llvm-svn: 69836

Real fix for PR3549, by using caching for predecessor counts in addition to the predecessors themselves.  This halves the time
to optimize the testcase, beyond what my previous patch did.

llvm-svn: 69792

Use PredIteratorCache in LCSSA, which gives a 37% overall speedup on the testcase from PR3549.  More improvements to come.

llvm-svn: 69788

llvm-svn: 69752

llvm-svn: 69680

and SCEVSignExtendExpr.

llvm-svn: 69649

the code to minimize dependencies on TargetData.

llvm-svn: 69644

GEP's don't usually become instructions.

llvm-svn: 69631

pointer type, make sure that the pointer size is a valid sequential index type.

llvm-svn: 69574

llvm-svn: 69450

llvm-svn: 69402

Use a safer iterator interface and get rid of std C++ library misuse.
This fixes a --enable-expensive-checks problem.

llvm-svn: 69353

regression in 403.gcc in PIC_CODEGEN=1 and DISABLE_LTO=1
mode.

llvm-svn: 69344

to get the correct answer for pointer types.

llvm-svn: 69321

incoming edges for a block with many predecessors.

llvm-svn: 69312

targets with pointers larger than 64 bits, due to the code not
yet being APInt clean.

llvm-svn: 69296

optimizer, which just happen to frequently involve optimizing GEPs.

llvm-svn: 69295

since the operand is always a constant.

llvm-svn: 69291

new instruction with SCEVExpander::InsertCastOfTo.

llvm-svn: 69290

have pointer types, though in contrast to C pointer types, SCEV
addition is never implicitly scaled. This not only eliminates the
need for special code like IndVars' EliminatePointerRecurrence
and LSR's own GEP expansion code, it also does a better job because
it lets the normal optimizations handle pointer expressions just
like integer expressions.

Also, since LLVM IR GEPs can't directly index into multi-dimensional
VLAs, moving the GEP analysis out of client code and into the SCEV
framework makes it easier for clients to handle multi-dimensional
VLAs the same way as other arrays.

Some existing regression tests show improved optimization.
test/CodeGen/ARM/2007-03-13-InstrSched.ll in particular improved to
the point where if-conversion started kicking in; I turned it off
for this test to preserve the intent of the test.

llvm-svn: 69258

and sext over (iv | const), if a longer iv is
available.  Allow expressions to have more than
one zext/sext parent.  All from OpenSSL.

llvm-svn: 69241

if a longer iv is available.  These subscript forms are
not common; they're a bottleneck in OpenSSL.

llvm-svn: 69215

sext around sext(shorter IV + constant), using a
longer IV instead, when it can figure out the
add can't overflow.  This comes up a lot in
subscripting; mainly affects 64 bit.

llvm-svn: 69123

llvm-svn: 69121

llvm.dbg.region.end instrinsic. This nested llvm.dbg.func.start/llvm.dbg.region.end pair now enables DW_TAG_inlined_subroutine support in code generator.

llvm-svn: 69118

This turns:

eq:
        %3 = icmp eq i32 %1, %2
        br label %join

ne:
        %4 = icmp ne i32 %1, %2
        br label %join

join:
        %5 = phi i1 [%3, %eq], [%4, %ne]
        br i1 %5, label %yes, label %no

=>

eq:
        %3 = icmp eq i32 %1, %2
        br i1 %3, label %yes, label %no

ne:
        %4 = icmp ne i32 %1, %2
        br i1 %4, label %yes, label %no

llvm-svn: 69102

deleting, not just the basic block.

llvm-svn: 69011

llvm-svn: 68939

strncat :(

strncat(foo, "bar", 99)
would be optimized to
memcpy(foo+strlen(foo), "bar", 100, 1)
instead of
memcpy(foo+strlen(foo), "bar", 4, 1)"

Patch by Benjamin Kramer!

llvm-svn: 68905

code.  Patch by Benjamin Kramer!

llvm-svn: 68885

integer types, unless they are already strange.  This prevents it from
turning the code produced by SROA into crazy libcalls and stuff that 
the code generator can't handle.  In the attached example, the result
was an i96 multiply that caused the x86 backend to assert.

Note that if TargetData had an idea of what the legal types are for
a target that this could be used to stop instcombine from introducing
i64 muls, as Scott wanted.

llvm-svn: 68598

llvm-svn: 68500

llvm-svn: 68485

instead of the place where it started to perform the string copy.

- PR3661
- Patch by Benjamin Kramer!

llvm-svn: 68443

llvm-svn: 68262

  Applications/Burg/burg
  Applications/ClamAV/clamscan
and many other tests.

llvm-svn: 68211

llvm-svn: 68172

if it dangles.

llvm-svn: 68150

Throttle back "fold select into operand" transformation. InstCombine should not generate selects of two constants unless they are selects of 0 and 1.

e.g.
define i32 @t1(i32 %c, i32 %x) nounwind {
       %t1 = icmp eq i32 %c, 0
       %t2 = lshr i32 %x, 18
       %t3 = select i1 %t1, i32 %t2, i32 %x
       ret i32 %t3
}

was turned into

define i32 @t2(i32 %c, i32 %x) nounwind {
       %t1 = icmp eq i32 %c, 0
       %t2 = select i1 %t1, i32 18, i32 0
       %t3 = lshr i32 %x, %t2
       ret i32 %t3
}

For most targets, that means materializing two constants and then a select. e.g. On x86-64

movl    %esi, %eax
shrl    $18, %eax
testl   %edi, %edi
cmovne  %esi, %eax
ret

=>

xorl    %eax, %eax
testl   %edi, %edi
movl    $18, %ecx
cmovne  %eax, %ecx
movl    %esi, %eax
shrl    %cl, %eax
ret

Also, the optimizer and codegen can reason about shl / and / add, etc. by a constant. This optimization will hinder optimizations using ComputeMaskedBits.

llvm-svn: 68142

Patch by Jay Foad.

llvm-svn: 68120