README.txt

	%.rle = phi i32 [ 0, %entry ], [ %.rle6, %bb2 ]	
	%i.05 = phi i32 [ 0, %entry ], [ %indvar.next, %bb2 ]
	%1 = load i32* %cond, align 4
	%2 = icmp eq i32 %1, 0
	br i1 %2, label %bb2, label %bb1

bb1:		; preds = %bb
	%3 = xor i32 %.rle, 234	
	store i32 %3, i32* %res, align 4
	br label %bb2

bb2:		; preds = %bb, %bb1
	%.rle6 = phi i32 [ %3, %bb1 ], [ %.rle, %bb ]	
	%indvar.next = add i32 %i.05, 1	
	%exitcond = icmp eq i32 %indvar.next, %n
	br i1 %exitcond, label %return, label %bb

DSE should sink partially dead stores to get the store out of the loop.

Here's another partial dead case:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=12395

//===---------------------------------------------------------------------===//

Scalar PRE hoists the mul in the common block up to the else:

int test (int a, int b, int c, int g) {
  int d, e;
  if (a)
    d = b * c;
  else
    d = b - c;
  e = b * c + g;
  return d + e;
}

It would be better to do the mul once to reduce codesize above the if.
This is GCC PR38204.


//===---------------------------------------------------------------------===//
This simple function from 179.art:

int winner, numf2s;
struct { double y; int   reset; } *Y;

void find_match() {
   int i;
   winner = 0;
   for (i=0;i<numf2s;i++)
       if (Y[i].y > Y[winner].y)
              winner =i;
}

Compiles into (with clang TBAA):

for.body:                                         ; preds = %for.inc, %bb.nph
  %indvar = phi i64 [ 0, %bb.nph ], [ %indvar.next, %for.inc ]
  %i.01718 = phi i32 [ 0, %bb.nph ], [ %i.01719, %for.inc ]
  %tmp4 = getelementptr inbounds %struct.anon* %tmp3, i64 %indvar, i32 0
  %tmp5 = load double* %tmp4, align 8, !tbaa !4
  %idxprom7 = sext i32 %i.01718 to i64
  %tmp10 = getelementptr inbounds %struct.anon* %tmp3, i64 %idxprom7, i32 0
  %tmp11 = load double* %tmp10, align 8, !tbaa !4
  %cmp12 = fcmp ogt double %tmp5, %tmp11
  br i1 %cmp12, label %if.then, label %for.inc

if.then:                                          ; preds = %for.body
  %i.017 = trunc i64 %indvar to i32
  br label %for.inc

for.inc:                                          ; preds = %for.body, %if.then
  %i.01719 = phi i32 [ %i.01718, %for.body ], [ %i.017, %if.then ]
  %indvar.next = add i64 %indvar, 1
  %exitcond = icmp eq i64 %indvar.next, %tmp22
  br i1 %exitcond, label %for.cond.for.end_crit_edge, label %for.body


It is good that we hoisted the reloads of numf2's, and Y out of the loop and
sunk the store to winner out.

However, this is awful on several levels: the conditional truncate in the loop
(-indvars at fault? why can't we completely promote the IV to i64?).

Beyond that, we have a partially redundant load in the loop: if "winner" (aka 
%i.01718) isn't updated, we reload Y[winner].y the next time through the loop.
Similarly, the addressing that feeds it (including the sext) is redundant. In
the end we get this generated assembly:

LBB0_2:                                 ## %for.body
                                        ## =>This Inner Loop Header: Depth=1
	movsd	(%rdi), %xmm0
	movslq	%edx, %r8
	shlq	$4, %r8
	ucomisd	(%rcx,%r8), %xmm0
	jbe	LBB0_4
	movl	%esi, %edx
LBB0_4:                                 ## %for.inc
	addq	$16, %rdi
	incq	%rsi
	cmpq	%rsi, %rax
	jne	LBB0_2

All things considered this isn't too bad, but we shouldn't need the movslq or
the shlq instruction, or the load folded into ucomisd every time through the
loop.

On an x86-specific topic, if the loop can't be restructure, the movl should be a
cmov.

