fix register corruption! (my god.) r15 is a scratch reg, using that as
a frame pointer is a pretty doofus thing to do. use r5 instead, and
mark it callee-saved, coz that's what it is!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25526 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor/genericize this, no functionality change

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25525 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure this tests the right thing, xfail on solaris, which doesn't have floorf

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25524 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bugs lowering stackrestore, fixing 2004-08-12-InlinerAndAllocas.c on
PPC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25522 91177308-0d34-0410-b5e6-96231b3b80d8

Speedup and simplify pass registration by the observation that there is
exactly one PassInfo object per RegisterPass object and that their lifetimes
are the same.  As such, there is no reason for the RegisterPass object to
dynamically allocate the PassInfo object at compiler startup time: just inline
the object by-value.  This should reduce codesize, heap size, and startup time. Yaay.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25521 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead #include

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25520 91177308-0d34-0410-b5e6-96231b3b80d8

It doesn't make sense to give llc a list of passes on the command line,
LLVM doesn't use it and it can't work anyway.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25519 91177308-0d34-0410-b5e6-96231b3b80d8

this doesn't work, remove it

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25518 91177308-0d34-0410-b5e6-96231b3b80d8

This only needs <iosfwd> not <iostream>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25517 91177308-0d34-0410-b5e6-96231b3b80d8

Add #include of <iostream>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25516 91177308-0d34-0410-b5e6-96231b3b80d8

Add explicit #includes of <iostream>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25515 91177308-0d34-0410-b5e6-96231b3b80d8

Make iostream #inclusion explicit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25514 91177308-0d34-0410-b5e6-96231b3b80d8

Add explicit iostream #includes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25513 91177308-0d34-0410-b5e6-96231b3b80d8

Make this more efficient in the following ways:
1. Do not statically construct a map when the program starts up, this
   is expensive and cannot be optimized.  Instead, create a list.
2. Do not insert entries for all function in the module into a hashmap
   that lives the full life of the compiler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25512 91177308-0d34-0410-b5e6-96231b3b80d8

This header should not pull in <iostream>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25511 91177308-0d34-0410-b5e6-96231b3b80d8

add explicit #includes of iostream

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25510 91177308-0d34-0410-b5e6-96231b3b80d8

Add explicit #includes of <iostream>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25509 91177308-0d34-0410-b5e6-96231b3b80d8

Several non-functionality changing changes:
1. Use the varargs version of getOrInsertFunction to simplify code.
2. remove #include
3. Reduce the number of #ifdef's.
4. remove extraneous vertical whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25508 91177308-0d34-0410-b5e6-96231b3b80d8

Visual Studio still has issues at being left out.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25507 91177308-0d34-0410-b5e6-96231b3b80d8

Visual Studio neurotic about inconsistent class/struct usage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25506 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in a recent refactor that caused a bunch of programs to miscompile
or the compiler to crash.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25503 91177308-0d34-0410-b5e6-96231b3b80d8

insignificant, but next up is proper stack frame layout!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25497 91177308-0d34-0410-b5e6-96231b3b80d8

Fix CodeGen/PowerPC/2006-01-20-ShiftPartsCrash.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25496 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase that crashes llc for ppc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25495 91177308-0d34-0410-b5e6-96231b3b80d8

Rename fcmovae to fcmovnb and fcmova to fcmovnbe (following Intel manual).
Some assemblers can't recognize the aliases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25494 91177308-0d34-0410-b5e6-96231b3b80d8

Do some code refactoring on Jim's scheduler in preparation of the new list
scheduler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25493 91177308-0d34-0410-b5e6-96231b3b80d8

trivial formatting improvement: don't insert extra blank lines between .comm
vars.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25492 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify search for abbreviations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25491 91177308-0d34-0410-b5e6-96231b3b80d8

Correct some simple errors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25490 91177308-0d34-0410-b5e6-96231b3b80d8

Right size integer values before emitting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25489 91177308-0d34-0410-b5e6-96231b3b80d8

Make the C writer work with packed types.  printContainedStructs is
still not quite right and will be fixed later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25488 91177308-0d34-0410-b5e6-96231b3b80d8

Reworked how Dwarf debug info entries and abbreviations are handled.  Added
pubnames and debuy str sections.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25487 91177308-0d34-0410-b5e6-96231b3b80d8

remove RET hack, add proper support for rets (watching out for ret voids)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25486 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify CWriter::printContainedStructs, also allowing it to work with
PackedTypes as a side-effect.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25485 91177308-0d34-0410-b5e6-96231b3b80d8

LowerReturn now doesn't have to handle f32 returns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25484 91177308-0d34-0410-b5e6-96231b3b80d8

remove some unintentionally committed code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25483 91177308-0d34-0410-b5e6-96231b3b80d8

