Emit getelementptr inbounds nuw flag where possible

driver/cl_options.cpp

@@ -721,6 +721,11 @@ cl::opt<unsigned>
                    cl::desc("Warn for stack size bigger than the given number"),
                    cl::value_desc("threshold"));
 
+cl::opt<bool>
+    enableGetElementPtrNuw("enable-getelementptr-nuw", cl::ZeroOrMore,
+                           cl::desc("enable nuw(no-unsigned-wrap) flag to "
+                                    "LLVM's getelementptr insturction"));
+
 #if LDC_LLVM_SUPPORTED_TARGET_SPIRV || LDC_LLVM_SUPPORTED_TARGET_NVPTX
 cl::list<std::string>
     dcomputeTargets("mdcompute-targets", cl::CommaSeparated,

driver/cl_options.h

@@ -138,6 +138,8 @@ extern cl::opt<std::string> saveOptimizationRecord;
 
 extern cl::opt<unsigned> fWarnStackSize;
 
+extern cl::opt<bool> enableGetElementPtrNuw;
+
 #if LDC_LLVM_SUPPORTED_TARGET_SPIRV || LDC_LLVM_SUPPORTED_TARGET_NVPTX
 extern cl::list<std::string> dcomputeTargets;
 extern cl::opt<std::string> dcomputeFilePrefix;

gen/llvmhelpers.cpp

@@ -17,6 +17,7 @@
 #include "dmd/init.h"
 #include "dmd/module.h"
 #include "dmd/template.h"
+#include "driver/cl_options.h"
 #include "gen/abi/abi.h"
 #include "gen/arrays.h"
 #include "gen/classes.h"
@@ -1868,9 +1869,18 @@ DLValue *DtoIndexAggregate(LLValue *src, AggregateDeclaration *ad,
 
   LLValue *ptr = src;
   LLType * ty = nullptr;
+#if LDC_LLVM_VER >= 2000
+  llvm::GEPNoWrapFlags nw = llvm::GEPNoWrapFlags::inBounds();
+  if (opts::enableGetElementPtrNuw)
+    nw |= llvm::GEPNoWrapFlags::noUnsignedWrap();
+#endif
   if (!isFieldIdx) {
     // apply byte-wise offset from object start
-    ptr = DtoGEP1(getI8Type(), ptr, off);
+    ptr = DtoGEP1(getI8Type(), ptr, off
+#if LDC_LLVM_VER >= 2000
+      , "", nullptr, nw
+#endif
+    );
     ty = DtoType(vd->type);
   } else {
     if (ad->structsize == 0) { // can happen for extern(C) structs
@@ -1884,7 +1894,11 @@ DLValue *DtoIndexAggregate(LLValue *src, AggregateDeclaration *ad,
       } else {
         st = irTypeAggr->getLLType();
       }
-      ptr = DtoGEP(st, ptr, 0, off);
+      ptr = DtoGEP(st, ptr, 0, off
+#if LDC_LLVM_VER >= 2000
+      , "", nullptr, nw
+#endif
+      );
       ty = isaStruct(st)->getElementType(off);
     }
   }

gen/toir.cpp

@@ -22,6 +22,7 @@
 #include "dmd/root/rmem.h"
 #include "dmd/target.h"
 #include "dmd/template.h"
+#include "driver/cl_options.h"
 #include "gen/aa.h"
 #include "gen/abi/abi.h"
 #include "gen/arrays.h"
@@ -1198,7 +1199,16 @@ class ToElemVisitor : public Visitor {
       }
       LLType *elt = DtoMemType(e1type->nextOf());
       LLType *arrty = llvm::ArrayType::get(elt, e1type->isTypeSArray()->dim->isIntegerExp()->getInteger());
-      arrptr = DtoGEP(arrty, DtoLVal(l), DtoConstUint(0), DtoRVal(r));
+#if LDC_LLVM_VER >= 2000
+      llvm::GEPNoWrapFlags nw = llvm::GEPNoWrapFlags::inBounds();
+      if (e->indexIsInBounds && opts::enableGetElementPtrNuw)
+        nw |= llvm::GEPNoWrapFlags::noUnsignedWrap();
+#endif
+      arrptr = DtoGEP(arrty, DtoLVal(l), DtoConstUint(0), DtoRVal(r)
+#if LDC_LLVM_VER >= 2000
+                    , "", nullptr, nw
+#endif
+      );
     } else if (e1type->ty == TY::Tarray) {
       if (p->emitArrayBoundsChecks() && !e->indexIsInBounds) {
         DtoIndexBoundsCheck(e->loc, l, r);
@@ -1292,8 +1302,16 @@ class ToElemVisitor : public Visitor {
       }
 