//===---------------------------------------------------------------------===//

[STORE SINKING]

GCC PR37810 is an interesting case where we should sink load/store reload
into the if block and outside the loop, so we don't reload/store it on the
non-call path.

for () {
  *P += 1;
  if ()
    call();
  else
    ...
->
tmp = *P
for () {
  tmp += 1;
  if () {
    *P = tmp;
    call();
    tmp = *P;
  } else ...
}
*P = tmp;

We now hoist the reload after the call (Transforms/GVN/lpre-call-wrap.ll), but
we don't sink the store.  We need partially dead store sinking.

//===---------------------------------------------------------------------===//

[LOAD PRE CRIT EDGE SPLITTING]

GCC PR37166: Sinking of loads prevents SROA'ing the "g" struct on the stack
leading to excess stack traffic. This could be handled by GVN with some crazy
symbolic phi translation.  The code we get looks like (g is on the stack):

bb2:		; preds = %bb1
..
	%9 = getelementptr %struct.f* %g, i32 0, i32 0		
	store i32 %8, i32* %9, align  bel %bb3

bb3:		; preds = %bb1, %bb2, %bb
	%c_addr.0 = phi %struct.f* [ %g, %bb2 ], [ %c, %bb ], [ %c, %bb1 ]
	%b_addr.0 = phi %struct.f* [ %b, %bb2 ], [ %g, %bb ], [ %b, %bb1 ]
	%10 = getelementptr %struct.f* %c_addr.0, i32 0, i32 0
	%11 = load i32* %10, align 4

%11 is partially redundant, an in BB2 it should have the value %8.

GCC PR33344 and PR35287 are similar cases.


//===---------------------------------------------------------------------===//

[LOAD PRE]

There are many load PRE testcases in testsuite/gcc.dg/tree-ssa/loadpre* in the
GCC testsuite, ones we don't get yet are (checked through loadpre25):

[CRIT EDGE BREAKING]
loadpre3.c predcom-4.c

[PRE OF READONLY CALL]
loadpre5.c

[TURN SELECT INTO BRANCH]
loadpre14.c loadpre15.c 

actually a conditional increment: loadpre18.c loadpre19.c

//===---------------------------------------------------------------------===//

[LOAD PRE / STORE SINKING / SPEC HACK]

This is a chunk of code from 456.hmmer:

int f(int M, int *mc, int *mpp, int *tpmm, int *ip, int *tpim, int *dpp,
     int *tpdm, int xmb, int *bp, int *ms) {
 int k, sc;
 for (k = 1; k <= M; k++) {
     mc[k] = mpp[k-1]   + tpmm[k-1];
     if ((sc = ip[k-1]  + tpim[k-1]) > mc[k])  mc[k] = sc;
     if ((sc = dpp[k-1] + tpdm[k-1]) > mc[k])  mc[k] = sc;
     if ((sc = xmb  + bp[k])         > mc[k])  mc[k] = sc;
     mc[k] += ms[k];
   }
}

It is very profitable for this benchmark to turn the conditional stores to mc[k]
into a conditional move (select instr in IR) and allow the final store to do the
store.  See GCC PR27313 for more details.  Note that this is valid to xform even
with the new C++ memory model, since mc[k] is previously loaded and later
stored.

//===---------------------------------------------------------------------===//

[SCALAR PRE]
There are many PRE testcases in testsuite/gcc.dg/tree-ssa/ssa-pre-*.c in the
GCC testsuite.

//===---------------------------------------------------------------------===//

There are some interesting cases in testsuite/gcc.dg/tree-ssa/pred-comm* in the
GCC testsuite.  For example, we get the first example in predcom-1.c, but 
miss the second one:

unsigned fib[1000];
unsigned avg[1000];

__attribute__ ((noinline))
void count_averages(int n) {
  int i;
  for (i = 1; i < n; i++)
    avg[i] = (((unsigned long) fib[i - 1] + fib[i] + fib[i + 1]) / 3) & 0xffff;
}

which compiles into two loads instead of one in the loop.

predcom-2.c is the same as predcom-1.c

predcom-3.c is very similar but needs loads feeding each other instead of
store->load.


