InstCombine: If we call llvm.objectsize on a malloc call we can replace it with the...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122959 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/README.txt b/lib/Target/README.txt
index 8ed8c598ddeb203a06dd4b450e67e20776a2e8db..d83cd8bb963cf4ae46ebe29087324901baf4d21b 100644 (file)
--- a/lib/Target/README.txt
+++ b/lib/Target/README.txt
@@ -2020,29 +2020,6 @@ define i1 @g(i32 a) nounwind readnone {
 
 //===---------------------------------------------------------------------===//
 
-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.
-
-Stuff like this occurs in drystone:
-
-  %call5 = call i8* @malloc(i32 48) optsize
-  %5 = getelementptr inbounds i8* %call5, i32 16
-  %6 = call i32 @llvm.objectsize.i32(i8* %5, i1 false)
-
-We should be able to constant fold that.
-
-//===---------------------------------------------------------------------===//
-
 This code (from Benchmarks/Dhrystone/dry.c):
 
 define i32 @Func1(i32, i32) nounwind readnone optsize ssp {

diff --git a/lib/Transforms/InstCombine/InstCombineCalls.cpp b/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 7068cb6669e834605bd4a5b9a1226e07bb8606d9..603a01075a54bc91e5592783315576b06944d7af 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -298,12 +298,16 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
         }
       }
     } else if (CallInst *MI = extractMallocCall(Op1)) {
-      // Get alloca size.
+      // Get allocation size.
       const Type* MallocType = getMallocAllocatedType(MI);
       if (MallocType && MallocType->isSized())
         if (Value *NElems = getMallocArraySize(MI, TD, true))
           if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
             Size = NElements->getZExtValue() * TD->getTypeAllocSize(MallocType);
+
+      // If there is no offset we can just return the size passed to malloc.
+      if (Offset == 0)
+        return ReplaceInstUsesWith(CI, MI->getArgOperand(0));
     }
 
     // Do not return "I don't know" here. Later optimization passes could

diff --git a/test/Transforms/InstCombine/objsize.ll b/test/Transforms/InstCombine/objsize.ll
index 043525b755566f8cf0d8a13e68022deb6c4cc21d..9950e5f4cc62836e266469b19c64d48f32fb679a 100644 (file)
--- a/test/Transforms/InstCombine/objsize.ll
+++ b/test/Transforms/InstCombine/objsize.ll
@@ -160,3 +160,19 @@ define i32 @test7() {
   ret i32 %objsize
 }
 
+define i32 @test8(i32 %x) {
+; CHECK: @test8
+  %alloc = call noalias i8* @malloc(i32 %x) nounwind
+  %objsize = call i32 @llvm.objectsize.i32(i8* %alloc, i1 false) nounwind readonly
+; CHECK-NEXT: ret i32 %x
+  ret i32 %objsize
+}
+
+define i32 @test9(i32 %x) {
+; CHECK: @test9
+  %alloc = call noalias i8* @malloc(i32 %x) nounwind
+  %gep = getelementptr inbounds i8* %alloc, i32 16
+  %objsize = call i32 @llvm.objectsize.i32(i8* %gep, i1 false) nounwind readonly
+; CHECK-NOT: ret i32 %x
+  ret i32 %objsize
+}