fix x86-64 static codegen to materialize the address of a global with movl instead
of lea.  It is better for code size (and presumably efficiency) to use:

  movl $foo, %eax

rather than:

  leal foo, eax

Both give a nice zero extending "move immediate" instruction, the former is just
smaller.  Note that global addresses should be handled different by the x86
backend, but I chose to follow the style already in place and add more fixme's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75403 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bug in my cleanup patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75402 91177308-0d34-0410-b5e6-96231b3b80d8

comment cleanup, reduce nesting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75398 91177308-0d34-0410-b5e6-96231b3b80d8

remove some dead patterns, WrapperRIP doesn't exist in -static mode
anymore, so these aren't needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75397 91177308-0d34-0410-b5e6-96231b3b80d8

silence vc++ warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75394 91177308-0d34-0410-b5e6-96231b3b80d8

silence a vc++ warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75393 91177308-0d34-0410-b5e6-96231b3b80d8

this test was incorrect for x86-64 static.  It passed on darwin, because darwin
doesn't have static x86-64 mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75392 91177308-0d34-0410-b5e6-96231b3b80d8

Tweak comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75391 91177308-0d34-0410-b5e6-96231b3b80d8

Update CMakeLists.txt

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75389 91177308-0d34-0410-b5e6-96231b3b80d8

Tweak svn:ignore

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75387 91177308-0d34-0410-b5e6-96231b3b80d8

Fix unused function warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75386 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r75252 which was causing some crashes at compile time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75384 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR4533, which is about buggy codegen in x86-64 -static mode.
Basically, using:
  lea symbol(%rip), %rax

is not valid in -static mode, because the current RIP may not be
within 32-bits of "symbol" when an app is built partially pic and
partially static.  The fix for this is to compile it to:

  lea symbol, %rax

It would be better to codegen this as:

  movq $symbol, %rax

but that will come next.

The hard part of fixing this bug was fixing abi-isel, which was actively
testing for the wrong behavior.  Also, the RUN lines are completely impossible
to understand what they are testing.  To help with this, convert the -static
x86-64 codegen tests to use filecheck.  This is much more stable and makes it
more clear what the codegen is expected to be.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75382 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75380 91177308-0d34-0410-b5e6-96231b3b80d8

assert(0) -> LLVM_UNREACHABLE.
Make llvm_unreachable take an optional string, thus moving the cerr<< out of
line.
LLVM_UNREACHABLE is now a simple wrapper that makes the message go away for
NDEBUG builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75379 91177308-0d34-0410-b5e6-96231b3b80d8

Stub out assembly matcher (.s -> MCInst) tblgen backend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75378 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75376 91177308-0d34-0410-b5e6-96231b3b80d8

Update documentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75375 91177308-0d34-0410-b5e6-96231b3b80d8

Delete the temp dir even when '--temp-dir' is specified.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75374 91177308-0d34-0410-b5e6-96231b3b80d8

Fix handling of max and full set.

A full set is a constant range that represents any number. If you take the
umax of that and [5, 10) you end up with [5, INT_MAX] because the values less
than 5 would be umax's against a value which is at least 5.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75372 91177308-0d34-0410-b5e6-96231b3b80d8

improve filecheck's "scanning from here" caret position.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75371 91177308-0d34-0410-b5e6-96231b3b80d8

make filecheck default to canonicalizing horizontal whitespace
away.  This way you can write a space and it matches arbitrary spaces and tabs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75370 91177308-0d34-0410-b5e6-96231b3b80d8

Break the world's largest unit test down a few logical lines. No semantic
changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75369 91177308-0d34-0410-b5e6-96231b3b80d8

Erase the temp dir before starting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75367 91177308-0d34-0410-b5e6-96231b3b80d8

Clarify and simplify.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75366 91177308-0d34-0410-b5e6-96231b3b80d8

Don't use a void return type with a function that returns a value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75364 91177308-0d34-0410-b5e6-96231b3b80d8

