3912 Commits

Author SHA1 Message Date
Devang Patel
a8d63b0b30 Do not strip llvm.used values.
llvm-svn: 46045
2008-01-16 03:33:05 +00:00
Duncan Sands
81e35b4d47 I noticed that the trampoline straightening transformation could
drop attributes on varargs call arguments.  Also, it could generate
invalid IR if the transformed call already had the 'nest' attribute
somewhere (this can never happen for code coming from llvm-gcc,
but it's a theoretical possibility).  Fix both problems.

llvm-svn: 45973
2008-01-14 19:52:09 +00:00
Chris Lattner
efb498eddb Fix the miscompilation of MiBench/consumer-lame that was exposed by Evan's
byval work.  This miscompilation is due to the program indexing an array out
of range and us doing a transformation that broke this.

llvm-svn: 45949
2008-01-14 02:09:12 +00:00
Chris Lattner
5d135cc20a The isNotSuitableForSRA property is now dead, don't compute it.
llvm-svn: 45948
2008-01-14 01:32:52 +00:00
Chris Lattner
760b05a331 Change SRAGlobal to not depend on isNotSuitableForSRA, which makes it very
difficult to understand the invariants.

llvm-svn: 45947
2008-01-14 01:31:05 +00:00
Chris Lattner
c2933482a0 Make the 'shrink global to bool' optimization more self contained, and thus
easier to show that its safe.  No functionality change.

llvm-svn: 45946
2008-01-14 01:17:44 +00:00
Chris Lattner
d22a5f6314 Turn a memcpy from a double* into a load/store of double instead of
a load/store of i64.  The later prevents promotion/scalarrepl of the
source and dest in many cases.

This fixes the 300% performance regression of the byval stuff on 
stepanov_v1p2.

llvm-svn: 45945
2008-01-14 00:28:35 +00:00
Chris Lattner
8560bb9d98 factor memcpy/memmove simplification out to its own SimplifyMemTransfer
method, no functionality change.

llvm-svn: 45944
2008-01-13 23:50:23 +00:00
Chris Lattner
5fbf76aaf4 simplify some code. If we can infer alignment for source and dest that are
greater than memcpy alignment, and if we lower to load/store, use the best 
alignment info we have.

llvm-svn: 45943
2008-01-13 22:30:28 +00:00
Chris Lattner
4f69f1a721 simplify some code by adding a InsertBitCastBefore method,
make memmove->memcpy conversion a bit simpler.

llvm-svn: 45942
2008-01-13 22:23:22 +00:00
Chris Lattner
32eae5daa5 Fix PR1907, a nasty miscompilation because instcombine didn't
realize that ne & sgt  was a signed comparison (it was only 
looking at whether the left compare was signed).

llvm-svn: 45937
2008-01-13 20:59:02 +00:00
Duncan Sands
560625b701 Small simplification.
llvm-svn: 45932
2008-01-13 08:12:17 +00:00
Duncan Sands
7414cc131b When turning a call to a bitcast function into a direct call,
if this becomes a varargs call then deal correctly with any
parameter attributes on the newly vararg call arguments.

llvm-svn: 45931
2008-01-13 08:02:44 +00:00
Chris Lattner
d90840eddc we don't have to make an explicit copy of a byval argument when
inlining a function if we know that the function does not write
to *any* memory.  This implements test/Transforms/Inline/byval2.ll

llvm-svn: 45912
2008-01-12 18:54:29 +00:00
Chris Lattner
fb5876f0be Allow clients to specify the inline threshold when creating
the inliner pass.  Patch by Robert Zeh.

llvm-svn: 45903
2008-01-12 06:49:13 +00:00
Duncan Sands
6f49217a5e When DAE drops the varargs part of a function, ensure any
attributes on the vararg call arguments are also dropped.

