour
at least vaguely consistent with the previous version and keep it as close to NFC as
I could.
There's no real reason not to merge sideeffect calls though, so let's do it! Small fixup
along the way to ensure we don't create indirect calls.
Should fix PR28964.
llvm-svn: 280215
Assuming the default FP env, we should not treat fdiv and frem any differently in terms of
trapping behavior than any other FP op. Ie, FP ops do not trap with the default FP env.
This matches how we treat the fdiv/frem in IR with isSafeToSpeculativelyExecute() and in
the backend after:
https://reviews.llvm.org/rL279970
llvm-svn: 279973
[Recommitting now an unrelated assertion in SROA is sorted out]
The new version has several advantages:
1) IMSHO it's more readable and neater
2) It handles loads and stores properly
3) It can handle any number of incoming blocks rather than just two. I'll be taking advantage of this in a followup patch.
With this change we can now finally sink load-modify-store idioms such as:
if (a)
return *b += 3;
else
return *b += 4;
=>
%z = load i32, i32* %y
%.sink = select i1 %a, i32 5, i32 7
%b = add i32 %z, %.sink
store i32 %b, i32* %y
ret i32 %b
When this works for switches it'll be even more powerful.
Round 4. This time we should handle all instructions correctly, and not replace any operands that need to be constant with variables.
This was really hard to determine safely, so the helper function should be put into the Instruction API. I'll do that as a followup.
llvm-svn: 279460
The new version has several advantages:
1) IMSHO it's more readable and neater
2) It handles loads and stores properly
3) It can handle any number of incoming blocks rather than just two. I'll be taking advantage of this in a followup patch.
With this change we can now finally sink load-modify-store idioms such as:
if (a)
return *b += 3;
else
return *b += 4;
=>
%z = load i32, i32* %y
%.sink = select i1 %a, i32 5, i32 7
%b = add i32 %z, %.sink
store i32 %b, i32* %y
ret i32 %b
When this works for switches it'll be even more powerful.
Round 4. This time we should handle all instructions correctly, and not replace any operands that need to be constant with variables.
This was really hard to determine safely, so the helper function should be put into the Instruction API. I'll do that as a followup.
llvm-svn: 279443
The new version has several advantages:
1) IMSHO it's more readable and neater
2) It handles loads and stores properly
3) It can handle any number of incoming blocks rather than just two. I'll be taking advantage of this in a followup patch.
With this change we can now finally sink load-modify-store idioms such as:
if (a)
return *b += 3;
else
return *b += 4;
=>
%z = load i32, i32* %y
%.sink = select i1 %a, i32 5, i32 7
%b = add i32 %z, %.sink
store i32 %b, i32* %y
ret i32 %b
When this works for switches it'll be even more powerful.
llvm-svn: 279229
This reverts commit r278660.
It causes downstream assertion failure in InstCombine on shuffle
instructions. Comes up in __mm_swizzle_epi32.
llvm-svn: 278672
The new version has several advantages:
1) IMSHO it's more readable and neater
2) It handles loads and stores properly
3) It can handle any number of incoming blocks rather than just two. I'll be taking advantage of this in a followup patch.
With this change we can now finally sink load-modify-store idioms such as:
if (a)
return *b += 3;
else
return *b += 4;
=>
%z = load i32, i32* %y
%.sink = select i1 %a, i32 5, i32 7
%b = add i32 %z, %.sink
store i32 %b, i32* %y
ret i32 %b
When this works for switches it'll be even more powerful.
llvm-svn: 278660
This generated IR based on the order of evaluation, which is different
between GCC and Clang. With that in mind you get bootstrap miscompares
if you compare a Clang built with GCC-built Clang vs. Clang built with
Clang-built Clang. Diagnosing that made my head hurt.