       // offset by lower
-      eptr = DtoGEP1(DtoMemType(etype->nextOf()), getBasePointer(), vlo, "lowerbound");
-
+#if LDC_LLVM_VER >= 2000
+      llvm::GEPNoWrapFlags nw = llvm::GEPNoWrapFlags::inBounds();
+      if (!needCheckUpper && !needCheckLower && opts::enableGetElementPtrNuw)
+        nw |= llvm::GEPNoWrapFlags::noUnsignedWrap();
+#endif
+      eptr = DtoGEP1(DtoMemType(etype->nextOf()), getBasePointer(), vlo, "lowerbound"
+#if LDC_LLVM_VER >= 2000
+                   , nullptr, nw
+#endif
+      );
       // adjust length
       elen = p->ir->CreateSub(vup, vlo);
     }

gen/tollvm.cpp

@@ -357,46 +357,103 @@ LLIntegerType *DtoSize_t() {
 namespace {
 llvm::GetElementPtrInst *DtoGEP(LLType *pointeeTy, LLValue *ptr,
                                 llvm::ArrayRef<LLValue *> indices,
-                                const char *name, llvm::BasicBlock *bb) {
+                                const char *name, llvm::BasicBlock *bb
+#if LDC_LLVM_VER >= 2000
+                              , llvm::GEPNoWrapFlags nw
+#endif
+                              ) {
   auto gep = llvm::GetElementPtrInst::Create(pointeeTy, ptr, indices, name,
                                              bb ? bb : gIR->scopebb());
+#if LDC_LLVM_VER >= 2000
+  gep->setNoWrapFlags(nw);
+#else
   gep->setIsInBounds(true);
+#endif
   return gep;
 }
 }
 
 LLValue *DtoGEP1(LLType *pointeeTy, LLValue *ptr, LLValue *i0, const char *name,
-                 llvm::BasicBlock *bb) {
-  return DtoGEP(pointeeTy, ptr, i0, name, bb);
+                 llvm::BasicBlock *bb
+#if LDC_LLVM_VER >= 2000
+               , llvm::GEPNoWrapFlags nw
+#endif
+                ) {
+  return DtoGEP(pointeeTy, ptr, i0, name, bb
+#if LDC_LLVM_VER >= 2000
+              , nw
+#endif
+  );
 }
 
 LLValue *DtoGEP(LLType *pointeeTy, LLValue *ptr, LLValue *i0, LLValue *i1,
-                const char *name, llvm::BasicBlock *bb) {
+                const char *name, llvm::BasicBlock *bb
+#if LDC_LLVM_VER >= 2000
+              , llvm::GEPNoWrapFlags nw
+#endif
+              ) {
   LLValue *indices[] = {i0, i1};
-  return DtoGEP(pointeeTy, ptr, indices, name, bb);
+  return DtoGEP(pointeeTy, ptr, indices, name, bb
+#if LDC_LLVM_VER >= 2000
+               , nw
+#endif
+  );
 }
 
 LLValue *DtoGEP1(LLType *pointeeTy, LLValue *ptr, unsigned i0, const char *name,
-                 llvm::BasicBlock *bb) {
-  return DtoGEP(pointeeTy, ptr, DtoConstUint(i0), name, bb);
+                 llvm::BasicBlock *bb
+#if LDC_LLVM_VER >= 2000
+               , llvm::GEPNoWrapFlags nw
+#endif
+                ) {
+  return DtoGEP(pointeeTy, ptr, DtoConstUint(i0), name, bb
+#if LDC_LLVM_VER >= 2000
+                , nw
+#endif
+  );
 }
 
 LLValue *DtoGEP(LLType *pointeeTy, LLValue *ptr, unsigned i0, unsigned i1,
-                const char *name, llvm::BasicBlock *bb) {
+                const char *name, llvm::BasicBlock *bb
+#if LDC_LLVM_VER >= 2000
+              , llvm::GEPNoWrapFlags nw
+#endif
+              ) {
   LLValue *indices[] = {DtoConstUint(i0), DtoConstUint(i1)};
-  return DtoGEP(pointeeTy, ptr, indices, name, bb);
+  return DtoGEP(pointeeTy, ptr, indices, name, bb
+#if LDC_LLVM_VER >= 2000
+              , nw
+#endif
+  );
 }
 
 LLConstant *DtoGEP(LLType *pointeeTy, LLConstant *ptr, unsigned i0,
-                   unsigned i1) {
+                   unsigned i1
+#if LDC_LLVM_VER >= 2000
+                 , llvm::GEPNoWrapFlags nw
+#endif
+                  ) {
   LLValue *indices[] = {DtoConstUint(i0), DtoConstUint(i1)};
   return llvm::ConstantExpr::getGetElementPtr(pointeeTy, ptr, indices,
-                                              /* InBounds = */ true);
+#if LDC_LLVM_VER >= 2000
+                                              nw
+#else
+                                              /* InBounds = */ true
+#endif
+                                            );
 }
 
