llvm-svn: 1128

llvm-svn: 1127

llvm-svn: 1125

llvm-svn: 1124

llvm-svn: 1123

* fix minor bug with taking datasize of unsized array type
* Insert code to support speculative changes later

llvm-svn: 1122

Convert backward conversion of expression trees into a new more powerful bidirectional approach.  Add transforms for more instructions.

llvm-svn: 1121

Fixed selection to create a TmpInstruction for each integer CC register
(since it is an implicit side-effect, unlike FP CC registers which are
explicit operands).

llvm-svn: 1120

llvm-svn: 1119

llvm-svn: 1118

llvm-svn: 1117

IGNode::pushOnStack().

llvm-svn: 1116

llvm-svn: 1115

//  1. PHI nodes with multiple entries for the same predecessor.  GCC sometimes
//     generates code that looks like this:
//
//  bb7:  br bool %cond1004, label %bb8, label %bb8
//  bb8: %reg119 = phi uint [ 0, %bb7 ], [ 1, %bb7 ]
//
//     which is completely illegal LLVM code.  To compensate for this, we insert
//     an extra basic block, and convert the code to look like this:
//
//  bb7: br bool %cond1004, label %bbX, label %bb8
//  bbX: br label bb8
//  bb8: %reg119 = phi uint [ 0, %bbX ], [ 1, %bb7 ]
//

llvm-svn: 1114

llvm-svn: 1112

llvm-svn: 1111

* Add code to fix PHI nodes that are missing arguments.  These PHI nodes can
  be generated by GCC if there is an uninitialized variable flowing into a merge
  point.  For example:

int foo(int y) {
  int X;
  if (y) X = 1;
  return X;
}

llvm-svn: 1110

llvm-svn: 1108

Moved InsertCallerSaveInstr to the SparcRegInfo.cpp

llvm-svn: 1106

CahedCopyMap was disabled to insert phi elimination code for all phi args

llvm-svn: 1105

llvm-svn: 1104

llvm-svn: 1103

This resolves definitions like this:
   %list * "foo"(...)
   %list * "foo"(int)

together which can often occur because C programmers don't put prototypes in like they should.  GRR

llvm-svn: 1102

llvm-svn: 1094

llvm-svn: 1093

llvm-svn: 1092

llvm-svn: 1091

Why does this stuff have to come up NOW?

llvm-svn: 1089

llvm-svn: 1088

Support floating point numbers in expodential form so that small numbers don't get truncated, which broke the health benchmark

llvm-svn: 1087

* Implement new peephole:
     // Peephole optimize the following instructions:
     // %t1 = cast {<...>} * %StructPtr to <ty> *
     //
     // Into: %t2 = getelementptr {<...>} * %StructPtr, <0, 0, 0, ...>
     //       %t1 = cast <eltype> * %t1 to <ty> *

  This peephole eliminated 9 evil casts in the health benchmark, and
  completely turned the addList method around.  :)

llvm-svn: 1085

llvm-svn: 1084

Refactor code to support it

llvm-svn: 1083

llvm-svn: 1081

* Fix a problem setting a name on a constant value that died because no symbol table was passed in
* Add some comments describing the passes
* Implement a new peephole:
     // Peephole optimize the following instructions:
     // %t = cast <T1>* %P to <T2> * ;; If T1 is losslessly convertable to T2
     // store <T2> %V, <T2>* %t
     //
     // Into:
     // %t = cast <T2> %V to <T1>
     // store <T1> %t2, <T1>* %P

llvm-svn: 1080

llvm-svn: 1079

llvm-svn: 1078

* Convert getelementptr/load pairs into a single load

llvm-svn: 1075

llvm-svn: 1074

llvm-svn: 1071