//===---------------------------------------------------------------------===//

[ALIAS ANALYSIS]

Type based alias analysis:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14705

We should do better analysis of posix_memalign.  At the least it should
no-capture its pointer argument, at best, we should know that the out-value
result doesn't point to anything (like malloc).  One example of this is in
SingleSource/Benchmarks/Misc/dt.c

//===---------------------------------------------------------------------===//

Interesting missed case because of control flow flattening (should be 2 loads):
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26629
With: llvm-gcc t2.c -S -o - -O0 -emit-llvm | llvm-as | 
             opt -mem2reg -gvn -instcombine | llvm-dis
we miss it because we need 1) CRIT EDGE 2) MULTIPLE DIFFERENT
VALS PRODUCED BY ONE BLOCK OVER DIFFERENT PATHS

//===---------------------------------------------------------------------===//

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19633
We could eliminate the branch condition here, loading from null is undefined:

struct S { int w, x, y, z; };
struct T { int r; struct S s; };
void bar (struct S, int);
void foo (int a, struct T b)
{
  struct S *c = 0;
  if (a)
    c = &b.s;
  bar (*c, a);
}

//===---------------------------------------------------------------------===//

simplifylibcalls should do several optimizations for strspn/strcspn:

strcspn(x, "a") -> inlined loop for up to 3 letters (similarly for strspn):

size_t __strcspn_c3 (__const char *__s, int __reject1, int __reject2,
                     int __reject3) {
  register size_t __result = 0;
  while (__s[__result] != '\0' && __s[__result] != __reject1 &&
         __s[__result] != __reject2 && __s[__result] != __reject3)
    ++__result;
  return __result;
}

This should turn into a switch on the character.  See PR3253 for some notes on
codegen.

456.hmmer apparently uses strcspn and strspn a lot.  471.omnetpp uses strspn.

//===---------------------------------------------------------------------===//

simplifylibcalls should turn these snprintf idioms into memcpy (GCC PR47917)

char buf1[6], buf2[6], buf3[4], buf4[4];
int i;

int foo (void) {
  int ret = snprintf (buf1, sizeof buf1, "abcde");
  ret += snprintf (buf2, sizeof buf2, "abcdef") * 16;
  ret += snprintf (buf3, sizeof buf3, "%s", i++ < 6 ? "abc" : "def") * 256;
  ret += snprintf (buf4, sizeof buf4, "%s", i++ > 10 ? "abcde" : "defgh")*4096;
  return ret;
}

//===---------------------------------------------------------------------===//

"gas" uses this idiom:
  else if (strchr ("+-/*%|&^:[]()~", *intel_parser.op_string))
..
  else if (strchr ("<>", *intel_parser.op_string)

Those should be turned into a switch.

//===---------------------------------------------------------------------===//

252.eon contains this interesting code:

        %3072 = getelementptr [100 x i8]* %tempString, i32 0, i32 0
        %3073 = call i8* @strcpy(i8* %3072, i8* %3071) nounwind
        %strlen = call i32 @strlen(i8* %3072)    ; uses = 1
        %endptr = getelementptr [100 x i8]* %tempString, i32 0, i32 %strlen
        call void @llvm.memcpy.i32(i8* %endptr, 
          i8* getelementptr ([5 x i8]* @"\01LC42", i32 0, i32 0), i32 5, i32 1)
        %3074 = call i32 @strlen(i8* %endptr) nounwind readonly 
        
This is interesting for a couple reasons.  First, in this:

The memcpy+strlen strlen can be replaced with:

        %3074 = call i32 @strlen([5 x i8]* @"\01LC42") nounwind readonly 

Because the destination was just copied into the specified memory buffer.  This,
in turn, can be constant folded to "4".

In other code, it contains:

        %endptr6978 = bitcast i8* %endptr69 to i32*            
        store i32 7107374, i32* %endptr6978, align 1
        %3167 = call i32 @strlen(i8* %endptr69) nounwind readonly    

Which could also be constant folded.  Whatever is producing this should probably
be fixed to leave this as a memcpy from a string.