Convert more assert(0)+abort() -> LLVM_UNREACHABLE,
and abort()/exit() -> llvm_report_error().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75363 91177308-0d34-0410-b5e6-96231b3b80d8

We get the P modifier wrong in a lot of cases, just add some more rigorous testing.
In addition to fixing this, I still need to do some more testing on darwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75362 91177308-0d34-0410-b5e6-96231b3b80d8

Don't put IT instruction before conditional branches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75361 91177308-0d34-0410-b5e6-96231b3b80d8

Smarter isel of ldrsb / ldrsh. Only make use of these when [r,r] address is feasible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75360 91177308-0d34-0410-b5e6-96231b3b80d8

Major changes to Thumb (not Thumb2). Many 16-bit instructions either modifies CPSR  when they are outside the IT blocks, or they can predicated when in Thumb2. Move the implicit def of CPSR to an optional def which defaults CPSR. This allows the 's' bit to be toggled dynamically.

A side-effect of this change is asm printer is now using unified assembly. There are some minor clean ups and fixes as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75359 91177308-0d34-0410-b5e6-96231b3b80d8

80 col violation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75358 91177308-0d34-0410-b5e6-96231b3b80d8

Move a method that creates constant ranges relative to another constant range
per icmp predicate out of predsimplify and into ConstantRange.

Add another utility method that determines whether one range is a subset of
another. Combine with the former to determine whether icmp pred range, range
is known to be true or not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75357 91177308-0d34-0410-b5e6-96231b3b80d8

Fix up support for OptionalDefOperand when it defaults to an actual register def. I need this to get ready for major Thumb1 surgery.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75328 91177308-0d34-0410-b5e6-96231b3b80d8

Use CreateStackStoreLoad helper in more places.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75320 91177308-0d34-0410-b5e6-96231b3b80d8

Two-address pass should use findCommutedOpIndices to determine what registers are commuted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75317 91177308-0d34-0410-b5e6-96231b3b80d8

Use findCommutedOpIndices to find the operands to commute.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75312 91177308-0d34-0410-b5e6-96231b3b80d8

Add superclasses of ARM Neon quad registers.  The Q2PR class contains pairs of
quad registers and the Q4PR class holds sets of 4 quad registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75309 91177308-0d34-0410-b5e6-96231b3b80d8

Add new vector types for 192-bit, 348-bit and 512-bit sizes.
These are needed to represent ARM Neon struct datatypes containing 2, 3 or 4
separate vectors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75308 91177308-0d34-0410-b5e6-96231b3b80d8

remove the "debug" modifier, it is only used by one instruction which can
never be generated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75305 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor TableGen's llvm::getName to share code with llvm::getEnumName,
since names are the same for almost all the types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75302 91177308-0d34-0410-b5e6-96231b3b80d8

add support for .zerofill, patch by Kevin Enderby!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75301 91177308-0d34-0410-b5e6-96231b3b80d8

Oops. s/#if defined(USE_OPROFILE)/#if USE_OPROFILE/. We #define
USE_OPROFILE to 0 on some paths through configure, which does the
wrong thing with #if defined().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75300 91177308-0d34-0410-b5e6-96231b3b80d8

make pcrel and non-pcrel global printing more similar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75298 91177308-0d34-0410-b5e6-96231b3b80d8

Support remote execute for ARM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75292 91177308-0d34-0410-b5e6-96231b3b80d8

Undo my brain cramp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75290 91177308-0d34-0410-b5e6-96231b3b80d8

Make changes suggested by Chris and eliminate newly-added raw_ostream
hooks as they're no longer needed.