 LLValue *DtoGEP1i64(LLType *pointeeTy, LLValue *ptr, uint64_t i0, const char *name,
-                    llvm::BasicBlock *bb) {
-  return DtoGEP(pointeeTy, ptr, DtoConstUlong(i0), name, bb);
+                    llvm::BasicBlock *bb
+#if LDC_LLVM_VER >= 2000
+                  , llvm::GEPNoWrapFlags nw
+#endif
+                  ) {
+  return DtoGEP(pointeeTy, ptr, DtoConstUlong(i0), name, bb
+#if LDC_LLVM_VER >= 2000
+              , nw
+#endif
+  );
 }
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -701,7 +758,7 @@ LLGlobalVariable *makeGlobal(LLStringRef name, LLType* type, LLStringRef section
 
   if (!section.empty())
     var->setSection(section);
-    
+
   return var;
 }
 
@@ -737,7 +794,7 @@ LLGlobalVariable *makeGlobalWithBytes(LLStringRef name, LLConstantList packedCon
     0u,
     externInit
   );
-  
+
   return var;
 }
 

gen/tollvm.h

@@ -77,19 +77,43 @@ LLStructType *DtoModuleReferenceType();
 
 // getelementptr helpers
 LLValue *DtoGEP1(LLType *pointeeTy, LLValue *ptr, LLValue *i0,
-                 const char *name = "", llvm::BasicBlock *bb = nullptr);
+                 const char *name = "", llvm::BasicBlock *bb = nullptr
+#if LDC_LLVM_VER >= 2000
+                 , llvm::GEPNoWrapFlags nw = llvm::GEPNoWrapFlags::inBounds()
+#endif
+                );
 LLValue *DtoGEP(LLType *pointeeTy, LLValue *ptr, LLValue *i0, LLValue *i1,
-                const char *name = "", llvm::BasicBlock *bb = nullptr);
+                const char *name = "", llvm::BasicBlock *bb = nullptr
+#if LDC_LLVM_VER >= 2000
+                 , llvm::GEPNoWrapFlags nw = llvm::GEPNoWrapFlags::inBounds()
+#endif
+              );
 
 LLValue *DtoGEP1(LLType *pointeeTy, LLValue *ptr, unsigned i0,
-                 const char *name = "", llvm::BasicBlock *bb = nullptr);
+                 const char *name = "", llvm::BasicBlock *bb = nullptr
+#if LDC_LLVM_VER >= 2000
+                 , llvm::GEPNoWrapFlags nw = llvm::GEPNoWrapFlags::inBounds()
+#endif
+                );
 LLValue *DtoGEP(LLType *pointeeTy, LLValue *ptr, unsigned i0, unsigned i1,
-                const char *name = "", llvm::BasicBlock *bb = nullptr);
+                const char *name = "", llvm::BasicBlock *bb = nullptr
+#if LDC_LLVM_VER >= 2000
+                 , llvm::GEPNoWrapFlags nw = llvm::GEPNoWrapFlags::inBounds()
+#endif
+              );
 LLConstant *DtoGEP(LLType *pointeeTy, LLConstant *ptr, unsigned i0,
-                   unsigned i1);
+                   unsigned i1
+#if LDC_LLVM_VER >= 2000
+                 , llvm::GEPNoWrapFlags nw = llvm::GEPNoWrapFlags::inBounds()
+#endif
+                  );
 
 LLValue *DtoGEP1i64(LLType *pointeeTy, LLValue *ptr, uint64_t i0,
-                    const char *name = "", llvm::BasicBlock *bb = nullptr);
+                    const char *name = "", llvm::BasicBlock *bb = nullptr
+#if LDC_LLVM_VER >= 2000
+                 , llvm::GEPNoWrapFlags nw = llvm::GEPNoWrapFlags::inBounds()
+#endif
+                  );
 
 // to constant helpers
 LLConstantInt *DtoConstSize_t(uint64_t);

tests/codegen/inbounds.d

@@ -1,4 +1,4 @@
-// RUN: %ldc -c -output-ll -of=%t.ll %s && FileCheck %s < %t.ll
+// RUN: %ldc -enable-getelementptr-nuw -c -output-ll -of=%t.ll %s && FileCheck %s < %t.ll
 
 struct S {
     float x, y;
@@ -9,7 +9,7 @@ extern(C):  // Avoid name mangling
 // IndexExp in static array with const exp
 // CHECK-LABEL: @foo1
 int foo1(int[3] a) {
-    // CHECK: getelementptr inbounds [3 x i32]
+    // CHECK: getelementptr inbounds{{( nuw)?}} [3 x i32]
     return a[1];
 }
 
@@ -58,7 +58,7 @@ int foo7(int* p, int i) {
 // Struct field
 // CHECK-LABEL: @foo8
 float foo8(S s) {
-    // CHECK: getelementptr inbounds
+    // CHECK: getelementptr inbounds{{( nuw)?}}
     return s.y;
 }
 
@@ -79,7 +79,7 @@ int foo10(int[] a, int i) {
 // SliceExp for static array with const lower bound
 // CHECK-LABEL: @foo11
 int[] foo11(ref int[3] a) {
-    // CHECK: getelementptr inbounds i32, ptr
+    // CHECK: getelementptr inbounds{{( nuw)?}} i32, ptr
     return a[1 .. $];
 }