Further, eon also has an interesting partially redundant strlen call:

bb8:            ; preds = %_ZN18eonImageCalculatorC1Ev.exit
        %682 = getelementptr i8** %argv, i32 6          ; <i8**> [#uses=2]
        %683 = load i8** %682, align 4          ; <i8*> [#uses=4]
        %684 = load i8* %683, align 1           ; <i8> [#uses=1]
        %685 = icmp eq i8 %684, 0               ; <i1> [#uses=1]
        br i1 %685, label %bb10, label %bb9

bb9:            ; preds = %bb8
        %686 = call i32 @strlen(i8* %683) nounwind readonly          
        %687 = icmp ugt i32 %686, 254           ; <i1> [#uses=1]
        br i1 %687, label %bb10, label %bb11

bb10:           ; preds = %bb9, %bb8
        %688 = call i32 @strlen(i8* %683) nounwind readonly          

This could be eliminated by doing the strlen once in bb8, saving code size and
improving perf on the bb8->9->10 path.

//===---------------------------------------------------------------------===//

I see an interesting fully redundant call to strlen left in 186.crafty:InputMove
which looks like:
       %movetext11 = getelementptr [128 x i8]* %movetext, i32 0, i32 0 
 

bb62:           ; preds = %bb55, %bb53
        %promote.0 = phi i32 [ %169, %bb55 ], [ 0, %bb53 ]             
        %171 = call i32 @strlen(i8* %movetext11) nounwind readonly align 1
        %172 = add i32 %171, -1         ; <i32> [#uses=1]
        %173 = getelementptr [128 x i8]* %movetext, i32 0, i32 %172       

...  no stores ...
       br i1 %or.cond, label %bb65, label %bb72

bb65:           ; preds = %bb62
        store i8 0, i8* %173, align 1
        br label %bb72

bb72:           ; preds = %bb65, %bb62
        %trank.1 = phi i32 [ %176, %bb65 ], [ -1, %bb62 ]            
        %177 = call i32 @strlen(i8* %movetext11) nounwind readonly align 1

Note that on the bb62->bb72 path, that the %177 strlen call is partially
redundant with the %171 call.  At worst, we could shove the %177 strlen call
up into the bb65 block moving it out of the bb62->bb72 path.   However, note
that bb65 stores to the string, zeroing out the last byte.  This means that on
that path the value of %177 is actually just %171-1.  A sub is cheaper than a
strlen!

This pattern repeats several times, basically doing:

  A = strlen(P);
  P[A-1] = 0;
  B = strlen(P);
  where it is "obvious" that B = A-1.

//===---------------------------------------------------------------------===//

186.crafty has this interesting pattern with the "out.4543" variable:

call void @llvm.memcpy.i32(
        i8* getelementptr ([10 x i8]* @out.4543, i32 0, i32 0),
       i8* getelementptr ([7 x i8]* @"\01LC28700", i32 0, i32 0), i32 7, i32 1) 
%101 = call@printf(i8* ...   @out.4543, i32 0, i32 0)) nounwind 

It is basically doing:

  memcpy(globalarray, "string");
  printf(...,  globalarray);
  
Anyway, by knowing that printf just reads the memory and forward substituting
the string directly into the printf, this eliminates reads from globalarray.
Since this pattern occurs frequently in crafty (due to the "DisplayTime" and
other similar functions) there are many stores to "out".  Once all the printfs
stop using "out", all that is left is the memcpy's into it.  This should allow
globalopt to remove the "stored only" global.

//===---------------------------------------------------------------------===//

This code:

define inreg i32 @foo(i8* inreg %p) nounwind {
  %tmp0 = load i8* %p
  %tmp1 = ashr i8 %tmp0, 5
  %tmp2 = sext i8 %tmp1 to i32
  ret i32 %tmp2
}

could be dagcombine'd to a sign-extending load with a shift.
For example, on x86 this currently gets this:

	movb	(%eax), %al
	sarb	$5, %al
	movsbl	%al, %eax

while it could get this:

	movsbl	(%eax), %eax
	sarl	$5, %eax

//===---------------------------------------------------------------------===//

GCC PR31029:

int test(int x) { return 1-x == x; }     // --> return false
int test2(int x) { return 2-x == x; }    // --> return x == 1 ?

Always foldable for odd constants, what is the rule for even?

//===---------------------------------------------------------------------===//

PR 3381: GEP to field of size 0 inside a struct could be turned into GEP
for next field in struct (which is at same address).