The major change with this patch is to make formatted_raw_ostream usable
by any client of raw_ostream.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75283 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate configure after r75279.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75280 91177308-0d34-0410-b5e6-96231b3b80d8

Add a --with-oprofile flag to configure, which uses OProfile's agent
library to tell it the addresses of JITted functions.  For a
particular program, this changes the opreport -l output from:

samples %    image name        symbol name
48182  98.9729 anon (tgid:19412 range:0x7f12ccaab000-0x7f12cdaab000) anon (tgid:19412 range:0x7f12ccaab000-0x7f12cdaab000)
11     0.0226  libstdc++.so.6.0.9    /usr/lib/libstdc++.so.6.0.9

to:

samples %    image name        symbol name
24565  60.7308 19814.jo        fib_left
15365  37.9861 19814.jo        fib_right
22     0.0544  ld-2.7.so       do_lookup_x

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75279 91177308-0d34-0410-b5e6-96231b3b80d8

code model is never set to default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75278 91177308-0d34-0410-b5e6-96231b3b80d8

fix indentation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75277 91177308-0d34-0410-b5e6-96231b3b80d8

remove the now-dead TM argument to these methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75276 91177308-0d34-0410-b5e6-96231b3b80d8

make PIC vs DynamicNoPIC be explicit in PICStyles.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75275 91177308-0d34-0410-b5e6-96231b3b80d8

some minor simplifications.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75274 91177308-0d34-0410-b5e6-96231b3b80d8

add a couple of predicates to test for "stub style pic in PIC mode" and "stub style pic in dynamic-no-pic" mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75273 91177308-0d34-0410-b5e6-96231b3b80d8

Remove ScalarEvolution::hasSCEV, which isn't being used, and which
breaks encapsulation. Also remove a dead prototype for setSCEV.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75272 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid compiler warnings when assertions are turned off.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75269 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid compiler warnings if assertions turned off.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75267 91177308-0d34-0410-b5e6-96231b3b80d8

CMOVxx doesn't swap operands which it's commuted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75266 91177308-0d34-0410-b5e6-96231b3b80d8

Remove TargetInstrInfo::CommuteChangesDestination and added findCommutedOpIndices which returns the operand indices which are swapped (when applicable). This allows for some code clean up and future enhancements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75264 91177308-0d34-0410-b5e6-96231b3b80d8

Link std.lib (.bc code) with llvm-ld.
Link devices.lib (processor specific variables) with mplink.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75263 91177308-0d34-0410-b5e6-96231b3b80d8

Fix unit tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75262 91177308-0d34-0410-b5e6-96231b3b80d8

convert test to use FileCheck, which is much more precise and faster than
the previous RUN lines.  Hopefully this will be an inspiration for future
tests :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75261 91177308-0d34-0410-b5e6-96231b3b80d8

add missing *, patch by Peter O'Gorman!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75259 91177308-0d34-0410-b5e6-96231b3b80d8

Push LLVMContext through the TypeBuilder API.  There are no users for this in-tree, so I can't really test it.
If you're using this, and it's broken, please send patches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75257 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate an unnecessary include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75256 91177308-0d34-0410-b5e6-96231b3b80d8

Push LLVMContext through the PatternMatch API.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75255 91177308-0d34-0410-b5e6-96231b3b80d8

Predicate VFP instructions on HasVFP2 instead of IsARM. This allows VFP instructions with thumb-2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75254 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize ScalarEvolution's cast-folding code to support more kinds
of loops. Add several new functions to for working with ScalarEvolution's
add-hoc value-range analysis functionality.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75252 91177308-0d34-0410-b5e6-96231b3b80d8

Update for GlobalVariables ctor change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75251 91177308-0d34-0410-b5e6-96231b3b80d8

t2LDM_RET does not fall-through.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75250 91177308-0d34-0410-b5e6-96231b3b80d8

Add Thumb2ITBlockPass.cpp to CMakeLists.txt, fixing
the cmake build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75246 91177308-0d34-0410-b5e6-96231b3b80d8

make this more like printOperand.  Perhaps some merging will happen
tomorrow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75245 91177308-0d34-0410-b5e6-96231b3b80d8

simplify fast isel by using ClassifyGlobalReference. This
elimiantes the last use of GVRequiresExtraLoad, so delete it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75244 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate GVRequiresRegister, replacing it with predicates we
need for other purposes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75243 91177308-0d34-0410-b5e6-96231b3b80d8

