cmpxchgXXb uses RBX as one of its implicit argument. I.e., when
we use that instruction we need to clobber RBX. This is generally
fine, expect when RBX is a reserved register because in that case,
the register allocator will not track its value and will not
save and restore it when interferences occur.
rdar://problem/24851412
llvm-svn: 262759
The x86 ret instruction has a 16 bit immediate indicating how many bytes
to pop off of the stack beyond the return address.
There is a problem when extremely large structs are passed by value: we
might not be able to fit the number of bytes to pop into the return
instruction.
To fix this, expand RET_FLAG a little later and use a special sequence
to clean the stack:
pop %ecx ; return address is now in %ecx
add $n, %esp ; clean the stack
push %ecx ; bring the return address back on the stack
ret ; pop the return address and jmp to it's value
llvm-svn: 262755
The diff is relatively large since I took a chance to rearrange the code I had to touch in a more obvious way, but the key bit is merely using the !range metadata when we can't analyze the instruction further. The previous !range metadata code was essentially just dead since no binary operator or cast will have !range metadata (per Verifier) and it was otherwise dropped on the floor.
llvm-svn: 262751
The divrem combine assumed the type of the div/rem is simple, which isn't
necessarily true. This probably worked fine until r250825, since it only
saw legal types, but now breaks when it runs as a pre-type-legalization
combine.
This fixes PR26835.
Differential Revision: http://reviews.llvm.org/D17878
llvm-svn: 262746
The thinlto_linkonceresolution.ll gold linker test introduced in r262727
included a target triple, but didn't set the emulation mode, which is
necessary since the default linker target may be different.
Patch by H.J. Lu
llvm-svn: 262745
When div+rem calls on the same arguments are found, the ARM back-end merges the
two calls into one __aeabi_divmod call for up to 32-bits values. However,
for 64-bit values, which also have a lib call (__aeabi_ldivmod), it wasn't
merging the calls, and thus calling ldivmod twice and spilling the temporary
results, which generated pretty bad code.
This patch legalises 64-bit lib calls for divmod, so that now all the spilling
and the second call are gone. It also relaxes the DivRem combiner a bit on the
legal type check, since it was already checking for isLegalOrCustom on every
value, so the extra check for isTypeLegal was redundant.
Second attempt, creating TLI.isOperationCustom like isOperationExpand, to make
sure we only emit valid types or the ones that were explicitly marked as custom.
Now, passing check-all and test-suite on x86, ARM and AArch64.
This patch fixes PR17193 (and a long time FIXME in the tests).
llvm-svn: 262738
Summary:
This is necessary for when we run out of VGPRs and can no
longer use v_{read,write}_lane for spilling SGPRs.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D17592
llvm-svn: 262732
The split code gen task ID should not be appended to save-temps output
file when the parallelism factor is 1 (not actually splitting).
llvm-svn: 262731
Summary:
This allows us to use virtual registers when we need extra registers
for inserting spill instructions in SIRegisterInfo:eliminateFrameIndex().
Once all the frame indices have been eliminated, the
PrologEpilogueInserter does an extra pass over the program to replace
all virtual registers with physical ones.
This allows us to make more efficient use of our emergency spill slots,
so we only need to create one.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D17591
llvm-svn: 262728
Summary:
Since IR files are all compiled into separate independent object files
in ThinLTO mode, the prevailing linkonce symbols must be emitted in its
object file even if it is no longer referenced there, e.g. if no
references remain in the module after inlining, since it may be
referenced by another ThinLTO compiled object file. This is done by
changing LDPR_PREVAILING_DEF_IRONLY* symbols to LDPR_PREVAILING_DEF,
which converts the prevailing linkonce to weak. We also don't need the
other prevailing IRONLY handling for internalization, which is not
currently performed for ThinLTO.
Test case included.
Reviewers: davidxl, rafael
Subscribers: rafael, joker.eph, llvm-commits
Differential Revision: http://reviews.llvm.org/D16173
llvm-svn: 262727
Summary:
Launch ThinLTO backends (LTO and codegen pipelines with importing) in
parallel using a ThreadPool, after creating the combined index.
The number of threads is controlled by the existing -jobs gold plugin
option, or the hardware concurrency if not specified.
The old behavior of exiting after creating the combined index can be
invoked via a new thinlto-index-only plugin option.