For example: store of float into { {{}}, float } could be turned into a store to
the float directly.

//===---------------------------------------------------------------------===//

The arg promotion pass should make use of nocapture to make its alias analysis
stuff much more precise.

//===---------------------------------------------------------------------===//

The following functions should be optimized to use a select instead of a
branch (from gcc PR40072):

char char_int(int m) {if(m>7) return 0; return m;}
int int_char(char m) {if(m>7) return 0; return m;}

//===---------------------------------------------------------------------===//

int func(int a, int b) { if (a & 0x80) b |= 0x80; else b &= ~0x80; return b; }

Generates this:

define i32 @func(i32 %a, i32 %b) nounwind readnone ssp {
entry:
  %0 = and i32 %a, 128                            ; <i32> [#uses=1]
  %1 = icmp eq i32 %0, 0                          ; <i1> [#uses=1]
  %2 = or i32 %b, 128                             ; <i32> [#uses=1]
  %3 = and i32 %b, -129                           ; <i32> [#uses=1]
  %b_addr.0 = select i1 %1, i32 %3, i32 %2        ; <i32> [#uses=1]
  ret i32 %b_addr.0
}

However, it's functionally equivalent to:

         b = (b & ~0x80) | (a & 0x80);

Which generates this:

define i32 @func(i32 %a, i32 %b) nounwind readnone ssp {
entry:
  %0 = and i32 %b, -129                           ; <i32> [#uses=1]
  %1 = and i32 %a, 128                            ; <i32> [#uses=1]
  %2 = or i32 %0, %1                              ; <i32> [#uses=1]
  ret i32 %2
}

This can be generalized for other forms:

     b = (b & ~0x80) | (a & 0x40) << 1;

//===---------------------------------------------------------------------===//

These two functions produce different code. They shouldn't:

#include <stdint.h>
 
uint8_t p1(uint8_t b, uint8_t a) {
  b = (b & ~0xc0) | (a & 0xc0);
  return (b);
}
 
uint8_t p2(uint8_t b, uint8_t a) {
  b = (b & ~0x40) | (a & 0x40);
  b = (b & ~0x80) | (a & 0x80);
  return (b);
}

define zeroext i8 @p1(i8 zeroext %b, i8 zeroext %a) nounwind readnone ssp {
entry:
  %0 = and i8 %b, 63                              ; <i8> [#uses=1]
  %1 = and i8 %a, -64                             ; <i8> [#uses=1]
  %2 = or i8 %1, %0                               ; <i8> [#uses=1]
  ret i8 %2
}

define zeroext i8 @p2(i8 zeroext %b, i8 zeroext %a) nounwind readnone ssp {
entry:
  %0 = and i8 %b, 63                              ; <i8> [#uses=1]
  %.masked = and i8 %a, 64                        ; <i8> [#uses=1]
  %1 = and i8 %a, -128                            ; <i8> [#uses=1]
  %2 = or i8 %1, %0                               ; <i8> [#uses=1]
  %3 = or i8 %2, %.masked                         ; <i8> [#uses=1]
  ret i8 %3
}

//===---------------------------------------------------------------------===//

IPSCCP does not currently propagate argument dependent constants through
functions where it does not not all of the callers.  This includes functions
with normal external linkage as well as templates, C99 inline functions etc.
Specifically, it does nothing to:

define i32 @test(i32 %x, i32 %y, i32 %z) nounwind {
entry:
  %0 = add nsw i32 %y, %z                         
  %1 = mul i32 %0, %x                             
  %2 = mul i32 %y, %z                             
  %3 = add nsw i32 %1, %2                         
  ret i32 %3
}

define i32 @test2() nounwind {
entry:
  %0 = call i32 @test(i32 1, i32 2, i32 4) nounwind
  ret i32 %0
}

It would be interesting extend IPSCCP to be able to handle simple cases like
this, where all of the arguments to a call are constant.  Because IPSCCP runs
before inlining, trivial templates and inline functions are not yet inlined.
The results for a function + set of constant arguments should be memoized in a
map.

