Implemented x86 inline asm b, h, w, k modifiers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28020 91177308-0d34-0410-b5e6-96231b3b80d8

Fix InstCombine/2006-04-28-ShiftShiftLongLong.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28019 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase miscompiled by instcombine

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28018 91177308-0d34-0410-b5e6-96231b3b80d8

Fix CodeGen/Generic/2006-04-28-Sign-extend-bool.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28017 91177308-0d34-0410-b5e6-96231b3b80d8

testcase that crashes the ppc backend, which can't sextinreg(i1)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28016 91177308-0d34-0410-b5e6-96231b3b80d8

Initial caller side support (for CCC only, not FastCC) of 128-bit vector
passing by value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28015 91177308-0d34-0410-b5e6-96231b3b80d8

Bare-bone X86 inline asm printer support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28014 91177308-0d34-0410-b5e6-96231b3b80d8

Update. It should use two shufps, not three!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28013 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the temporary option: -no-isel-fold-inflight

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28012 91177308-0d34-0410-b5e6-96231b3b80d8

Implement four-wide shuffle with 2 shufps if no more than two elements come
from each vector. e.g.
        shuffle(G1, G2, 7, 1, 5, 2)
==>
        movaps _G2, %xmm0
        shufps $151, _G1, %xmm0
        shufps $216, %xmm0, %xmm0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28011 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR743: emit -help output of a tool to cout, not cerr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28010 91177308-0d34-0410-b5e6-96231b3b80d8

TargetLowering::LowerArguments should return a VBIT_CONVERT of
FORMAL_ARGUMENTS SDOperand in the return result vector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28009 91177308-0d34-0410-b5e6-96231b3b80d8

Mapping of physregs can make it so that the designated and input physregs are
the same.  In this case, don't emit a noop copy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28008 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Transforms/Reassociate/2006-04-27-ReassociateVector.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28007 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28006 91177308-0d34-0410-b5e6-96231b3b80d8

Use movaps instead of movapd for spill / restore.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28005 91177308-0d34-0410-b5e6-96231b3b80d8

Added a temporary option -no-isel-fold-inflight to control whether a "inflight"
node can be folded.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28003 91177308-0d34-0410-b5e6-96231b3b80d8

When isel'ing a node, mark its operands "InFlight" before selecting them. These
nodes should not be folded into other nodes.
This fixes the miscompilation of PR 749.
Temporarily under flag control.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28002 91177308-0d34-0410-b5e6-96231b3b80d8

When we have a two-address instruction where the input cannot be clobbered
and is already available, instead of falling back to emitting a load, fall
back to emitting a reg-reg copy.  This generates significantly better code
for some SSE testcases, as SSE has lots of two-address instructions and
none of them are read/modify/write.  As one example, this change does:

        pshufd %XMM5, XMMWORD PTR [%ESP + 84], 255
        xorps %XMM2, %XMM5
        cmpltps %XMM1, %XMM0
-       movaps XMMWORD PTR [%ESP + 52], %XMM0
-       movapd %XMM6, XMMWORD PTR [%ESP + 52]
+       movaps %XMM6, %XMM0
        cmpltps %XMM6, XMMWORD PTR [%ESP + 68]
        movapd XMMWORD PTR [%ESP + 52], %XMM6
        movaps %XMM6, %XMM0
        cmpltps %XMM6, XMMWORD PTR [%ESP + 36]
        cmpltps %XMM3, %XMM0
-       movaps XMMWORD PTR [%ESP + 20], %XMM0
-       movapd %XMM7, XMMWORD PTR [%ESP + 20]
+       movaps %XMM7, %XMM0
        cmpltps %XMM7, XMMWORD PTR [%ESP + 4]
        movapd XMMWORD PTR [%ESP + 20], %XMM7
        cmpltps %XMM4, %XMM0

... which is far better than a store followed by a load!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28001 91177308-0d34-0410-b5e6-96231b3b80d8

Test case for PR748

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28000 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27999 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27998 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for inserting undef into a vector.  This implements
Transforms/InstCombine/vec_insert_to_shuffle.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27997 91177308-0d34-0410-b5e6-96231b3b80d8

This should turn into one vector shuffle instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27996 91177308-0d34-0410-b5e6-96231b3b80d8

Make x86 isel lowering produce tailcall nodes. They are match to normal calls
for now.

Patch contributed by Alexander Friedman.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27994 91177308-0d34-0410-b5e6-96231b3b80d8

A couple of new entries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27993 91177308-0d34-0410-b5e6-96231b3b80d8

Support for passing 128-bit vector arguments via XMM registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27992 91177308-0d34-0410-b5e6-96231b3b80d8

Insert a VBIT_CONVERT between a FORMAL_ARGUMENT node and its vector uses
(VAND, VADD, etc.). Legalizer will assert otherwise.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27991 91177308-0d34-0410-b5e6-96231b3b80d8

For PR747:
If we fail to find a required program, simply set that program to echo
out something that tells the user the situation. That is, instead of just
"true runtest" we now get "echo 'Skipped: runtest not found'".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27990 91177308-0d34-0410-b5e6-96231b3b80d8

Oops

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27989 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fix: not updating NumIntRegs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27988 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Regression/CodeGen/Generic/2006-04-26-SetCCAnd.ll and
PR748.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27987 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27986 91177308-0d34-0410-b5e6-96231b3b80d8

- Clean up formal argument lowering code. Prepare for vector pass by value work.
- Fixed vararg support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27985 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some nondeterminstic behavior in the mem2reg pass that (in addition to
nondeterminism being bad) could cause some trivial missed optimizations (dead
phi nodes being left around for later passes to clean up).

