Add enable-getelementptr-nuw flag

Author: kubo39

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"
@@ -1868,12 +1869,16 @@ 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
 #if LDC_LLVM_VER >= 2000
-      , "", nullptr
-      , llvm::GEPNoWrapFlags::inBounds() | llvm::GEPNoWrapFlags::noUnsignedWrap()
+      , "", nullptr, nw
 #endif
     );
     ty = DtoType(vd->type);
@@ -1891,8 +1896,7 @@ DLValue *DtoIndexAggregate(LLValue *src, AggregateDeclaration *ad,
       }
       ptr = DtoGEP(st, ptr, 0, off
 #if LDC_LLVM_VER >= 2000
-      , "", nullptr
-      , llvm::GEPNoWrapFlags::inBounds() | llvm::GEPNoWrapFlags::noUnsignedWrap()
+      , "", 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"
@@ -1200,7 +1201,7 @@ class ToElemVisitor : public Visitor {
       LLType *arrty = llvm::ArrayType::get(elt, e1type->isTypeSArray()->dim->isIntegerExp()->getInteger());
 #if LDC_LLVM_VER >= 2000
       llvm::GEPNoWrapFlags nw = llvm::GEPNoWrapFlags::inBounds();
-      if (e->indexIsInBounds)
+      if (e->indexIsInBounds && opts::enableGetElementPtrNuw)
         nw |= llvm::GEPNoWrapFlags::noUnsignedWrap();
 #endif
       arrptr = DtoGEP(arrty, DtoLVal(l), DtoConstUint(0), DtoRVal(r)
@@ -1303,7 +1304,7 @@ class ToElemVisitor : public Visitor {
       // offset by lower
 #if LDC_LLVM_VER >= 2000
       llvm::GEPNoWrapFlags nw = llvm::GEPNoWrapFlags::inBounds();
-      if (!needCheckUpper && !needCheckLower)
+      if (!needCheckUpper && !needCheckLower && opts::enableGetElementPtrNuw)
         nw |= llvm::GEPNoWrapFlags::noUnsignedWrap();
 #endif
       eptr = DtoGEP1(DtoMemType(etype->nextOf()), getBasePointer(), vlo, "lowerbound"

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;