change a bunch of logic in LowerGlobalAddress to leverage the work
done in ClassifyGlobalReference instead of reconstructing the info
awkwardly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75240 91177308-0d34-0410-b5e6-96231b3b80d8

add a predicate to determine if a global var reference requires a
PIC-base to be added in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75238 91177308-0d34-0410-b5e6-96231b3b80d8

move some classification logic around.  Now GVRequiresExtraLoad
is just a trivial wrapper around "ClassifyGlobalReference", which
stole a ton of logic from LowerGlobalAddress.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75237 91177308-0d34-0410-b5e6-96231b3b80d8

change isGlobalStubReference to take target flags instead of a MachineOperand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75236 91177308-0d34-0410-b5e6-96231b3b80d8

convert some late code (called by regalloc and code emission)
to use isGlobalStubReference instead of GVRequiresExtraLoad
(which should really be part of isel).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75234 91177308-0d34-0410-b5e6-96231b3b80d8

add a new predicate method that says whether a GlobalValue
MachineOperand is a reference to a stub, not a reference to the
global variable itself.  Look no context needed!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75233 91177308-0d34-0410-b5e6-96231b3b80d8

GVRequiresExtraLoad is now never used for calls, simplify it based on this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75232 91177308-0d34-0410-b5e6-96231b3b80d8

actually, just eliminate PCRelGVRequiresExtraLoad.  It makes the code
more complex and slow than just directly testing what we care about.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75231 91177308-0d34-0410-b5e6-96231b3b80d8

There is only one case where GVRequiresExtraLoad returns true for calls:
split its handling out to PCRelGVRequiresExtraLoad, and simplify code
based on this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75230 91177308-0d34-0410-b5e6-96231b3b80d8

the "isDirectCall" operand of GVRequiresRegister is always false, eliminate it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75229 91177308-0d34-0410-b5e6-96231b3b80d8

split call handling out of X86SelectAddress into X86SelectCallAddress

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75228 91177308-0d34-0410-b5e6-96231b3b80d8

convert a helper method to be a static function instead of a
template.  Also convert it to take a MachineOperand instead of a GV*

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75227 91177308-0d34-0410-b5e6-96231b3b80d8

80 col violation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75226 91177308-0d34-0410-b5e6-96231b3b80d8

More info about Thumb1 predication support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75220 91177308-0d34-0410-b5e6-96231b3b80d8

We don't need separate thumb1 instructions tADDSi3 etc. for addc and subc. The "normal" version always modify condition register CPSR so we should just use def : pat to match to the same instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75219 91177308-0d34-0410-b5e6-96231b3b80d8

Add a thumb2 pass to insert IT blocks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75218 91177308-0d34-0410-b5e6-96231b3b80d8

Move isPredicated from .cpp to .h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75217 91177308-0d34-0410-b5e6-96231b3b80d8

Finish pushing LLVMContext through the IRBuilder/ConstantFolder interface.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75213 91177308-0d34-0410-b5e6-96231b3b80d8

80 col violation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75212 91177308-0d34-0410-b5e6-96231b3b80d8

Fix #include specification.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75208 91177308-0d34-0410-b5e6-96231b3b80d8

ImmutableMap/ImmutableSet: Allow caching of ImutAVLTree digests while the tree
is still mutable. My experiments show this reduces the amount of times we
compute a full tree digest by over 50% on very small cases, and potentially much
larger on others.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75207 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a bogus assertion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75206 91177308-0d34-0410-b5e6-96231b3b80d8

Replace TM.getRegisterInfo() calls by TRI instance variable.
Use getAsmName() method instead of accessing AsmName field directly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75205 91177308-0d34-0410-b5e6-96231b3b80d8

Redesign this to avoid standard stream classes.  This stream class
provides pretty -printing of comments and other such things in asm
files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75202 91177308-0d34-0410-b5e6-96231b3b80d8

Handle 'a' modifier on inline assembly operands.
This is part of the fix for pr4521.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75201 91177308-0d34-0410-b5e6-96231b3b80d8