With this, llvm-gcc4 now bootstraps and correctly compares.  I don't know
why I never tried to do it before... :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27984 91177308-0d34-0410-b5e6-96231b3b80d8

Actually, semantical doesn't appear to be a word.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27983 91177308-0d34-0410-b5e6-96231b3b80d8

Implement Transforms/IndVarsSimplify/complex-scev.ll, a case where we didn't
recognize some simple affine IV's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27982 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27981 91177308-0d34-0410-b5e6-96231b3b80d8

Fix fastcc failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27980 91177308-0d34-0410-b5e6-96231b3b80d8

Unfortunately this really isn't the place for advertisement.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27979 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27978 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some anchors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27977 91177308-0d34-0410-b5e6-96231b3b80d8

Add some notes about the current state of source (front end) languages
so we can point to them on llvm-dev.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27976 91177308-0d34-0410-b5e6-96231b3b80d8

Switching over FORMAL_ARGUMENTS mechanism to lower call arguments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27975 91177308-0d34-0410-b5e6-96231b3b80d8

Don't forget return void.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27974 91177308-0d34-0410-b5e6-96231b3b80d8

Keep the stack from on darwin 16-byte aligned.  This fixes many JIT
failres.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27973 91177308-0d34-0410-b5e6-96231b3b80d8

Separate LowerOperation() into multiple functions, one per opcode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27972 91177308-0d34-0410-b5e6-96231b3b80d8

slightly more useful error message

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27971 91177308-0d34-0410-b5e6-96231b3b80d8

better c99 struct handling

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27970 91177308-0d34-0410-b5e6-96231b3b80d8

another c99 style problem

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27969 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27968 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27967 91177308-0d34-0410-b5e6-96231b3b80d8

No functionality changes, but cleaner code with correct comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27966 91177308-0d34-0410-b5e6-96231b3b80d8

Explicitly specify result type for def : Pat<> patterns (if it produces a vector
result). Otherwise tblgen will pick the default (v16i8 for 128-bit vector).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27965 91177308-0d34-0410-b5e6-96231b3b80d8

Added X86 SSE2 intrinsics which can be represented as vector_shuffles. This is
a temporary workaround for the 2-wide vector_shuffle problem (i.e. its mask
would have type v2i32 which is not legal).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27964 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27963 91177308-0d34-0410-b5e6-96231b3b80d8

Added a movq test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27962 91177308-0d34-0410-b5e6-96231b3b80d8

Special case handling two wide build_vector(0, x).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27961 91177308-0d34-0410-b5e6-96231b3b80d8

Some missing movlps, movhps, movlpd, and movhpd patterns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27960 91177308-0d34-0410-b5e6-96231b3b80d8

A little bit more build_vector enhancement for v8i16 cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27959 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a completed entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27958 91177308-0d34-0410-b5e6-96231b3b80d8

Suggest looking at the SPARC backend.  How weird is that?? :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27957 91177308-0d34-0410-b5e6-96231b3b80d8

Added addJumpTableIndex

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27956 91177308-0d34-0410-b5e6-96231b3b80d8

MakeMIInst() should handle jump table index operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27955 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27954 91177308-0d34-0410-b5e6-96231b3b80d8

MOVL shuffle (i.e. movd or movss / movsd from memory) of undef, V2 == V2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27953 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the updating of the machine CFG when a PHI node was in a successor of
the jump table's range check block.  This re-enables 100% dense jump tables
by default on PPC & x86

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27952 91177308-0d34-0410-b5e6-96231b3b80d8

Keep Visual Studio informed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27951 91177308-0d34-0410-b5e6-96231b3b80d8

Code cleanup associated with jump tables, thanks to Chris for noticing
these.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27950 91177308-0d34-0410-b5e6-96231b3b80d8

Turn of jump tables for a bit, there are still some issues to work out with
updating the machine CFG.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27949 91177308-0d34-0410-b5e6-96231b3b80d8

Optimized stores to the constant pool, while cool, are unnecessary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27948 91177308-0d34-0410-b5e6-96231b3b80d8

JumpTable support!  What this represents is working asm and jit support for
x86 and ppc for 100% dense switch statements when relocations are non-PIC.
This support will be extended and enhanced in the coming days to support
PIC, and less dense forms of jump tables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27947 91177308-0d34-0410-b5e6-96231b3b80d8

Don't do all the lowering stuff for 2-wide build_vector's. Also, minor optimization for shuffle of undef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27946 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a performance regression. Use {p}shuf* when there are only two distinct elements in a build_vector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27945 91177308-0d34-0410-b5e6-96231b3b80d8

Two more build_vector tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27944 91177308-0d34-0410-b5e6-96231b3b80d8

Teach the JIT how to relocate LI, this fixes the JIT on Prolangs-C/TimberWolfMC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27943 91177308-0d34-0410-b5e6-96231b3b80d8

This no longer needs the JIT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27942 91177308-0d34-0410-b5e6-96231b3b80d8

Fix JIT support for static ctors, which was apparently completely broken!

