README.txt

The expression should optimize to something like
"!((start|end)&~PMD_MASK). Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

From GCC Bug 15241:
unsigned int
foo (unsigned int a, unsigned int b)
{
 if (a <= 7 && b <= 7)
   baz ();
}
Should combine to "(a|b) <= 7".  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

From GCC Bug 3756:
int
pn (int n)
{
 return (n >= 0 ? 1 : -1);
}
Should combine to (n >> 31) | 1.  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts | llc".

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

From GCC Bug 28685:
int test(int a, int b)
{
 int lt = a < b;
 int eq = a == b;

 return (lt || eq);
}
Should combine to "a <= b".  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts | llc".

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

void a(int variable)
{
 if (variable == 4 || variable == 6)
   bar();
}
This should optimize to "if ((variable | 2) == 6)".  Currently not
optimized with "clang -emit-llvm-bc | opt -std-compile-opts | llc".

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

unsigned int f(unsigned int i, unsigned int n) {++i; if (i == n) ++i; return
i;}
unsigned int f2(unsigned int i, unsigned int n) {++i; i += i == n; return i;}
These should combine to the same thing.  Currently, the first function
produces better code on X86.

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

From GCC Bug 15784:
#define abs(x) x>0?x:-x
int f(int x, int y)
{
 return (abs(x)) >= 0;
}
This should optimize to x == INT_MIN. (With -fwrapv.)  Currently not
optimized with "clang -emit-llvm-bc | opt -std-compile-opts".

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

From GCC Bug 14753:
void
rotate_cst (unsigned int a)
{
 a = (a << 10) | (a >> 22);
 if (a == 123)
   bar ();
}
void
minus_cst (unsigned int a)
{
 unsigned int tem;

 tem = 20 - a;
 if (tem == 5)
   bar ();
}
void
mask_gt (unsigned int a)
{
 /* This is equivalent to a > 15.  */
 if ((a & ~7) > 8)
   bar ();
}
void
rshift_gt (unsigned int a)
{
 /* This is equivalent to a > 23.  */
 if ((a >> 2) > 5)
   bar ();
}
All should simplify to a single comparison.  All of these are
currently not optimized with "clang -emit-llvm-bc | opt
-std-compile-opts".

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

From GCC Bug 32605:
int c(int* x) {return (char*)x+2 == (char*)x;}
Should combine to 0.  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts" (although llc can optimize it).

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

int a(unsigned char* b) {return *b > 99;}
There's an unnecessary zext in the generated code with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

int a(unsigned b) {return ((b << 31) | (b << 30)) >> 31;}
Should be combined to  "((b >> 1) | b) & 1".  Currently not optimized
with "clang -emit-llvm-bc | opt -std-compile-opts".

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

unsigned a(unsigned x, unsigned y) { return x | (y & 1) | (y & 2);}
Should combine to "x | (y & 3)".  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

unsigned a(unsigned a) {return ((a | 1) & 3) | (a & -4);}
Should combine to "a | 1".  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

int a(int a, int b, int c) {return (~a & c) | ((c|a) & b);}
Should fold to "(~a & c) | (a & b)".  Currently not optimized with
"clang -emit-llvm-bc | opt -std-compile-opts".

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

int a(int a,int b) {return (~(a|b))|a;}
Should fold to "a|~b".  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

int a(int a, int b) {return (a&&b) || (a&&!b);}
Should fold to "a".  Currently not optimized with "clang -emit-llvm-bc
| opt -std-compile-opts".

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

int a(int a, int b, int c) {return (a&&b) || (!a&&c);}
Should fold to "a ? b : c", or at least something sane.  Currently not
optimized with "clang -emit-llvm-bc | opt -std-compile-opts".