llvm-svn: 45892
2008-01-11 23:13:45 +00:00
Chris Lattner
a1246ba8ed Teach argpromote to ruthlessly hack small byval structs when it can
get away with it, which exposes opportunities to eliminate the memory
objects entirely.  For example, we now compile byval.ll to:

define internal void @f1(i32 %b.0, i64 %b.1) {
entry:
	%tmp2 = add i32 %b.0, 1		; <i32> [#uses=0]
	ret void
}

define i32 @main() nounwind  {
entry:
	call void @f1( i32 1, i64 2 )
	ret i32 0
}

This seems like it would trigger a lot for code that passes around small
structs (e.g. SDOperand's or _Complex)...

llvm-svn: 45886
2008-01-11 22:31:41 +00:00
Chris Lattner
8644bdca87 Use smallptrset instead of std::set for efficiency.
llvm-svn: 45878
2008-01-11 19:36:30 +00:00
Chris Lattner
44aaf42d14 a byval argument is guaranteed to be valid to load.
llvm-svn: 45877
2008-01-11 19:34:32 +00:00
Chris Lattner
129a0e4f7d Update this code to use eraseFromParent where possible. Compute
whether an argument is byval and pass into isSafeToPromoteArgument.

llvm-svn: 45876
2008-01-11 19:20:39 +00:00
Chris Lattner
85a0b511cc replace a loop with a constant time check.
llvm-svn: 45875
2008-01-11 18:55:10 +00:00
Chris Lattner
c9666b967f another minor datastructure tweak.
llvm-svn: 45874
2008-01-11 18:47:45 +00:00
Chris Lattner
a6b0783f14 start using smallvector to avoid vector heap thrashing.
llvm-svn: 45873
2008-01-11 18:43:58 +00:00
Chris Lattner
bf51fecdc4 When inlining a functino with a byval argument, make an explicit
copy of it in case the callee modifies the struct.

llvm-svn: 45853
2008-01-11 06:09:30 +00:00
Chris Lattner
67f581b344 Implement PR1795, an instcombine hack for forming GEPs with integer pointer arithmetic.
llvm-svn: 45745
2008-01-08 07:23:51 +00:00
Duncan Sands
7955cf0cd7 Small cleanup for handling of type/parameter attribute
incompatibility.

llvm-svn: 45704
2008-01-07 17:16:06 +00:00
Gordon Henriksen
f0803127c6 Deleting an empty file. Thanks, /usr/bin/patch!
llvm-svn: 45675
2008-01-07 02:29:04 +00:00
Gordon Henriksen
db4f51e1b9 With this patch, the LowerGC transformation becomes the
ShadowStackCollector, which additionally has reduced overhead with
no sacrifice in portability.

Considering a function @fun with 8 loop-local roots,
ShadowStackCollector introduces the following overhead
(x86):

; shadowstack prologue
        movl    L_llvm_gc_root_chain$non_lazy_ptr, %eax
        movl    (%eax), %ecx
        movl    $___gc_fun, 20(%esp)
        movl    $0, 24(%esp)
        movl    $0, 28(%esp)
        movl    $0, 32(%esp)
        movl    $0, 36(%esp)
        movl    $0, 40(%esp)
        movl    $0, 44(%esp)
        movl    $0, 48(%esp)
        movl    $0, 52(%esp)
        movl    %ecx, 16(%esp)
        leal    16(%esp), %ecx
        movl    %ecx, (%eax)

; shadowstack loop overhead
        (none)

; shadowstack epilogue
        movl    48(%esp), %edx
        movl    %edx, (%ecx)

; shadowstack metadata
        .align  3
___gc_fun:                              # __gc_fun
        .long   8
        .space  4

In comparison to LowerGC:

; lowergc prologue
        movl    L_llvm_gc_root_chain$non_lazy_ptr, %eax
        movl    (%eax), %ecx
        movl    %ecx, 48(%esp)
        movl    $8, 52(%esp)
        movl    $0, 60(%esp)
        movl    $0, 56(%esp)
        movl    $0, 68(%esp)
        movl    $0, 64(%esp)
        movl    $0, 76(%esp)
        movl    $0, 72(%esp)
        movl    $0, 84(%esp)
        movl    $0, 80(%esp)
        movl    $0, 92(%esp)
        movl    $0, 88(%esp)
        movl    $0, 100(%esp)
        movl    $0, 96(%esp)
        movl    $0, 108(%esp)
        movl    $0, 104(%esp)
        movl    $0, 116(%esp)
        movl    $0, 112(%esp)

