[ValueTracking] Improve pointer offset computation for cases of same base

This patch improves the handling of pointer offset in GEP expressions where
one argument is the base pointer. isPointerOffset() is being used by memcpyopt
where current code synthesizes consecutive 32 bytes stores to one store and
two memset intrinsic calls. With this patch, we convert the stores to one
memset intrinsic.

Differential Revision: https://reviews.llvm.org/D67989

llvm-svn: 374454

2 files changed, 116 insertions, 9 deletions

diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index f0b87810ef9..aaebba00527 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -5755,17 +5755,47 @@ Optional<int64_t> llvm::isPointerOffset(const Value *Ptr1, const Value *Ptr2,
   const GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1);
   const GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2);
 
-  // If one pointer is a GEP and the other isn't, then see if the GEP is a
-  // constant offset from the base, as in "P" and "gep P, 1".
-  if (GEP1 && !GEP2 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) {
-    auto Offset = getOffsetFromIndex(GEP1, 1, DL);
-    if (!Offset)
+  // If one pointer is a GEP see if the GEP is a constant offset from the base,
+  // as in "P" and "gep P, 1".
+  // Also do this iteratively to handle the the following case:
+  //   Ptr_t1 = GEP Ptr1, c1
+  //   Ptr_t2 = GEP Ptr_t1, c2
+  //   Ptr2 = GEP Ptr_t2, c3
+  // where we will return c1+c2+c3.
+  // TODO: Handle the case when both Ptr1 and Ptr2 are GEPs of some common base
+  // -- replace getOffsetFromBase with getOffsetAndBase, check that the bases
+  // are the same, and return the difference between offsets.
+  auto getOffsetFromBase = [&DL](const GEPOperator *GEP,
+                                 const Value *Ptr) -> Optional<int64_t> {
+    const GEPOperator *GEP_T = GEP;
+    int64_t OffsetVal = 0;
+    bool HasSameBase = false;
+    while (GEP_T) {
+      auto Offset = getOffsetFromIndex(GEP_T, 1, DL);
+      if (!Offset)
+        return None;
+      OffsetVal += *Offset;
+      auto Op0 = GEP_T->getOperand(0)->stripPointerCasts();
+      if (Op0 == Ptr) {
+        HasSameBase = true;
+        break;
+      }
+      GEP_T = dyn_cast<GEPOperator>(Op0);
+    }
+    if (!HasSameBase)
       return None;
-    return -*Offset;
-  }
+    return OffsetVal;
+  };
 
-  if (GEP2 && !GEP1 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) {
-    return getOffsetFromIndex(GEP2, 1, DL);
+  if (GEP1) {
+    auto Offset = getOffsetFromBase(GEP1, Ptr2);
+    if (Offset)
+      return -*Offset;
+  }
+  if (GEP2) {
+    auto Offset = getOffsetFromBase(GEP2, Ptr1);
+    if (Offset)
+      return Offset;
   }
 
   // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
diff --git a/llvm/test/Transforms/MemCpyOpt/store-to-memset.ll b/llvm/test/Transforms/MemCpyOpt/store-to-memset.ll
new file mode 100644
index 00000000000..59ed892b60e
--- /dev/null
+++ b/llvm/test/Transforms/MemCpyOpt/store-to-memset.ll
@@ -0,0 +1,77 @@
+; RUN: opt < %s -memcpyopt -S | FileCheck %s
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-grtev4-linux-gnu"
+
+define i8* @foo(i8* returned %0, i32 %1, i64 %2) {
+entry:
+  %3 = getelementptr inbounds i8, i8* %0, i64 %2
+  %4 = getelementptr inbounds i8, i8* %3, i64 -32
+  %vv = trunc i32 %1 to i8
+  store i8 %vv, i8* %4, align 1
+  %5 = getelementptr inbounds i8, i8* %4, i64 1
+  store i8 %vv, i8* %5, align 1
+  %6= getelementptr inbounds i8, i8* %4, i64 2
+  store i8 %vv, i8* %6, align 1
+  %7= getelementptr inbounds i8, i8* %4, i64 3
+  store i8 %vv, i8* %7, align 1
+  %8= getelementptr inbounds i8, i8* %4, i64 4
+  store i8 %vv, i8* %8, align 1
+  %9= getelementptr inbounds i8, i8* %4, i64 5
+  store i8 %vv, i8* %9, align 1
+  %10= getelementptr inbounds i8, i8* %4, i64 6
+  store i8 %vv, i8* %10, align 1
+  %11= getelementptr inbounds i8, i8* %4, i64 7
+  store i8 %vv, i8* %11, align 1
+  %12= getelementptr inbounds i8, i8* %4, i64 8
+  store i8 %vv, i8* %12, align 1
+  %13= getelementptr inbounds i8, i8* %4, i64 9
+  store i8 %vv, i8* %13, align 1
+  %14= getelementptr inbounds i8, i8* %4, i64 10
+  store i8 %vv, i8* %14, align 1
+  %15= getelementptr inbounds i8, i8* %4, i64 11
+  store i8 %vv, i8* %15, align 1
+  %16= getelementptr inbounds i8, i8* %4, i64 12
+  store i8 %vv, i8* %16, align 1
+  %17= getelementptr inbounds i8, i8* %4, i64 13
+  store i8 %vv, i8* %17, align 1
+  %18= getelementptr inbounds i8, i8* %4, i64 14
+  store i8 %vv, i8* %18, align 1
+  %19= getelementptr inbounds i8, i8* %4, i64 15
+  store i8 %vv, i8* %19, align 1
+  %20= getelementptr inbounds i8, i8* %4, i64 16
+  store i8 %vv, i8* %20, align 1
+  %21= getelementptr inbounds i8, i8* %20, i64 1
+  store i8 %vv, i8* %21, align 1
+  %22= getelementptr inbounds i8, i8* %20, i64 2
+  store i8 %vv, i8* %22, align 1
+  %23= getelementptr inbounds i8, i8* %20, i64 3
+  store i8 %vv, i8* %23, align 1
+  %24= getelementptr inbounds i8, i8* %20, i64 4
+  store i8 %vv, i8* %24, align 1
+  %25= getelementptr inbounds i8, i8* %20, i64 5
+  store i8 %vv, i8* %25, align 1
+  %26= getelementptr inbounds i8, i8* %20, i64 6
+  store i8 %vv, i8* %26, align 1
+  %27= getelementptr inbounds i8, i8* %20, i64 7
+  store i8 %vv, i8* %27, align 1
+  %28= getelementptr inbounds i8, i8* %20, i64 8
+  store i8 %vv, i8* %28, align 1
+  %29= getelementptr inbounds i8, i8* %20, i64 9
+  store i8 %vv, i8* %29, align 1
+  %30= getelementptr inbounds i8, i8* %20, i64 10
+  store i8 %vv, i8* %30, align 1
+  %31 = getelementptr inbounds i8, i8* %20, i64 11
+  store i8 %vv, i8* %31, align 1
+  %32 = getelementptr inbounds i8, i8* %20, i64 12
+  store i8 %vv, i8* %32, align 1
+  %33 = getelementptr inbounds i8, i8* %20, i64 13
+  store i8 %vv, i8* %33, align 1
+  %34 = getelementptr inbounds i8, i8* %20, i64 14
+  store i8 %vv, i8* %34, align 1
+  %35 = getelementptr inbounds i8, i8* %20, i64 15
+  store i8 %vv, i8* %35, align 1
+  ret i8* %0
+; CHECK-LABEL: @foo
+; CHECK: call void @llvm.memset.p0i8.i64(i8* align 1 %4, i8 %vv, i64 32, i1 false)
+}
+