This also reverts commit r277337, which "fixed" the test case.
llvm-svn: 277820
Using RAUW was wrong here; if we have a switch transform such as:
18 -> 6 then
6 -> 0
If we use RAUW, while performing the second transform the *transformed* 6
from the first will be also replaced, so we end up with:
18 -> 0
6 -> 0
Found by clang stage2 bootstrap; testcase added.
llvm-svn: 277332
It looks like the two independent parts of the rotate operation (a lshr and shl) are being reordered on some bots. Add CHECK-DAGs to account for this.
llvm-svn: 277329
If a switch is sparse and all the cases (once sorted) are in arithmetic progression, we can extract the common factor out of the switch and create a dense switch. For example:
switch (i) {
case 5: ...
case 9: ...
case 13: ...
case 17: ...
}
can become:
if ( (i - 5) % 4 ) goto default;
switch ((i - 5) / 4) {
case 0: ...
case 1: ...
case 2: ...
case 3: ...
}
or even better:
switch ( ROTR(i - 5, 2) {
case 0: ...
case 1: ...
case 2: ...
case 3: ...
}
The division and remainder operations could be costly so we only do this if the factor is a power of two, and emit a right-rotate instead of a divide/remainder sequence. Dense switches can be lowered significantly better than sparse switches and can even be transformed into lookup tables.
llvm-svn: 277325
SimplifyCFG had logic to insert calls to llvm.trap for two very
particular IR patterns: stores and invokes of undef/null.
While InstCombine canonicalizes certain undefined behavior IR patterns
to stores of undef, phase ordering means that this cannot be relied upon
in general.
There are much better tools than llvm.trap: UBSan and ASan.
N.B. I could be argued into reverting this change if a clear argument as
to why it is important that we synthesize llvm.trap for stores, I'd be
hard pressed to see why it'd be useful for invokes...
llvm-svn: 273778
reduce the number of comparisons.
Specifically, InstCombine can turn:
(i == 5334 || i == 5335)
into:
((i | 1) == 5335)
SimplifyCFG was already able to detect the pattern:
(i == 5334 || i == 5335)
to:
((i & -2) == 5334)
This patch supersedes D21315 and resolves PR27555
(https://llvm.org/bugs/show_bug.cgi?id=27555).
Thanks to David and Chandler for the suggestions!
Author: Thomas Jablin (tjablin)
Reviewers: majnemer chandlerc halfdan cycheng
http://reviews.llvm.org/D21397
llvm-svn: 273639
(i == 5334 || i == 5335)
to:
((i & -2) == 5334)
This transformation has some incorrect side conditions. Specifically, the
transformation is only applied when the right-hand side constant (5334 in
the example) is a power of two not equal and not equal to the negated mask.
These side conditions were added in r258904 to fix PR26323. The correct side
condition is that: ((Constant & Mask) == Constant)[(5334 & -2) == 5334].
It's a little bit hard to see why these transformations are correct and what
the side conditions ought to be. Here is a CVC3 program to verify them for
64-bit values:
ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
x : BITVECTOR(64);
y : BITVECTOR(64);
z : BITVECTOR(64);
mask : BITVECTOR(64) = BVSHL(ONE, z);
QUERY( (y & ~mask = y) =>
((x & ~mask = y) <=> (x = y OR x = (y | mask)))
);
Please note that each pattern must be a dual implication (<--> or iff). One
directional implication can create spurious matches. If the implication is
only one-way, an unsatisfiable condition on the left side can imply a
satisfiable condition on the right side. Dual implication ensures that
satisfiable conditions are transformed to other satisfiable conditions and
unsatisfiable conditions are transformed to other unsatisfiable conditions.
Here is a concrete example of a unsatisfiable condition on the left
implying a satisfiable condition on the right:
mask = (1 << z)
(x & ~mask) == y --> (x == y || x == (y | mask))
Substituting y = 3, z = 0 yields:
(x & -2) == 3 --> (x == 3 || x == 2)
The version of this code before r258904 had no side-conditions and
incorrectly justified itself in comments through one-directional
implication.
Thanks to Chandler for the suggestion!
Author: Thomas Jablin (tjablin)
Reviewers: chandlerc majnemer hfinkel cycheng
http://reviews.llvm.org/D21417
llvm-svn: 272873
Mandeep Singh Grang