Regularize the interface for conversion functions to/from float/double.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34881 91177308-0d34-0410-b5e6-96231b3b80d8

Fix constant folding of fp->int bitcast for vectors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34880 91177308-0d34-0410-b5e6-96231b3b80d8

Add APIntVal as a possible GenericeValue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34879 91177308-0d34-0410-b5e6-96231b3b80d8

Only propagate IsKill if the last use is a kill.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34878 91177308-0d34-0410-b5e6-96231b3b80d8

Implement APInt support for the binary operators.
Move the getConstantExpr function towards the end of the file so we don't
need a dozen forward declarations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34877 91177308-0d34-0410-b5e6-96231b3b80d8

1. Have the ExecutionContext keep track of the APInt's allocated and
   ensure they are cleaned up when the stack frame exits.
2. Move a function to the Execution.cpp file where it belongs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34876 91177308-0d34-0410-b5e6-96231b3b80d8

http://llvm.org/bugs/show_bug.cgi?id=1237

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34875 91177308-0d34-0410-b5e6-96231b3b80d8

Implement loading and storing of APInt values from memory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34874 91177308-0d34-0410-b5e6-96231b3b80d8

Make getNumWords public so that those using getRawData stand a chance of
not reading beyond the end of the buffer returned.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34873 91177308-0d34-0410-b5e6-96231b3b80d8

Add names for some of the operators. This is needed for the macros in
the Interpreter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34872 91177308-0d34-0410-b5e6-96231b3b80d8

Emit low/high immediate loads properly for Linux/PPC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34871 91177308-0d34-0410-b5e6-96231b3b80d8

my recent change caused a failure in a bswap testcase, because it changed
the order that instcombine processed instructions in the testcase.  The end
result is that instcombine finished with:

define i16 @test1(i16 %a) {
        %tmp = zext i16 %a to i32               ; <i32> [#uses=2]
        %tmp21 = lshr i32 %tmp, 8               ; <i32> [#uses=1]
        %tmp5 = shl i32 %tmp, 8         ; <i32> [#uses=1]
        %tmp.upgrd.32 = or i32 %tmp21, %tmp5            ; <i32> [#uses=1]
        %tmp.upgrd.3 = trunc i32 %tmp.upgrd.32 to i16           ; <i16> [#uses=1]
        ret i16 %tmp.upgrd.3
}

which can't get matched as a bswap.

This patch makes instcombine more sophisticated about removing truncating
casts, allowing it to turn this into:

define i16 @test2(i16 %a) {
        %tmp211 = lshr i16 %a, 8
        %tmp52 = shl i16 %a, 8
        %tmp.upgrd.323 = or i16 %tmp211, %tmp52
        ret i16 %tmp.upgrd.323
}

which then matches as bswap.  This fixes bswap.ll and implements
InstCombine/cast2.ll:test[12].  This also implements cast elimination of
add/sub.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34870 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase: instcombine should remove all the casts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34869 91177308-0d34-0410-b5e6-96231b3b80d8

Translate bit operations to English.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34868 91177308-0d34-0410-b5e6-96231b3b80d8

instcombine doesn't do CSE, simplify unrelated detail

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34867 91177308-0d34-0410-b5e6-96231b3b80d8

Watch out for cases like this:

entry (0x8b056f0, LLVM BB @0x8b01b30, ID#0):
Live Ins: %r0 %r1 %r2 %r3
        %reg1032 = tMOVrr %r3<kill>
        %reg1033 = tMOVri8 1
        %reg1034 = tMOVri8 0
        tCMPi8 %reg1029<kill>, 0
        tBcc mbb<entry,0x8b06a10>, 0
    Successors according to CFG: 0x8b06980 0x8b06a10

entry (0x8b06980, LLVM BB @0x8b01b30, ID#12):
    Predecessors according to CFG: 0x8b056f0
        %reg1036 = tMOVrr %reg1034<kill>
    Successors according to CFG: 0x8b06a10

entry (0x8b06a10, LLVM BB @0x8b01b30, ID#13):
    Predecessors according to CFG: 0x8b056f0 0x8b06980
        %reg1024<dead> = tMOVrr %reg1030<kill>
        ...

reg1030 and r1 have already been joined. When reg1024 and reg1030 are joined,
r1 live range from function entry to the tMOVrr instruction are dead. Eliminate
r1 from the livein set of the entry BB, not the BB where the copy is.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34866 91177308-0d34-0410-b5e6-96231b3b80d8

add a top-level iteration loop to instcombine.  This means that it will never
finish without combining something it is capable of.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34865 91177308-0d34-0410-b5e6-96231b3b80d8

APIntify this pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34863 91177308-0d34-0410-b5e6-96231b3b80d8

Add isPositive for symmetry with isNegative.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34862 91177308-0d34-0410-b5e6-96231b3b80d8

Finally get this patch right :)
Replace expensive getZExtValue() == 0 calls with isZero() calls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34861 91177308-0d34-0410-b5e6-96231b3b80d8

Dang, I've done that twice now! Undo previous commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34860 91177308-0d34-0410-b5e6-96231b3b80d8