//===---------------------------------------------------------------------===//

The libcall constant folding stuff should be moved out of SimplifyLibcalls into
libanalysis' constantfolding logic.  This would allow IPSCCP to be able to
handle simple things like this:

static int foo(const char *X) { return strlen(X); }
int bar() { return foo("abcd"); }

//===---------------------------------------------------------------------===//

functionattrs doesn't know much about memcpy/memset.  This function should be
marked readnone rather than readonly, since it only twiddles local memory, but
functionattrs doesn't handle memset/memcpy/memmove aggressively:

struct X { int *p; int *q; };
int foo() {
 int i = 0, j = 1;
 struct X x, y;
 int **p;
 y.p = &i;
 x.q = &j;
 p = __builtin_memcpy (&x, &y, sizeof (int *));
 return **p;
}

This can be seen at:
$ clang t.c -S -o - -mkernel -O0 -emit-llvm | opt -functionattrs -S


//===---------------------------------------------------------------------===//

Missed instcombine transformation:
define i1 @a(i32 %x) nounwind readnone {
entry:
  %cmp = icmp eq i32 %x, 30
  %sub = add i32 %x, -30
  %cmp2 = icmp ugt i32 %sub, 9
  %or = or i1 %cmp, %cmp2
  ret i1 %or
}
This should be optimized to a single compare.  Testcase derived from gcc.

//===---------------------------------------------------------------------===//

Missed instcombine or reassociate transformation:
int a(int a, int b) { return (a==12)&(b>47)&(b<58); }

The sgt and slt should be combined into a single comparison. Testcase derived
from gcc.

//===---------------------------------------------------------------------===//

Missed instcombine transformation:

  %382 = srem i32 %tmp14.i, 64                    ; [#uses=1]
  %383 = zext i32 %382 to i64                     ; [#uses=1]
  %384 = shl i64 %381, %383                       ; [#uses=1]
  %385 = icmp slt i32 %tmp14.i, 64                ; [#uses=1]

The srem can be transformed to an and because if %tmp14.i is negative, the
shift is undefined.  Testcase derived from 403.gcc.

//===---------------------------------------------------------------------===//

This is a range comparison on a divided result (from 403.gcc):

  %1337 = sdiv i32 %1336, 8                       ; [#uses=1]
  %.off.i208 = add i32 %1336, 7                   ; [#uses=1]
  %1338 = icmp ult i32 %.off.i208, 15             ; [#uses=1]
  
We already catch this (removing the sdiv) if there isn't an add, we should
handle the 'add' as well.  This is a common idiom with it's builtin_alloca code.
C testcase:

int a(int x) { return (unsigned)(x/16+7) < 15; }

Another similar case involves truncations on 64-bit targets:

  %361 = sdiv i64 %.046, 8                        ; [#uses=1]
  %362 = trunc i64 %361 to i32                    ; [#uses=2]
...
  %367 = icmp eq i32 %362, 0                      ; [#uses=1]

//===---------------------------------------------------------------------===//

Missed instcombine/dagcombine transformation:
define void @lshift_lt(i8 zeroext %a) nounwind {
entry:
  %conv = zext i8 %a to i32
  %shl = shl i32 %conv, 3
  %cmp = icmp ult i32 %shl, 33
  br i1 %cmp, label %if.then, label %if.end

if.then:
  tail call void @bar() nounwind
  ret void

if.end:
  ret void
}
declare void @bar() nounwind

The shift should be eliminated.  Testcase derived from gcc.

//===---------------------------------------------------------------------===//

These compile into different code, one gets recognized as a switch and the
other doesn't due to phase ordering issues (PR6212):

int test1(int mainType, int subType) {
  if (mainType == 7)
    subType = 4;
  else if (mainType == 9)
    subType = 6;
  else if (mainType == 11)
    subType = 9;
  return subType;
}

int test2(int mainType, int subType) {
  if (mainType == 7)
    subType = 4;
  if (mainType == 9)
    subType = 6;
  if (mainType == 11)
    subType = 9;
  return subType;
}

