[IslExprBuilder] Do not generate RTC with more than 64 bit

Such RTCs may introduce integer wrapping intrinsics with more than 64 bit,
which are translated to library calls on AOSP that are not part of the
runtime and will consequently cause linker errors.

Thanks to Eli Friedman for reporting this issue and reducing the test case.

llvm-svn: 314065

diff --git a/polly/include/polly/CodeGen/IslExprBuilder.h b/polly/include/polly/CodeGen/IslExprBuilder.h
index ce606c1..c036463 100644 (file)
--- a/polly/include/polly/CodeGen/IslExprBuilder.h
+++ b/polly/include/polly/CodeGen/IslExprBuilder.h
@@ -17,6 +17,7 @@
 
 #include "llvm/ADT/MapVector.h"
 #include "isl/ast.h"
+#include "isl/isl-noexceptions.h"
 
 namespace llvm {
 class DataLayout;
@@ -184,6 +185,13 @@ public:
   /// @return The llvm::Value* containing the result of the computation.
   llvm::Value *createAccessAddress(__isl_take isl_ast_expr *Expr);
 
+  /// Check if an @p Expr contains integer constants larger than 64 bit.
+  ///
+  /// @param Expr The expression to check.
+  ///
+  /// @return True if the ast expression is larger than 64 bit.
+  bool hasLargeInts(isl::ast_expr Expr);
+
 private:
   Scop &S;
 

diff --git a/polly/lib/CodeGen/IslExprBuilder.cpp b/polly/lib/CodeGen/IslExprBuilder.cpp
index 0155881..0a4dd91 100644 (file)
--- a/polly/lib/CodeGen/IslExprBuilder.cpp
+++ b/polly/lib/CodeGen/IslExprBuilder.cpp
@@ -73,6 +73,32 @@ Value *IslExprBuilder::getOverflowState() const {
   return OverflowState;
 }
 
+bool IslExprBuilder::hasLargeInts(isl::ast_expr Expr) {
+  enum isl_ast_expr_type Type = isl_ast_expr_get_type(Expr.get());
+
+  if (Type == isl_ast_expr_id)
+    return false;
+
+  if (Type == isl_ast_expr_int) {
+    isl::val Val = Expr.get_val();
+    APInt APValue = APIntFromVal(Val);
+    auto BitWidth = APValue.getBitWidth();
+    return BitWidth >= 64;
+  }
+
+  assert(Type == isl_ast_expr_op && "Expected isl_ast_expr of type operation");
+
+  int NumArgs = isl_ast_expr_get_op_n_arg(Expr.get());
+
+  for (int i = 0; i < NumArgs; i++) {
+    isl::ast_expr Operand = Expr.get_op_arg(i);
+    if (hasLargeInts(Operand))
+      return true;
+  }
+
+  return false;
+}
+
 Value *IslExprBuilder::createBinOp(BinaryOperator::BinaryOps Opc, Value *LHS,
                                    Value *RHS, const Twine &Name) {
   // Handle the plain operation (without overflow tracking) first.

diff --git a/polly/lib/CodeGen/IslNodeBuilder.cpp b/polly/lib/CodeGen/IslNodeBuilder.cpp
index 4e7a8ee..35210b7 100644 (file)
--- a/polly/lib/CodeGen/IslNodeBuilder.cpp
+++ b/polly/lib/CodeGen/IslNodeBuilder.cpp
@@ -1590,6 +1590,17 @@ Value *IslNodeBuilder::generateSCEV(const SCEV *Expr) {
 /// of this run-time check to false to be conservatively correct,
 Value *IslNodeBuilder::createRTC(isl_ast_expr *Condition) {
   auto ExprBuilder = getExprBuilder();
+
+  // In case the AST expression has integers larger than 64 bit, bail out. The
+  // resulting LLVM-IR will contain operations on types that use more than 64
+  // bits. These are -- in case wrapping intrinsics are used -- translated to
+  // runtime library calls that are not available on all systems (e.g., Android)
+  // and consequently will result in linker errors.
+  if (ExprBuilder.hasLargeInts(isl::manage(isl_ast_expr_copy(Condition)))) {
+    isl_ast_expr_free(Condition);
+    return Builder.getFalse();
+  }
+
   ExprBuilder.setTrackOverflow(true);
   Value *RTC = ExprBuilder.create(Condition);
   if (!RTC->getType()->isIntegerTy(1))

diff --git a/polly/test/Isl/CodeGen/large-numbers-in-boundary-context.ll b/polly/test/Isl/CodeGen/large-numbers-in-boundary-context.ll
index ccd581c..eb97855 100644 (file)
--- a/polly/test/Isl/CodeGen/large-numbers-in-boundary-context.ll
+++ b/polly/test/Isl/CodeGen/large-numbers-in-boundary-context.ll
@@ -1,13 +1,12 @@
 ; RUN: opt %loadPolly -S -polly-codegen < %s | FileCheck %s
 ;
 ; The boundary context contains a constant that does not fit in 64 bits. Hence,
-; we will check that we use an appropriaty typed constant, here with 65 bits.
-; An alternative would be to bail out early but that would not be as easy.
-;
-; CHECK: {{.*}} = icmp sle i65 {{.*}}, -9223372036854775810
-;
-; CHECK: polly.start
+; make sure we bail out. On certain systems, e.g. AOSP, no runtime support for
+; 128bit operations is available and consequently the code generation of large
+; values might cause linker errors.
 ;
+; CHECK: br i1 false, label %polly.start, label %bb11.pre_entry_bb
+
 target triple = "x86_64-unknown-linux-gnu"
 