Use more efficient test for one value in a ConstantInt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34859 91177308-0d34-0410-b5e6-96231b3b80d8

Guard against huge loop trip counts in an APInt safe way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34858 91177308-0d34-0410-b5e6-96231b3b80d8

X86-64 VACOPY needs custom expansion. va_list is a struct { i32, i32, i8*, i8* }.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34857 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure debug code is not evaluated in non-debug case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34856 91177308-0d34-0410-b5e6-96231b3b80d8

1. Sort switch cases using APInt safe comparison.
2. Make sure debug output of APInt values is safe for all bit widths.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34855 91177308-0d34-0410-b5e6-96231b3b80d8

Use APInt safe isOne() method on ConstantInt instead of getZExtValue()==1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34854 91177308-0d34-0410-b5e6-96231b3b80d8

Make sorting of ConstantInt be APInt clean through use of ult function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34853 91177308-0d34-0410-b5e6-96231b3b80d8

Fix ashr for bitwidths > 64. This is now validated up to 1024 bits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34852 91177308-0d34-0410-b5e6-96231b3b80d8

Fix uninitialized use of variable. Remove tabs and fix identation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34850 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify things

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34849 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a significant algorithm problem with the instcombine worklist.  removing
a value from the worklist required scanning the entire worklist to remove all
entries.  We now use a combination map+vector to prevent duplicates from
happening and prevent the scan.  This speeds up instcombine on a large file
from the llvm-gcc bootstrap from 189.7s to 4.84s in a debug build and from
5.04s to 1.37s in a release build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34848 91177308-0d34-0410-b5e6-96231b3b80d8

minor cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34846 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR1234 by working around a compiler bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34845 91177308-0d34-0410-b5e6-96231b3b80d8

Mark dead def as unused.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34844 91177308-0d34-0410-b5e6-96231b3b80d8

Dead live-in detection bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34843 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34842 91177308-0d34-0410-b5e6-96231b3b80d8

- Keep track all def and uses of stack slot available in register.
- Available value use may be deleted (e.g. noop move).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34841 91177308-0d34-0410-b5e6-96231b3b80d8

testcase for PR1233

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34840 91177308-0d34-0410-b5e6-96231b3b80d8

Invalidate last use of a reused register if the use is a deleted noop copy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34839 91177308-0d34-0410-b5e6-96231b3b80d8

argument lowering should copy from the vreg shadows of live-in arguments
passed in registers, not directly from the pregs themselves.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34838 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34837 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34836 91177308-0d34-0410-b5e6-96231b3b80d8

Use a better algorithm for rounding sqrt results. Change the FIXME about
this to a NOTE: because pari/gp results start to get rounded incorrectly
after 192 bits of precision. APInt and pari/gp never differ by more than
1, but APInt is more accurate because it does not lose precision after 192
bits as does pari/gp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34834 91177308-0d34-0410-b5e6-96231b3b80d8

Implement unionWith.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34833 91177308-0d34-0410-b5e6-96231b3b80d8

switch the inliner from being recursive to being iterative.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34832 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an unequal bitwidth issue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34831 91177308-0d34-0410-b5e6-96231b3b80d8

grammaro

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34830 91177308-0d34-0410-b5e6-96231b3b80d8

Make it 64-bit safe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34829 91177308-0d34-0410-b5e6-96231b3b80d8

Add a FIXME

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34828 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34826 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a problem where shifting by 64-bits leads to incorrect results on PPC
but not on X86 becuase shift by word size is "undefined".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34825 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate unnecessary reset of SP in epilog on darwin

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34824 91177308-0d34-0410-b5e6-96231b3b80d8

Reverse a premature commital.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34822 91177308-0d34-0410-b5e6-96231b3b80d8

Prefer non-virtual calls to ConstantInt::isZero over virtual calls to
Constant::isNullValue() in situations where it is possible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34821 91177308-0d34-0410-b5e6-96231b3b80d8

Add a non-virtual test for zero, isZero, and rename isUnitValue isOne.
These will be used in upcoming patches to avoid virtual function call when
the client knows it is a ConstantInt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34820 91177308-0d34-0410-b5e6-96231b3b80d8

Add an abs() function to get the absolute value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34819 91177308-0d34-0410-b5e6-96231b3b80d8

Combine two lines that can be.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34818 91177308-0d34-0410-b5e6-96231b3b80d8

Remove virtual keyword from method that doesn't need to be virtual.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34817 91177308-0d34-0410-b5e6-96231b3b80d8

Make it possible to create an SCEVUnknown from an APInt as well as an int.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34816 91177308-0d34-0410-b5e6-96231b3b80d8

Although probably not necessary, guard against a potential assertion by
using isNullValue() instead of getZExtValue() == 0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34815 91177308-0d34-0410-b5e6-96231b3b80d8

Use isUnitValue() instead of getZExtValue() == 1 which will prevent an
assert if the ConstantInt's value is large.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34814 91177308-0d34-0410-b5e6-96231b3b80d8

