consistently use specializeToContext to specialize scop to context

Using specializeToContext improves consistency of the representation,
both by not having outside users modify the domain, but also
by having only a single location that takes into account the context.

Author: Sven Verdoolaege

tc/core/polyhedral/scop.h

@@ -102,12 +102,9 @@ struct Scop {
     // strategy.proto.fix_parameters_before_scheduling to true.
     // If you want to fix the parameters in the support domain,
     // then you need to do it explicitly.
-    // Note that the access relations must be intersect with the context as
-    // well to obtain consistent dependences.
     // TODO: this is still subject to interpretation but intersecting seems
     // final here so probably we're right not doing it.
-    // res->domain() =
-    // res->domain().intersect_params(res->globalParameterContext);
+    // res->specializeToContext();
     return res;
   }
 

test/test_cuda_mapper.cc

@@ -520,10 +520,9 @@ constexpr auto kExpectedMatmul_64_64_64 =
 TEST_F(PolyhedralMapperTest, MergedContexts) {
   auto scop = PrepareAndJoinBandsMatMul();
 
-  // Unit test claims to use scop->globalParameterContext properly
+  // Unit test claims to use the specialized context properly
   scop->fixParameters<int>({{"M", 64}, {"N", 64}, {"K", 64}});
-  auto globalParameterContext = scop->globalParameterContext;
-  scop->domain() = scop->domain().intersect_params(globalParameterContext);
+  scop->specializeToContext();
 
   auto mscop = TileAndMapThreads(std::move(scop), {16, 16}, {32ul, 8ul});
   auto res = std::get<0>(mscop->codegen(specializedName));

test/test_cuda_mapper_memory_promotion.cc

@@ -46,8 +46,7 @@ class TestMapper : public ::testing::Test {
     auto ctx = isl::with_exceptions::globalIslCtx();
     auto scop = Scop::makeScop(ctx, tc);
     scop = Scop::makeSpecializedScop(*scop, problemSizes);
-    scop->domain() =
-        scop->domain().intersect_params(scop->globalParameterContext);
+    scop->specializeToContext();
     return MappedScop::makeWithOuterBlockInnerThreadStrategy(
         std::move(scop), mappingOptions);
   }
@@ -251,7 +250,7 @@ def fun(float(N, M) A, float(N, M) B) -> (C) {
         {tile1, tile2});
     auto& scop = const_cast<Scop&>(mscop->scop());
     // Must force domain intersection for overapproximation to work
-    scop.domain() = scop.domain().intersect_params(scop.globalParameterContext);
+    scop.specializeToContext();
     auto ctx = scop.domain().get_ctx();
     auto groups = TensorReferenceGroup::accessedBySubtree(
         scop.scheduleRoot()->child(childPos), scop);
@@ -331,7 +330,7 @@ def fun(float(N, M) A) -> (B, C) {
         {tile1, tile2});
     auto& scop = const_cast<Scop&>(mscop->scop());
     // Must force domain intersection for overapproximation to work
-    scop.domain() = scop.domain().intersect_params(scop.globalParameterContext);
+    scop.specializeToContext();
     auto ctx = scop.domain().get_ctx();
     auto groups = TensorReferenceGroup::accessedBySubtree(
         scop.scheduleRoot()->child(childPos), scop);