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

int a(int a, int b, int c) {return (a&&b) || (a&&c) || (a&&b&&c);}
Should fold to a && (b || c).  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

int a(int x) {return x | ((x & 8) ^ 8);}
Should combine to x | 8.  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

int a(int x) {return x ^ ((x & 8) ^ 8);}
Should also combine to x | 8.  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

int a(int x) {return (x & 8) == 0 ? -1 : -9;}
Should combine to (x | -9) ^ 8.  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

int a(int x) {return (x & 8) == 0 ? -9 : -1;}
Should combine to x | -9.  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

int a(int x) {return ((x | -9) ^ 8) & x;}
Should combine to x & -9.  Currently not optimized with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

unsigned a(unsigned a) {return a * 0x11111111 >> 28 & 1;}
Should combine to "a * 0x88888888 >> 31".  Currently not optimized
with "clang -emit-llvm-bc | opt -std-compile-opts".

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

unsigned a(char* x) {if ((*x & 32) == 0) return b();}
There's an unnecessary zext in the generated code with "clang
-emit-llvm-bc | opt -std-compile-opts".

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

unsigned a(unsigned long long x) {return 40 * (x >> 1);}
Should combine to "20 * (((unsigned)x) & -2)".  Currently not
optimized with "clang -emit-llvm-bc | opt -std-compile-opts".

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

We would like to do the following transform in the instcombiner:

  -X/C -> X/-C

However, this isn't valid if (-X) overflows. We can implement this when we
have the concept of a "C signed subtraction" operator that which is undefined
on overflow.

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

This was noticed in the entryblock for grokdeclarator in 403.gcc:

        %tmp = icmp eq i32 %decl_context, 4          
        %decl_context_addr.0 = select i1 %tmp, i32 3, i32 %decl_context 
        %tmp1 = icmp eq i32 %decl_context_addr.0, 1 
        %decl_context_addr.1 = select i1 %tmp1, i32 0, i32 %decl_context_addr.0

tmp1 should be simplified to something like:
  (!tmp || decl_context == 1)

This allows recursive simplifications, tmp1 is used all over the place in
the function, e.g. by:

        %tmp23 = icmp eq i32 %decl_context_addr.1, 0            ; <i1> [#uses=1]
        %tmp24 = xor i1 %tmp1, true             ; <i1> [#uses=1]
        %or.cond8 = and i1 %tmp23, %tmp24               ; <i1> [#uses=1]

later.

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

Store sinking: This code:

void f (int n, int *cond, int *res) {
    int i;
    *res = 0;
    for (i = 0; i < n; i++)
        if (*cond)
            *res ^= 234; /* (*) */
}

On this function GVN hoists the fully redundant value of *res, but nothing
moves the store out.  This gives us this code:

bb:		; preds = %bb2, %entry
	%.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.

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

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.

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

[PHI TRANSLATE GEPs]

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 fully redundant, an in BB2 it should have the value %8.

GCC PR33344 is a similar case.

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

There are many load PRE testcases in testsuite/gcc.dg/tree-ssa/loadpre* in the
GCC testsuite.  There are many pre testcases as ssa-pre-*.c

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

There are some interesting cases in testsuite/gcc.dg/tree-ssa/pred-comm* in the
GCC testsuite.  For example, predcom-1.c is:

 for (i = 2; i < 1000; i++)
    fib[i] = (fib[i-1] + fib[i - 2]) & 0xffff;

which compiles into:

bb1:		; preds = %bb1, %bb1.thread
	%indvar = phi i32 [ 0, %bb1.thread ], [ %0, %bb1 ]	
	%i.0.reg2mem.0 = add i32 %indvar, 2		
	%0 = add i32 %indvar, 1		; <i32> [#uses=3]
	%1 = getelementptr [1000 x i32]* @fib, i32 0, i32 %0		
	%2 = load i32* %1, align 4		; <i32> [#uses=1]
	%3 = getelementptr [1000 x i32]* @fib, i32 0, i32 %indvar	
	%4 = load i32* %3, align 4		; <i32> [#uses=1]
	%5 = add i32 %4, %2		; <i32> [#uses=1]
	%6 = and i32 %5, 65535		; <i32> [#uses=1]
	%7 = getelementptr [1000 x i32]* @fib, i32 0, i32 %i.0.reg2mem.0
	store i32 %6, i32* %7, align 4
	%exitcond = icmp eq i32 %0, 998		; <i1> [#uses=1]
	br i1 %exitcond, label %return, label %bb1