//===---------------------------------------------------------------------===//

The following test case (from PR6576):

define i32 @mul(i32 %a, i32 %b) nounwind readnone {
entry:
 %cond1 = icmp eq i32 %b, 0                      ; <i1> [#uses=1]
 br i1 %cond1, label %exit, label %bb.nph
bb.nph:                                           ; preds = %entry
 %tmp = mul i32 %b, %a                           ; <i32> [#uses=1]
 ret i32 %tmp
exit:                                             ; preds = %entry
 ret i32 0
}

could be reduced to:

define i32 @mul(i32 %a, i32 %b) nounwind readnone {
entry:
 %tmp = mul i32 %b, %a
 ret i32 %tmp
}

//===---------------------------------------------------------------------===//

We should use DSE + llvm.lifetime.end to delete dead vtable pointer updates.
See GCC PR34949

Another interesting case is that something related could be used for variables
that go const after their ctor has finished.  In these cases, globalopt (which
can statically run the constructor) could mark the global const (so it gets put
in the readonly section).  A testcase would be:

#include <complex>
using namespace std;
const complex<char> should_be_in_rodata (42,-42);
complex<char> should_be_in_data (42,-42);
complex<char> should_be_in_bss;

Where we currently evaluate the ctors but the globals don't become const because
the optimizer doesn't know they "become const" after the ctor is done.  See
GCC PR4131 for more examples.

//===---------------------------------------------------------------------===//

In this code:

long foo(long x) {
  return x > 1 ? x : 1;
}

LLVM emits a comparison with 1 instead of 0. 0 would be equivalent
and cheaper on most targets.

LLVM prefers comparisons with zero over non-zero in general, but in this
case it choses instead to keep the max operation obvious.

//===---------------------------------------------------------------------===//

Switch lowering generates less than ideal code for the following switch:
define void @a(i32 %x) nounwind {
entry:
  switch i32 %x, label %if.end [
    i32 0, label %if.then
    i32 1, label %if.then
    i32 2, label %if.then
    i32 3, label %if.then
    i32 5, label %if.then
  ]
if.then:
  tail call void @foo() nounwind
  ret void
if.end:
  ret void
}
declare void @foo()

Generated code on x86-64 (other platforms give similar results):
a:
	cmpl	$5, %edi
	ja	.LBB0_2
	movl	%edi, %eax
	movl	$47, %ecx
	btq	%rax, %rcx
	jb	.LBB0_3
.LBB0_2:
	ret
.LBB0_3:
	jmp	foo  # TAILCALL

The movl+movl+btq+jb could be simplified to a cmpl+jne.

Or, if we wanted to be really clever, we could simplify the whole thing to
something like the following, which eliminates a branch:
	xorl    $1, %edi
	cmpl	$4, %edi
	ja	.LBB0_2
	ret
.LBB0_2:
	jmp	foo  # TAILCALL

//===---------------------------------------------------------------------===//

We compile this:

int foo(int a) { return (a & (~15)) / 16; }

Into:

define i32 @foo(i32 %a) nounwind readnone ssp {
entry:
  %and = and i32 %a, -16
  %div = sdiv i32 %and, 16
  ret i32 %div
}

but this code (X & -A)/A is X >> log2(A) when A is a power of 2, so this case
should be instcombined into just "a >> 4".

We do get this at the codegen level, so something knows about it, but 
instcombine should catch it earlier:

_foo:                                   ## @foo
## BB#0:                                ## %entry
	movl	%edi, %eax
	sarl	$4, %eax
	ret

//===---------------------------------------------------------------------===//

This code (from GCC PR28685):

int test(int a, int b) {
  int lt = a < b;
  int eq = a == b;
  if (lt)
    return 1;
  return eq;
}

Is compiled to:

define i32 @test(i32 %a, i32 %b) nounwind readnone ssp {
entry:
  %cmp = icmp slt i32 %a, %b
  br i1 %cmp, label %return, label %if.end

if.end:                                           ; preds = %entry
  %cmp5 = icmp eq i32 %a, %b
  %conv6 = zext i1 %cmp5 to i32
  ret i32 %conv6

return:                                           ; preds = %entry
  ret i32 1
}

it could be:

define i32 @test__(i32 %a, i32 %b) nounwind readnone ssp {
entry:
  %0 = icmp sle i32 %a, %b
  %retval = zext i1 %0 to i32
  ret i32 %retval
}

//===---------------------------------------------------------------------===//

This code can be seen in viterbi:

  %64 = call noalias i8* @malloc(i64 %62) nounwind
...
  %67 = call i64 @llvm.objectsize.i64(i8* %64, i1 false) nounwind
  %68 = call i8* @__memset_chk(i8* %64, i32 0, i64 %62, i64 %67) nounwind

llvm.objectsize.i64 should be taught about malloc/calloc, allowing it to
fold to %62.  This is a security win (overflows of malloc will get caught)
and also a performance win by exposing more memsets to the optimizer.