If the target doesn't support f32 natively, insert the FP_EXTEND in target-indep
code, so that the LowerReturn code doesn't have to handle it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25482 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed InitializeMemory to handle ConstantPacked.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25481 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead expression.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25480 91177308-0d34-0410-b5e6-96231b3b80d8

Don't use invalidated use_iterator's.  This fixes a crash compiling povray

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25479 91177308-0d34-0410-b5e6-96231b3b80d8

fix sext breakage: now we correctly deal with functions that return
int vs uint

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25478 91177308-0d34-0410-b5e6-96231b3b80d8

Update Windows version of config.h to note availability of floor/floorf/ceil/ceilf.
Also pulled Windows-specific modifications to config.h.in to a separate file, so
that changes to it trigger regeneration of config.h (which doesn't happen when the
custom build step for it changes).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25477 91177308-0d34-0410-b5e6-96231b3b80d8

fix storing bools! eek!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25476 91177308-0d34-0410-b5e6-96231b3b80d8

Add a test case for PR697, ensuring the linker works with ConstantPacked.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25475 91177308-0d34-0410-b5e6-96231b3b80d8

Stop doing that accidental commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25474 91177308-0d34-0410-b5e6-96231b3b80d8

A few more SH{L|R}D peepholes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25473 91177308-0d34-0410-b5e6-96231b3b80d8

Added a load folding bug test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25472 91177308-0d34-0410-b5e6-96231b3b80d8

Prevent folding of a node with multiple uses if the node already folds a load!
Here is an example where the load ended up being done twice:

%A = global uint 0

uint %test(uint %B, ubyte %C) {
%tmp = load uint *%A;
%X = shl uint %tmp, ubyte %C
%Cv = sub ubyte 32, %C
%Y = shr uint %B, ubyte %Cv
%Z = or uint %Y, %X
store uint %Z, uint* %A
ret uint %Z
}

==>

subl $4, %esp
movl %ebx, (%esp)
movl 8(%esp), %edx
movl A, %eax
movb 12(%esp), %bl
movb %bl, %cl
shldl %cl, %edx, %eax
movb %bl, %cl
shldl %cl, %edx, A
movl (%esp), %ebx
addl $4, %esp
ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25471 91177308-0d34-0410-b5e6-96231b3b80d8

ConstantFoldLoadThroughGEPConstantExpr wasn't handling pointers to
packed types correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25470 91177308-0d34-0410-b5e6-96231b3b80d8

Didn't mean to commit the last one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25469 91177308-0d34-0410-b5e6-96231b3b80d8

Added i16 SH{L|R}D patterns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25468 91177308-0d34-0410-b5e6-96231b3b80d8

add support for ConstantPacked to the linker

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25467 91177308-0d34-0410-b5e6-96231b3b80d8

Don't include generated lexer .cpp files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25466 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fix. Flag operand number may be calculated incorrectly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25465 91177308-0d34-0410-b5e6-96231b3b80d8

typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25464 91177308-0d34-0410-b5e6-96231b3b80d8

nasty nasty patterns

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25463 91177308-0d34-0410-b5e6-96231b3b80d8

fix boolean XOR (which fixes up comparisons..)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25462 91177308-0d34-0410-b5e6-96231b3b80d8

BOOM!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25460 91177308-0d34-0410-b5e6-96231b3b80d8

click click

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25459 91177308-0d34-0410-b5e6-96231b3b80d8

Use pattern information to determine whether the use expects this
instruction to produce a result. e.g MUL8m, the instruction does not
produce a explicit result. However it produces an implicit result in
AL which would be copied to a temp. The root operator of the matching
pattern is a mul so the use would expect it to produce a result.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25458 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid generating a redundant setcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25457 91177308-0d34-0410-b5e6-96231b3b80d8

For PR696:
Don't do floor->floorf conversion if floorf is not available. This checks
the compiler's host, not its target, which is incorrect for cross-compilers
Not sure that's important as we don't build many cross-compilers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25456 91177308-0d34-0410-b5e6-96231b3b80d8

fix calls that return f32

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25455 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure that libm is used during config tests so that ceil, floor, and
friends are actually detected.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25454 91177308-0d34-0410-b5e6-96231b3b80d8

For PR696:
Add checks for ceil, ceilf, floor, and floorf

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25453 91177308-0d34-0410-b5e6-96231b3b80d8

Implement casts.ll:test26: a cast from float -> double -> integer, doesn't
need the float->double part.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25452 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new testcase for an eliminable cast