This is basically:
  LOAD fib[i+1]
  LOAD fib[i]
  STORE fib[i+2]

instead of handling this as a loop or other xform, all we'd need to do is teach
load PRE to phi translate the %0 add (i+1) into the predecessor as (i'+1+1) =
(i'+2) (where i' is the previous iteration of i).  This would find the store
which feeds it.

predcom-2.c is apparently the same as predcom-1.c
predcom-3.c is very similar but needs loads feeding each other instead of
store->load.
predcom-4.c seems the same as the rest.


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

Other simple load PRE cases:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35287 [LPRE crit edge splitting]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34677 (licm does this, LPRE crit edge)
  llvm-gcc t2.c -S -o - -O0 -emit-llvm | llvm-as | opt -mem2reg -simplifycfg -gvn | llvm-dis

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

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

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

When GVN/PRE finds a store of float* to a must aliases pointer when expecting
an int*, it should turn it into a bitcast.  This is a nice generalization of
the SROA hack that would apply to other cases, e.g.:

int foo(int C, int *P, float X) {
  if (C) {
    bar();
    *P = 42;
  } else
    *(float*)P = X;

   return *P;
}


One example (that requires crazy phi translation) is:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16799 [BITCAST PHI TRANS]

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

A/B get pinned to the stack because we turn an if/then into a select instead
of PRE'ing the load/store.  This may be fixable in instcombine:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37892



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) GEP PHI TRAN, 2) CRIT EDGE 3) 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, "") -> strlen(x)
strcspn("", x) -> 0
strspn("", x) -> 0
strspn(x, "") -> strlen(x)
strspn(x, "a") -> strchr(x, 'a')-x

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.

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

"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:

        %3073 = call i8* @strcpy(i8* %3072, i8* %3071) nounwind
        %strlen = call i32 @strlen(i8* %3072)  

The strlen could be replaced with: %strlen = sub %3072, %3073, because the
strcpy call returns a pointer to the end of the string.  Based on that, the
endptr GEP just becomes equal to 3073, which eliminates a strlen call and GEP.

Second, 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 contains this interesting pattern:

%77 = call i8* @strstr(i8* getelementptr ([6 x i8]* @"\01LC5", i32 0, i32 0),
                       i8* %30)
%phitmp648 = icmp eq i8* %77, getelementptr ([6 x i8]* @"\01LC5", i32 0, i32 0)
br i1 %phitmp648, label %bb70, label %bb76

bb70:           ; preds = %OptionMatch.exit91, %bb69
        %78 = call i32 @strlen(i8* %30) nounwind readonly align 1               ; <i32> [#uses=1]

This is basically:
  cststr = "abcdef";
  if (strstr(cststr, P) == cststr) {
     x = strlen(P);
     ...

The strstr call would be significantly cheaper written as:

cststr = "abcdef";
if (memcmp(P, str, strlen(P)))
  x = strlen(P);

This is memcmp+strlen instead of strstr.  This also makes the strlen fully
redundant.

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

186.crafty also contains this code:

%1906 = call i32 @strlen(i8* getelementptr ([32 x i8]* @pgn_event, i32 0,i32 0))
%1907 = getelementptr [32 x i8]* @pgn_event, i32 0, i32 %1906
%1908 = call i8* @strcpy(i8* %1907, i8* %1905) nounwind align 1
%1909 = call i32 @strlen(i8* getelementptr ([32 x i8]* @pgn_event, i32 0,i32 0))
%1910 = getelementptr [32 x i8]* @pgn_event, i32 0, i32 %1909         

The last strlen is computable as 1908-@pgn_event, which means 1910=1908.

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

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.

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