; lowergc loop overhead
        leal    44(%esp), %eax
        movl    %eax, 56(%esp)
        leal    40(%esp), %eax
        movl    %eax, 64(%esp)
        leal    36(%esp), %eax
        movl    %eax, 72(%esp)
        leal    32(%esp), %eax
        movl    %eax, 80(%esp)
        leal    28(%esp), %eax
        movl    %eax, 88(%esp)
        leal    24(%esp), %eax
        movl    %eax, 96(%esp)
        leal    20(%esp), %eax
        movl    %eax, 104(%esp)
        leal    16(%esp), %eax
        movl    %eax, 112(%esp)

; lowergc epilogue
        movl    48(%esp), %edx
        movl    %edx, (%ecx)

; lowergc metadata
        (none)

llvm-svn: 45670
2008-01-07 01:30:53 +00:00
Duncan Sands
fd975e4b3d The transform that tries to turn calls to bitcast functions into
direct calls bails out unless caller and callee have essentially
equivalent parameter attributes.  This is illogical - the callee's
attributes should be of no relevance here.  Rework the logic, which
incidentally fixes a crash when removed arguments have attributes.

llvm-svn: 45658
2008-01-06 18:27:01 +00:00
Duncan Sands
b8489f09a2 When transforming a call to a bitcast function into
a direct call with cast parameters and cast return
value (if any), instcombine was prepared to cast any
non-void return value into any other, whether castable
or not.  Add a new predicate for testing whether casting
is valid, and check it both for the return value and
(as a cleanup) for the parameters.

llvm-svn: 45657
2008-01-06 10:12:28 +00:00
Chris Lattner
7e1c3aa702 remove a couple more unsafe xforms in the face of overflow.
llvm-svn: 45613
2008-01-05 01:22:42 +00:00
Chris Lattner
983697dfac remove the (x-y) < 0 comparison xform, it miscompiles
things that are not equality comparisons, for example:
   (2147479553+4096)-2147479553 < 0    !=   (2147479553+4096) < 2147479553

llvm-svn: 45612
2008-01-05 01:18:20 +00:00
Wojciech Matyjewicz
9ec15f974f fix typo
llvm-svn: 45594
2008-01-04 20:02:18 +00:00
Chris Lattner
d4c66656a1 Fix PR1896
llvm-svn: 45568
2008-01-04 05:04:53 +00:00
Chris Lattner
26b89fd30a don't hoist FP additions into unconditional adds + selects. This
could theoretically introduce a trap, but is also a performance issue.
This speeds up ptrdist/ks by 8%.

llvm-svn: 45533
2008-01-03 07:25:26 +00:00
Chris Lattner
028f584087 add missing #include
llvm-svn: 45516
2008-01-02 23:41:05 +00:00
Chris Lattner
ad9a6ccb83 Remove attribution from file headers, per discussion on llvmdev.
llvm-svn: 45418
2007-12-29 20:36:04 +00:00
Chris Lattner
8193d4af33 remove attribution from lib Makefiles.
llvm-svn: 45415
2007-12-29 20:09:26 +00:00
Christopher Lamb
dfad5f19b4 Disable null pointer folding transforms for non-generic address spaces. This should probably be a target-specific predicate based on address space. That way for targets where this isn't applicable the predicate can be optimized away.
llvm-svn: 45403
2007-12-29 07:56:53 +00:00
Chris Lattner
2369d2f4ab dead calls to llvm.stacksave can be deleted, even though they
have potential side-effects.