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25451 91177308-0d34-0410-b5e6-96231b3b80d8

implement support for f32 arguments past the first 6 words

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25450 91177308-0d34-0410-b5e6-96231b3b80d8

This test passes now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25449 91177308-0d34-0410-b5e6-96231b3b80d8

1. Identify bytecode modules that have upgraded intrinsics by setting a
   boolean flag if we read a function prototype that needs upgrading.
2. Don't upgrade the CallInst instruction until after its been inserted
   into the basic block, and only if we know that we have seen an
   upgraded intrinsic function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25448 91177308-0d34-0410-b5e6-96231b3b80d8

Don't forget about casting the result of an upgraded call to an intrinsic
in the case where it needs to cast back to a signed type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25447 91177308-0d34-0410-b5e6-96231b3b80d8

1. Documentation upgrade.
2. Have UpgradeInstrinicCall return an Instruction* instead of a CallInst*
   and return the needed CastInst* if the result of the upgraded call needs
   to be casted back to a signed type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25446 91177308-0d34-0410-b5e6-96231b3b80d8

Add a flag to identify bytecode files that have intrinsic functions that
need to be upgraded.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25445 91177308-0d34-0410-b5e6-96231b3b80d8

adc and sbb need an incoming flag to ensure it reads the carry flag
from add / sub.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25444 91177308-0d34-0410-b5e6-96231b3b80d8

make this work for Reid, who renamed his CVS module dir :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25443 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase the autoupgrader in the bc reader fails to handle

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25442 91177308-0d34-0410-b5e6-96231b3b80d8

Make get_suffix faster by using a switch on getTypeID rather than a series
of comparisons on the various type objects.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25441 91177308-0d34-0410-b5e6-96231b3b80d8

Another typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25440 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25439 91177308-0d34-0410-b5e6-96231b3b80d8

Two peepholes:
(or (x >> c) | (y << (32 - c))) ==> (shrd x, y, c)
(or (x << c) | (y >> (32 - c))) ==> (shld x, y, c)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25438 91177308-0d34-0410-b5e6-96231b3b80d8

Prevent unnecessary CopyToReg when the same HW register appears in two spots
in the pattern.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25437 91177308-0d34-0410-b5e6-96231b3b80d8

Didn't mean to check that in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25436 91177308-0d34-0410-b5e6-96231b3b80d8

A obvious typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25435 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure intrinsic auto-upgrade is invoked correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25434 91177308-0d34-0410-b5e6-96231b3b80d8

Don't accept the ctpop, cttz, or ctlz intrinsics with signed arguments. The
interface requires unsigned arguments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25433 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new interface function to AutoUpgrade for simultaneously upgrading
the Function and the CallInst: UpgradeCallsToIntrinsic(Function*). Also,
re-factor the AutoUpgrade implementation to eliminate some duplication of
code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25432 91177308-0d34-0410-b5e6-96231b3b80d8

Expand test to cover the non-overloaded variants of the arithmetic and
floating point intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25431 91177308-0d34-0410-b5e6-96231b3b80d8

If not internalizing, don't mark llvm.global[cd]tors const, as a fix for a
hypothetical future boog.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25430 91177308-0d34-0410-b5e6-96231b3b80d8

Don't internalize llvm.global[cd]tor unless there are uses of it.  This
unbreaks front-ends that don't use __main (like the new CFE).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25429 91177308-0d34-0410-b5e6-96231b3b80d8

was ignoring the legalized chain in this case, fixed SPASS on alpha

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25428 91177308-0d34-0410-b5e6-96231b3b80d8

a reduced test case, the GlobalAddress is never legalized

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25427 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of code in the DAGCombiner that is duplicated in SelectionDAG.cpp
Now all constant folding in the code generator is in one place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25426 91177308-0d34-0410-b5e6-96231b3b80d8

Temporary work around for a libcall insertion bug: If a target doesn't
support FSIN/FCOS nodes, do not lower sin/cos to them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25425 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure that cloning a module clones its target triple and dependent
library list as well.  This should help bugpoint.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25424 91177308-0d34-0410-b5e6-96231b3b80d8

Don't assert on 'select_cc SETUO'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25423 91177308-0d34-0410-b5e6-96231b3b80d8

fix out of date comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25422 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Regression/CodeGen/PowerPC/2006-01-18-InvalidBranchOpcodeAssert.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25421 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25420 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a backwards conditional that caused an inf loop in some cases.  This
fixes: test/Regression/CodeGen/Generic/2005-01-18-SetUO-InfLoop.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25419 91177308-0d34-0410-b5e6-96231b3b80d8

this testcase sent the dag combiner spinning into an infinite loop

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25418 91177308-0d34-0410-b5e6-96231b3b80d8