This occurs several times in viterbi.

Note that this would change the semantics of @llvm.objectsize which by its
current definition always folds to a constant. We also should make sure that
we remove checking in code like

  char *p = malloc(strlen(s)+1);
  __strcpy_chk(p, s, __builtin_objectsize(p, 0));

//===---------------------------------------------------------------------===//

This code (from Benchmarks/Dhrystone/dry.c):

define i32 @Func1(i32, i32) nounwind readnone optsize ssp {
entry:
  %sext = shl i32 %0, 24
  %conv = ashr i32 %sext, 24
  %sext6 = shl i32 %1, 24
  %conv4 = ashr i32 %sext6, 24
  %cmp = icmp eq i32 %conv, %conv4
  %. = select i1 %cmp, i32 10000, i32 0
  ret i32 %.
}

Should be simplified into something like:

define i32 @Func1(i32, i32) nounwind readnone optsize ssp {
entry:
  %sext = shl i32 %0, 24
  %conv = and i32 %sext, 0xFF000000
  %sext6 = shl i32 %1, 24
  %conv4 = and i32 %sext6, 0xFF000000
  %cmp = icmp eq i32 %conv, %conv4
  %. = select i1 %cmp, i32 10000, i32 0
  ret i32 %.
}

and then to:

define i32 @Func1(i32, i32) nounwind readnone optsize ssp {
entry:
  %conv = and i32 %0, 0xFF
  %conv4 = and i32 %1, 0xFF
  %cmp = icmp eq i32 %conv, %conv4
  %. = select i1 %cmp, i32 10000, i32 0
  ret i32 %.
}
//===---------------------------------------------------------------------===//

clang -O3 currently compiles this code

int g(unsigned int a) {
  unsigned int c[100];
  c[10] = a;
  c[11] = a;
  unsigned int b = c[10] + c[11];
  if(b > a*2) a = 4;
  else a = 8;
  return a + 7;
}

into

define i32 @g(i32 a) nounwind readnone {
  %add = shl i32 %a, 1
  %mul = shl i32 %a, 1
  %cmp = icmp ugt i32 %add, %mul
  %a.addr.0 = select i1 %cmp, i32 11, i32 15
  ret i32 %a.addr.0
}

The icmp should fold to false. This CSE opportunity is only available
after GVN and InstCombine have run.

//===---------------------------------------------------------------------===//

memcpyopt should turn this:

define i8* @test10(i32 %x) {
  %alloc = call noalias i8* @malloc(i32 %x) nounwind
  call void @llvm.memset.p0i8.i32(i8* %alloc, i8 0, i32 %x, i32 1, i1 false)
  ret i8* %alloc
}

into a call to calloc.  We should make sure that we analyze calloc as
aggressively as malloc though.

//===---------------------------------------------------------------------===//

clang -O3 doesn't optimize this:

void f1(int* begin, int* end) {
  std::fill(begin, end, 0);
}

into a memset.  This is PR8942.

//===---------------------------------------------------------------------===//

clang -O3 -fno-exceptions currently compiles this code:

void f(int N) {
  std::vector<int> v(N);

  extern void sink(void*); sink(&v);
}

into

define void @_Z1fi(i32 %N) nounwind {
entry:
  %v2 = alloca [3 x i32*], align 8
  %v2.sub = getelementptr inbounds [3 x i32*]* %v2, i64 0, i64 0
  %tmpcast = bitcast [3 x i32*]* %v2 to %"class.std::vector"*
  %conv = sext i32 %N to i64
  store i32* null, i32** %v2.sub, align 8, !tbaa !0
  %tmp3.i.i.i.i.i = getelementptr inbounds [3 x i32*]* %v2, i64 0, i64 1