This allows Prolangs-C++/city and probably a bunch of other stuff to work
well with the new front-end

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27941 91177308-0d34-0410-b5e6-96231b3b80d8

Revamp build_vector lowering to take advantage of movss and movd instructions.
movd always clear the top 96 bits and movss does so when it's loading the
value from memory.
The net result is codegen for 4-wide shuffles is much improved. It is near
optimal if one or more elements is a zero. e.g.

__m128i test(int a, int b) {
  return _mm_set_epi32(0, 0, b, a);
}

compiles to

_test:
movd 8(%esp), %xmm1
movd 4(%esp), %xmm0
punpckldq %xmm1, %xmm0
ret

compare to gcc:

_test:
subl $12, %esp
movd 20(%esp), %xmm0
movd 16(%esp), %xmm1
punpckldq %xmm0, %xmm1
movq %xmm1, %xmm0
movhps LC0, %xmm0
addl $12, %esp
ret

or icc:

_test:
        movd      4(%esp), %xmm0                                #5.10
        movd      8(%esp), %xmm3                                #5.10
        xorl      %eax, %eax                                    #5.10
        movd      %eax, %xmm1                                   #5.10
        punpckldq %xmm1, %xmm0                                  #5.10
        movd      %eax, %xmm2                                   #5.10
        punpckldq %xmm2, %xmm3                                  #5.10
        punpckldq %xmm3, %xmm0                                  #5.10
        ret                                                     #5.10

There are still room for improvement, for example the FP variant of the above example:

__m128 test(float a, float b) {
  return _mm_set_ps(0.0, 0.0, b, a);
}

_test:
movss 8(%esp), %xmm1
movss 4(%esp), %xmm0
unpcklps %xmm1, %xmm0
xorps %xmm1, %xmm1
movlhps %xmm1, %xmm0
ret

The xorps and movlhps are unnecessary. This will require post legalizer optimization to handle.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27939 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27938 91177308-0d34-0410-b5e6-96231b3b80d8

Change the PPC JIT to use a Static relocation model

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27937 91177308-0d34-0410-b5e6-96231b3b80d8

typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27936 91177308-0d34-0410-b5e6-96231b3b80d8

fix thinko

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27935 91177308-0d34-0410-b5e6-96231b3b80d8

add some low-prio notes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27934 91177308-0d34-0410-b5e6-96231b3b80d8

The BFS scheduler is apparently nondeterminstic (causes many llvmgcc bootstrap
miscompares).  Switch RISC targets to use the list-td scheduler, which isn't.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27933 91177308-0d34-0410-b5e6-96231b3b80d8

movddup is a SSE3 instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27932 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a hack required by V9.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27931 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a couple more memory issues

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27930 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the extraneous --defined-only option to nm. This is the default and
some versions of nm don't recognize it (its a gnu option).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27928 91177308-0d34-0410-b5e6-96231b3b80d8

Add && to each RUN: line (except the last one).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27927 91177308-0d34-0410-b5e6-96231b3b80d8

More build_vector tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27926 91177308-0d34-0410-b5e6-96231b3b80d8

Check for llc crash.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27925 91177308-0d34-0410-b5e6-96231b3b80d8

To be replaced with another test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27924 91177308-0d34-0410-b5e6-96231b3b80d8

Now generating perfect (I think) code for "vector set" with a single non-zero
scalar value.

e.g.
        _mm_set_epi32(0, a, 0, 0);
==>
movd 4(%esp), %xmm0
pshufd $69, %xmm0, %xmm0

        _mm_set_epi8(0, 0, 0, 0, 0, a, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
==>
movzbw 4(%esp), %ax
movzwl %ax, %eax
pxor %xmm0, %xmm0
pinsrw $5, %eax, %xmm0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27923 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a really subtle and obnoxious memory bug that caused issues with an
llvm-gcc4 boostrap.  Whenever a node is deleted by the dag combiner, it
*must* be returned by the visit function, or the dag combiner will not
know that the node has been processed (and will, e.g., send it to the
target dag combine xforms).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27922 91177308-0d34-0410-b5e6-96231b3b80d8

Don't require the "dot" program if -flat option is given, and don't produce
any of the fancy graphs or other output. Just produce the flat, makefile
style output on stdout.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27921 91177308-0d34-0410-b5e6-96231b3b80d8

Keep Visual Studio informed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27920 91177308-0d34-0410-b5e6-96231b3b80d8

Configure llvm-config in tools, not utils.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27919 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the llvm-config directory. Its now in tools.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27918 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-config doesn't live here anymore ==> tools

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27917 91177308-0d34-0410-b5e6-96231b3b80d8