From d9d4889ccc3e9a753142fe16ff298b5c707d0872 Mon Sep 17 00:00:00 2001
From: Dan Gohman <gohman@apple.com>
Date: Wed, 10 Mar 2010 19:31:51 +0000
Subject: [PATCH] Constant-fold GEP-of-GEP into a single GEP.

llvm-svn: 98178
---
 lib/Analysis/ConstantFolding.cpp | 11 +++++++++++
 test/Other/constant-fold-gep.ll  | 31 +++++++++++++++++++++++++++++++
 2 files changed, 42 insertions(+)

diff --git a/lib/Analysis/ConstantFolding.cpp b/lib/Analysis/ConstantFolding.cpp
index 114db2d3705..62020afb723 100644
--- a/lib/Analysis/ConstantFolding.cpp
+++ b/lib/Analysis/ConstantFolding.cpp
@@ -589,6 +589,17 @@ static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps,
   APInt Offset = APInt(BitWidth,
                        TD->getIndexedOffset(Ptr->getType(),
                                             (Value**)Ops+1, NumOps-1));
+
+  // If this is a GEP of a GEP, fold it all into a single GEP.
+  while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
+    SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
+    Ptr = cast<Constant>(GEP->getOperand(0));
+    Offset += APInt(BitWidth,
+                    TD->getIndexedOffset(Ptr->getType(),
+                                         (Value**)NestedOps.data(),
+                                         NestedOps.size()));
+  }
+
   // If the base value for this address is a literal integer value, fold the
   // getelementptr to the resulting integer value casted to the pointer type.
   if (BaseIsInt) {
diff --git a/test/Other/constant-fold-gep.ll b/test/Other/constant-fold-gep.ll
index 650b4af139e..ecef9c48492 100644
--- a/test/Other/constant-fold-gep.ll
+++ b/test/Other/constant-fold-gep.ll
@@ -21,10 +21,12 @@
 ; PLAIN: %1 = type { double, float, double, double }
 ; PLAIN: %2 = type { i1, i1* }
 ; PLAIN: %3 = type { i64, i64 }
+; PLAIN: %4 = type { i32, i32 }
 ; OPT: %0 = type { i1, double }
 ; OPT: %1 = type { double, float, double, double }
 ; OPT: %2 = type { i1, i1* }
 ; OPT: %3 = type { i64, i64 }
+; OPT: %4 = type { i32, i32 }
 
 ; The automatic constant folder in opt does not have targetdata access, so
 ; it can't fold gep arithmetic, in general. However, the constant folder run
@@ -122,6 +124,16 @@
 @N = constant i64* getelementptr ({ i64, i64 }* null, i32 0, i32 1)
 @O = constant i64* getelementptr ([2 x i64]* null, i32 0, i32 1)
 
+; Fold GEP of a GEP. Theoretically some of these cases could be folded
+; without using targetdata, however that's not implemented yet.
+
+; PLAIN: @Z = global i32* getelementptr inbounds (i32* getelementptr inbounds ([3 x %4]* @ext, i64 0, i64 1, i32 0), i64 1)
+; OPT: @Z = global i32* getelementptr (i32* getelementptr inbounds ([3 x %4]* @ext, i64 0, i64 1, i32 0), i64 1)
+; TO: @Z = global i32* getelementptr inbounds ([3 x %0]* @ext, i64 0, i64 1, i32 1)
+
+@ext = external global [3 x { i32, i32 }]
+@Z = global i32* getelementptr inbounds (i32* getelementptr inbounds ([3 x { i32, i32 }]* @ext, i64 0, i64 1, i32 0), i64 1)
+
 ; Duplicate all of the above as function return values rather than
 ; global initializers.
 
@@ -468,3 +480,22 @@ define i64* @fO() nounwind {
   %t = bitcast i64* getelementptr ([2 x i64]* null, i32 0, i32 1) to i64*
   ret i64* %t
 }
+
+; PLAIN: define i32* @fZ() nounwind {
+; PLAIN:   %t = bitcast i32* getelementptr inbounds (i32* getelementptr inbounds ([3 x %4]* @ext, i64 0, i64 1, i32 0), i64 1) to i32*
+; PLAIN:   ret i32* %t
+; PLAIN: }
+; OPT: define i32* @fZ() nounwind {
+; OPT:   ret i32* getelementptr inbounds (i32* getelementptr inbounds ([3 x %4]* @ext, i64 0, i64 1, i32 0), i64 1)
+; OPT: }
+; TO: define i32* @fZ() nounwind {
+; TO:   ret i32* getelementptr inbounds ([3 x %0]* @ext, i64 0, i64 1, i32 1)
+; TO: }
+; SCEV: Classifying expressions for: @fZ
+; SCEV:   %t = bitcast i32* getelementptr inbounds (i32* getelementptr inbounds ([3 x %4]* @ext, i64 0, i64 1, i32 0), i64 1) to i32*
+; SCEV:   -->  ((3 * sizeof(i32)) + @ext)
+
+define i32* @fZ() nounwind {
+  %t = bitcast i32* getelementptr inbounds (i32* getelementptr inbounds ([3 x { i32, i32 }]* @ext, i64 0, i64 1, i32 0), i64 1) to i32*
+  ret i32* %t
+}