 @global = external global i32, align 4

diff --git a/polly/test/ScopInfo/int2ptr_ptr2int.ll b/polly/test/ScopInfo/int2ptr_ptr2int.ll
index 0979eba..586ee3f 100644 (file)
--- a/polly/test/ScopInfo/int2ptr_ptr2int.ll
+++ b/polly/test/ScopInfo/int2ptr_ptr2int.ll
@@ -30,12 +30,11 @@
 ; IR-NEXT:   store i64 %p_add4, i64* %p_arrayidx3
 ;
 ; IR:      polly.loop_preheader:
-; IR-NEXT:   %scevgep = getelementptr i64, i64* %ptr, i64 1
+; IR-NEXT:   %scevgep = getelementptr i64, i64* %ptr, i32 1
 ; IR-NEXT:   %10 = add i64 %val, 1
 ; IR-NEXT:   br label %polly.loop_header
 ;
-;
-target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target datalayout = "e-p:32:32:32-m:e-i64:64-f80:128-n8:16:32:64-S128"
 
 define void @f(i64* %A, i64* %ptr, i64 %val) {
 entry:

diff --git a/polly/test/ScopInfo/int2ptr_ptr2int_2.ll b/polly/test/ScopInfo/int2ptr_ptr2int_2.ll
index 3fc64d4..951f493 100644 (file)
--- a/polly/test/ScopInfo/int2ptr_ptr2int_2.ll
+++ b/polly/test/ScopInfo/int2ptr_ptr2int_2.ll
@@ -13,7 +13,7 @@
 ;
 ; CHECK:        ReadAccess :=  [Reduction Type: NONE] [Scalar: 0]
 ; CHECK-NEXT:       [val, ptr] -> { Stmt_for_body[i0] -> MemRef_B[9 + val] };
-; CHECK-NEXT:   Execution Context: [val, ptr] -> {  : val <= 9223372036854775806 }
+; CHECK-NEXT:   Execution Context: [val, ptr] -> {  : val <= 32766 }
 ;
 ; CHECK:   ReadAccess :=       [Reduction Type: +] [Scalar: 0]
 ; CHECK-NEXT:       [val, ptr] -> { Stmt_for_body[i0] -> MemRef_A[9 + ptr] };
@@ -21,22 +21,22 @@
 ; CHECK-NEXT:       [val, ptr] -> { Stmt_for_body[i0] -> MemRef_A[9 + ptr] };
 ;
 ; IR:      polly.stmt.for.body:
-; IR-NEXT:  %p_tmp = ptrtoint i64* %scevgep to i64
-; IR-NEXT:  %p_add = add nsw i64 %p_tmp, 1
-; IR-NEXT:  %p_arrayidx3 = getelementptr inbounds i64, i64* %A, i64 %p_add
+; IR-NEXT:  %p_tmp = ptrtoint i64* %scevgep to i16
+; IR-NEXT:  %p_add = add nsw i16 %p_tmp, 1
+; IR-NEXT:  %p_arrayidx3 = getelementptr inbounds i64, i64* %A, i16 %p_add
 ; IR-NEXT:  %tmp4_p_scalar_ = load i64, i64* %p_arrayidx3
 ; IR-NEXT:  %p_add4 = add nsw i64 %tmp4_p_scalar_, %polly.preload.tmp3.merge
 ; IR-NEXT:  store i64 %p_add4, i64* %p_arrayidx3
 ;
 ; IR:      polly.loop_preheader:
-; IR-NEXT:   %scevgep = getelementptr i64, i64* %ptr, i64 1
-; IR-NEXT:   %26 = add i64 %val, 1
+; IR-NEXT:   %scevgep = getelementptr i64, i64* %ptr, i16 1
+; IR-NEXT:   %35 = add i16 %val, 1
 ; IR-NEXT:   br label %polly.loop_header
 ;
 ;
-target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target datalayout = "e-p:16:16:16-m:e-i64:64-f80:128-n8:16:16:64-S128"
 
-define void @f(i64* %A, i64* %B, i64* %ptr, i64 %val) {
+define void @f(i64* %A, i64* %B, i64* %ptr, i16 %val) {
 entry:
   br label %for.cond
 
@@ -47,15 +47,15 @@ for.cond:                                         ; preds = %for.inc, %entry
 
 for.body:                                         ; preds = %for.cond
   %add.ptr = getelementptr inbounds i64, i64* %ptr, i64 1
-  %tmp = ptrtoint i64* %add.ptr to i64
-  %add = add nsw i64 %tmp, 1
-  %add1 = add nsw i64 %val, 1
-  %tmp1 = inttoptr i64 %add1 to i64*
+  %tmp = ptrtoint i64* %add.ptr to i16
+  %add = add nsw i16 %tmp, 1
+  %add1 = add nsw i16 %val, 1
+  %tmp1 = inttoptr i16 %add1 to i64*
   %add.ptr2 = getelementptr inbounds i64, i64* %tmp1, i64 1
-  %tmp2 = ptrtoint i64* %add.ptr2 to i64
-  %arrayidx = getelementptr inbounds i64, i64* %B, i64 %tmp2
+  %tmp2 = ptrtoint i64* %add.ptr2 to i16
+  %arrayidx = getelementptr inbounds i64, i64* %B, i16 %tmp2
   %tmp3 = load i64, i64* %arrayidx
-  %arrayidx3 = getelementptr inbounds i64, i64* %A, i64 %add
+  %arrayidx3 = getelementptr inbounds i64, i64* %A, i16 %add
   %tmp4 = load i64, i64* %arrayidx3
   %add4 = add nsw i64 %tmp4, %tmp3
   store i64 %add4, i64* %arrayidx3