Add an isUnitValue method for comparison against 1. This just follows a
common pattern with isNullValue, isAllOnesValue, etc. It also helps to
prevent things like getZExtValue() == 1, which could assert for large
bitwidths.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34813 91177308-0d34-0410-b5e6-96231b3b80d8

Use GCC intrinsics when available, and use smarter fallbacks when not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34812 91177308-0d34-0410-b5e6-96231b3b80d8

Clarify the use of getValue/getSExtValue/getZExtValue and add the new
APInt constructor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34811 91177308-0d34-0410-b5e6-96231b3b80d8

Use APInt conversion to string so the result is correct regardless of the
bit width of the ConstantInt being converted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34810 91177308-0d34-0410-b5e6-96231b3b80d8

Make output for ConstantInt construction correct for any bitwidth.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34809 91177308-0d34-0410-b5e6-96231b3b80d8

Use the APInt versions of the bit-wise conversions of float/double to int
and back. While this is not strictly necessary, it does pave the way for
future changes in casting. It should now be possible to improve vector
casting to deal with un-equal fp/int sizes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34808 91177308-0d34-0410-b5e6-96231b3b80d8

Add doubleToBits and floatToBits methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34807 91177308-0d34-0410-b5e6-96231b3b80d8

Use modern variable name. ConstantUnsignedInt is long since dead. No
functional change with this patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34806 91177308-0d34-0410-b5e6-96231b3b80d8

Emit eh filter info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34805 91177308-0d34-0410-b5e6-96231b3b80d8

Collect eh filter info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34804 91177308-0d34-0410-b5e6-96231b3b80d8

Use a simpler constructor for ConstantInt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34803 91177308-0d34-0410-b5e6-96231b3b80d8

Lower eh filter intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34802 91177308-0d34-0410-b5e6-96231b3b80d8

Add eh filter intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34801 91177308-0d34-0410-b5e6-96231b3b80d8

Add bitsToDouble and bitsToFloat methods for re-interpretation of bits as FP.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34800 91177308-0d34-0410-b5e6-96231b3b80d8

Wrap a long line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34799 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid a potential assert out if the loop increment is > 64 bits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34798 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34797 91177308-0d34-0410-b5e6-96231b3b80d8

The 64-bit constructor for ConstantInt changes from int64_t to uint64_t.
This caused a warning for construction with -1. Avoid the warning by using
-1ULL instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34796 91177308-0d34-0410-b5e6-96231b3b80d8

Construct ConstantInt with simpler constructor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34795 91177308-0d34-0410-b5e6-96231b3b80d8

Use a simpler constructor when constructing ConstantInt. Also, replace
verbose code to sext/trunc or zext/trunc and APInt with new methods on
that class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34794 91177308-0d34-0410-b5e6-96231b3b80d8

Use a simpler constructor when constructing ConstantInst.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34793 91177308-0d34-0410-b5e6-96231b3b80d8

Drop the ConstantInt(const Type&, const APInt&) constructor. It is
redundant and more verbose than the ConstantInt(const APInt&) constructor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34792 91177308-0d34-0410-b5e6-96231b3b80d8

Make the static table of results in sqrt const.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34791 91177308-0d34-0410-b5e6-96231b3b80d8

Fix last night's 445.gobmk breakage which was caused by comparison of
APInt's of unequal bitwidth.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34790 91177308-0d34-0410-b5e6-96231b3b80d8

Add methods for bit width modification: sextOrTrunc, zextOrTrunc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34789 91177308-0d34-0410-b5e6-96231b3b80d8

Ensure that fastcall'ed function is correctly mangled & stack is
properly aligned

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34788 91177308-0d34-0410-b5e6-96231b3b80d8

Implemented the frameaddress intrinsic for PPC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34787 91177308-0d34-0410-b5e6-96231b3b80d8

Delete register scavenger when done with it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34786 91177308-0d34-0410-b5e6-96231b3b80d8

Use a spilled free callee-saved register as scratch register.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34785 91177308-0d34-0410-b5e6-96231b3b80d8

Add a version of FindUnusedReg that restrict search to a specific set of registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34784 91177308-0d34-0410-b5e6-96231b3b80d8

- Track which callee-saved registers are spilled.
- Some code clean up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34783 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the "isSigned" parameters from ConstantRange. It turns out they
are not needed as the results are the same with or without it.

Patch by Nicholas Lewycky.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34782 91177308-0d34-0410-b5e6-96231b3b80d8

Switch from std::vector<bool> to BitVector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34781 91177308-0d34-0410-b5e6-96231b3b80d8

APIntify various computations in ScalarEvolution

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34780 91177308-0d34-0410-b5e6-96231b3b80d8

Use a real table in sqrt to shorten and quicken the code.
Thanks for the idea Chris.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34779 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of verboten <iostream> include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34777 91177308-0d34-0410-b5e6-96231b3b80d8

Use the GCC built-in for PopulationCount when it's available, and use a faster
algorithm when it's not.  This should be particularly noticeable in the
64-bit case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34776 91177308-0d34-0410-b5e6-96231b3b80d8

Add a square root function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34775 91177308-0d34-0410-b5e6-96231b3b80d8