This commit involves just the ThinLTO-specific pieces of D15390, the NFC
and other restructuring pieces were committed independently:
r262677: Add hardware_concurrency interface to llvm::thread (NFC)
r262719: Change split code gen to use ThreadPool
r262721: Refactor gold-plugin codegen to prepare for ThinLTO threads (NFC)
Reviewers: pcc, joker.eph, rafael
Subscribers: rafael, davidxl, llvm-commits, joker.eph
Differential Revision: http://reviews.llvm.org/D15390
llvm-svn: 262724
This is the NFC part remaining from D15390, which refactors the
current codegen() into a CodeGen class with various modular methods and
other helper functions that will be used by the follow-on ThinLTO piece.
llvm-svn: 262721
These correspond to IMAGE_ATOMIC_* and are going to be used by Mesa for the
GL_ARB_shader_image_load_store extension.
Initial change by Nicolai H.hnle
Differential Revision: http://reviews.llvm.org/D17401
llvm-svn: 262701
We have known UB in some ilists where we static cast half nodes to
(larger) derived types and use the address. See llvm.org/PR26753.
This needs to be fixed, but in the meantime it'd be nice if running
ubsan didn't complain. This adds annotations in the two places where
ubsan complains while running check-all of a sanitized clang build.
llvm-svn: 262683
The vast majority of LiveRanges (ie, 4/5) have exactly 1 segment and 1
value number, and a good chunk of the rest have 2 of each, so
allocating space for 4 is wasteful. This is especially noticeable when
dealing with a very large number of vregs, and I have an internal case
where dropping this to 2 shaves over 5% off of peak memory when
compiling a particularly large function.
llvm-svn: 262681
Gold has a newly added LDPT_GET_SYMBOLS_V3 callback that can
distinguish between a module that is not included in the link, and
one that is included but has its entire interface preempted by others.
Fixes PR26674.
llvm-svn: 262676
The RTDyldMemoryManager::getSymbolAddressInProcess method accepts a
linker-mangled symbol name, but it calls through to dlsym to do the lookup (via
DynamicLibrary::SearchForAddressOfSymbol), and dlsym expects an unmangled
symbol name.
Historically we've attempted to "demangle" by removing leading '_'s on all
platforms, and fallen back to an extra search if that failed. That's broken, as
it can cause symbols to resolve incorrectly on platforms that don't do mangling
if you query '_foo' and the process also happens to contain a 'foo'.
Fix this by demangling conditionally based on the host platform. That's safe
here because this function is specifically for symbols in the host process, so
the usual cross-process JIT looking concerns don't apply.
M unittests/ExecutionEngine/ExecutionEngineTest.cpp
M lib/ExecutionEngine/RuntimeDyld/RTDyldMemoryManager.cpp
llvm-svn: 262657
This experiment was originally about trying to use facts implied dominating conditions to infer more precise known bits. While the compile time was found to be acceptable on several large code bases, we never found sufficiently profitable examples to justify turning on the code by default. Given this, it's time to abandon the experiment.
Several folks have commented that they've found this useful for experimentation, but nothing has come of those experiments. Given how easy the patch is to apply, there's no reason to leave the code in tree.
For anyone interested in further investigation in this area, I recommend finding the summary email I sent on one of the original review threads. In particular, I now believe the use-list based approach is strictly worse than the dom-tree-walking approach.
llvm-svn: 262646
Given that we're not actually reducing the instruction count in the included
regression tests, I think we would call this a canonicalization step.
The motivation comes from the example in PR26702:
https://llvm.org/bugs/show_bug.cgi?id=26702
If we hoist the bitwise logic ahead of the bitcast, the previously unoptimizable
example of:
define <4 x i32> @is_negative(<4 x i32> %x) {
%lobit = ashr <4 x i32> %x, <i32 31, i32 31, i32 31, i32 31>
%not = xor <4 x i32> %lobit, <i32 -1, i32 -1, i32 -1, i32 -1>
%bc = bitcast <4 x i32> %not to <2 x i64>
%notnot = xor <2 x i64> %bc, <i64 -1, i64 -1>
%bc2 = bitcast <2 x i64> %notnot to <4 x i32>
ret <4 x i32> %bc2
}
Simplifies to the expected:
define <4 x i32> @is_negative(<4 x i32> %x) {
%lobit = ashr <4 x i32> %x, <i32 31, i32 31, i32 31, i32 31>
ret <4 x i32> %lobit
}
Differential Revision: http://reviews.llvm.org/D17583
llvm-svn: 262645
After r262438 we can have provably positive NSW SCEV expressions whose
zero extensions cannot be simplified (since r262438 makes SCEV better at
computing constant ranges). This means demoting sexts of positive add
recurrences eagerly can result in an unsimplified zero extension where
we could have had a simplified sign extension. This change fixes the
issue by teaching SCEV to demote sext of a positive SCEV expression to a
zext only if the sext could not be simplified.
llvm-svn: 262638