llvm-svn: 45392
2007-12-29 00:59:12 +00:00
Owen Anderson
ebd3e9c500 Repair a transform that Chris noticed a bug in. Thanks to Nicholas for pointing out my stupid mistakes when writing this patch. :-)
llvm-svn: 45384
2007-12-28 07:42:12 +00:00
Chris Lattner
2456399ce5 disable this instcombine xform, it miscompiles:
define i32 @main() {
entry:
	%z = alloca i32		; <i32*> [#uses=2]
	store i32 0, i32* %z
	%tmp = load i32* %z		; <i32> [#uses=1]
	%sub = sub i32 %tmp, 1		; <i32> [#uses=1]
	%cmp = icmp ult i32 %sub, 0		; <i1> [#uses=1]
	%retval = select i1 %cmp, i32 1, i32 0		; <i32> [#uses=1]
	ret i32 %retval
}

into ret 1, instead of ret 0.

Christopher, please investigate.

llvm-svn: 45383
2007-12-28 06:24:31 +00:00
Gordon Henriksen
90d48b077d Fixing several transforms which would drop the collector attribute
when copying functions.

llvm-svn: 45356
2007-12-25 22:16:06 +00:00
Chris Lattner
90df7f7424 Don't break critical edges for single-bb loops, this helps with PR1877, though
it is only a partial fix.  This change is noise for most programs, but
speeds up Shootout-C++/matrix by 20%, Ptrdist/ks by 24%, smg2000 by 8%,
hexxagon by 9%, bzip2 by 9% (not sure I trust this), ackerman by 13%, etc.

OTOH, it slows down Shootout/fib2 by 40% (I'll update PR1877 with this info).

llvm-svn: 45354
2007-12-25 19:06:45 +00:00
Gordon Henriksen
c0a3899bbf GC poses hazards to the inliner. Consider:
define void @f() {
            ...
            call i32 @g()
            ...
    }

    define void @g() {
            ...
    }

The hazards are:

  - @f and @g have GC, but they differ GC. Inlining is invalid. This
    may never occur.
  - @f has no GC, but @g does. g's GC must be propagated to @f.

The other scenarios are safe:

  - @f and @g have the same GC.
  - @f and @g have no GC.
  - @g has no GC.

This patch adds inliner checks for the former two scenarios.

llvm-svn: 45351
2007-12-25 03:10:07 +00:00
Chris Lattner
7e1e1f2933 add a -backedge-hack llc-beta option to codegenprepare.
When specified, don't split backedges of single-bb loops.
This helps address PR1877

llvm-svn: 45344
2007-12-24 19:32:55 +00:00
Chris Lattner
d64df490ca implement InstCombine/shift-trunc-shift.ll. This allows
us to compile:
#include <math.h>
int t1(double d) { return signbit(d); }

into:

_t1:
	movd	%xmm0, %rax
	shrq	$63, %rax
	ret

instead of:

_t1:
	movd	%xmm0, %rax
	shrq	$32, %rax
	shrl	$31, %eax
	ret

on x86-64.

llvm-svn: 45311
2007-12-22 09:07:47 +00:00
Devang Patel
e035f776e9 If succ has succ itself as one of the predecessors then do
not merge current bb and succ even if bb's terminator is
unconditional branch to succ.

llvm-svn: 45305
2007-12-22 01:32:53 +00:00
Duncan Sands
85ca85c070 Make DAE not wipe out attributes on calls, and not drop
return attributes on the floor.  In the case of a call
to a varargs function where the varargs arguments are
being removed, any call attributes on those arguments
need to be dropped.  I didn't do this because I plan to
make it illegal to have such attributes (see next patch).
With this change, compiling the gcc filter2 eh test at -O0
and then running opt -std-compile-opts on it results in
a correctly working program (compiling at -O1 or higher
results in the test failing due to a problem with how we
output eh info into the IR).

llvm-svn: 45285
2007-12-21 19:16:16 +00:00
Christopher Lamb
7ca648a7b1 Implement review feedback, including additional transforms
(icmp slt (sub A B) 1) -> (icmp sle A B)
icmp sgt (sub A B) -1) -> (icmp sge A B)

and add testcase.

llvm-svn: 45256
2007-12-20 07:21:11 +00:00