Use ScalarEvolution to create tight bounds on the parameters

[grosser]

Hi Johannes,

this patch LGTM besides two minor comments:

1. The LLVM language reference says "The range is allowed to wrap."

In case the lower bound is larger than the upper bound (e.g. [10, 5) we would add constraints p >= 10 && p < 5. However, this should rather be p >= 10 || p < 5.

2. Add a comment that this change exploits LLVM's range metadata to the commit message

Sebastian pointed out that from a look at the commit message or the test case it is not immediately clear where the information about the bounded range of p is derived from. It would be nice to point this out in the commit message.

[sebpop]

2. Add a comment that this change exploits LLVM's range metadata to the commit message

Sebastian pointed out that from a look at the commit message or the test case it is not immediately clear where the information about the bounded range of p is derived from. It would be nice to point this out in the commit message.

I would like the comment to also be included in a comment in the
testcase close by the check line:

+; CHECK: [p_0] -> { : p_0 >= 0 and p_0 <= 255 }

pointing out that the constraint is deduced from the stmt using the
range !0 metadata:

+ %tmp = load i32* %p, align 4, !range !0
+!0 = !{ i32 0, i32 256 }

Thanks,
Sebastian

[jdoerfert]

landed in r230230.

lib/Analysis/ScopInfo.cpp

 
 void Scop::buildContext() {
   isl_space *Space = isl_space_params_alloc(IslCtx, 0);
   Context = isl_set_universe(isl_space_copy(Space));
   AssumedContext = isl_set_universe(Space);
 }
 
 void Scop::addParameterBounds() {
-  for (unsigned i = 0; i < isl_set_dim(Context, isl_dim_param); ++i) {
+  for (const auto &ParamID : ParameterIds) {
     isl_val *V;
-    isl_id *Id;
-    const SCEV *Scev;
-    const IntegerType *T;
-    int Width;
+    int dim = ParamID.second;
 
-    Id = isl_set_get_dim_id(Context, isl_dim_param, i);
-    Scev = (const SCEV *)isl_id_get_user(Id);
-    isl_id_free(Id);
+    ConstantRange SRange = SE->getSignedRange(ParamID.first);
 
-    T = dyn_cast<IntegerType>(Scev->getType());
-
-    if (!T)
+    // TODO: Find a case where the full set is actually helpful.
+    if (SRange.isFullSet())
       continue;
 
-    Width = T->getBitWidth();
-
-    V = isl_val_int_from_si(IslCtx, Width - 1);
-    V = isl_val_2exp(V);
-    V = isl_val_neg(V);
-    Context = isl_set_lower_bound_val(Context, isl_dim_param, i, V);
+    V = isl_valFromAPInt(IslCtx, SRange.getLower(), true);
+    Context = isl_set_lower_bound_val(Context, isl_dim_param, dim, V);
 
-    V = isl_val_int_from_si(IslCtx, Width - 1);
-    V = isl_val_2exp(V);
+    V = isl_valFromAPInt(IslCtx, SRange.getUpper(), true);
     V = isl_val_sub_ui(V, 1);
-    Context = isl_set_upper_bound_val(Context, isl_dim_param, i, V);
+    Context = isl_set_upper_bound_val(Context, isl_dim_param, dim, V);
   }
 }
 
 void Scop::realignParams() {
   // Add all parameters into a common model.
   isl_space *Space = isl_space_params_alloc(IslCtx, ParameterIds.size());
 
   for (const auto &ParamID : ParameterIds) {
void Scop::simplifyAssumedContext() {
   //       for (long j = 0; j < m; j++)
   //         A[i+p][j] = 1.0;
   //
   //   we assume that the condition m <= 0 or (m >= 1 and p >= 0) holds as
   //   otherwise we would access out of bound data. Now, knowing that code is
   //   only executed for the case m >= 0, it is sufficient to assume p >= 0.
   AssumedContext =
       isl_set_gist_params(AssumedContext, isl_union_set_params(getDomains()));
+  AssumedContext = isl_set_gist_params(AssumedContext, getContext());
 }
 
 /// @brief Add the minimal/maximal access in @p Set to @p User.
 static int buildMinMaxAccess(__isl_take isl_set *Set, void *User) {
   Scop::MinMaxVectorTy *MinMaxAccesses = (Scop::MinMaxVectorTy *)User;
   isl_pw_multi_aff *MinPMA, *MaxPMA;
   isl_pw_aff *LastDimAff;
   isl_aff *OneAff;

test/ScopInfo/ranged_parameter.ll

+; RUN: opt %loadPolly -polly-scops -analyze < %s | FileCheck %s
+;
+; CHECK: Context:
+; CHECK:   [p_0] -> {  : p_0 >= 0 and p_0 <= 255 }
+;
+;    void jd(int *A, int *p /* in [0,256) */) {
+;      for (int i = 0; i < 1024; i++)
+;        A[i + *p] = i;
+;    }
+;
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+define void @jd(i32* %A, i32* %p) {
+entry:
+  %tmp = load i32* %p, align 4, !range !0
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.inc ]
+  %exitcond = icmp ne i32 %i.0, 1024
+  br i1 %exitcond, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %add = add i32 %i.0, %tmp
+  %idxprom = sext i32 %add to i64
+  %arrayidx = getelementptr inbounds i32* %A, i64 %idxprom
+  store i32 %i.0, i32* %arrayidx, align 4
+  br label %for.inc
+
+for.inc:                                          ; preds = %for.body
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+!0 =  !{ i32 0, i32 256 }