Stronger chain detection in LoopCarry pass

src/LoopCarry.cpp

@@ -283,11 +283,34 @@ class LoopCarryOverLoop : public IRMutator {
 
         // For each load, move the load index forwards by one loop iteration
         vector<Expr> indices, next_indices, predicates, next_predicates;
+        // CSE-d versions of the above, so can_prove can be safely used on them.
+        vector<Expr> indices_csed, next_indices_csed, predicates_csed, next_predicates_csed;
         for (const vector<const Load *> &v : loads) {
             indices.push_back(v[0]->index);
             next_indices.push_back(step_forwards(v[0]->index, linear));
             predicates.push_back(v[0]->predicate);
             next_predicates.push_back(step_forwards(v[0]->predicate, linear));
+
+            if (indices.back().defined()) {
+                indices_csed.push_back(common_subexpression_elimination(indices.back()));
+            } else {
+                indices_csed.emplace_back();
+            }
+            if (next_indices.back().defined()) {
+                next_indices_csed.push_back(common_subexpression_elimination(next_indices.back()));
+            } else {
+                next_indices_csed.emplace_back();
+            }
+            if (predicates.back().defined()) {
+                predicates_csed.push_back(common_subexpression_elimination(predicates.back()));
+            } else {
+                predicates_csed.emplace_back();
+            }
+            if (next_predicates.back().defined()) {
+                next_predicates_csed.push_back(common_subexpression_elimination(next_predicates.back()));
+            } else {
+                next_predicates_csed.emplace_back();
+            }
         }
 
         // Find loads done on this loop iteration that will be
@@ -299,11 +322,16 @@ class LoopCarryOverLoop : public IRMutator {
                 if (i == j) {
                     continue;
                 }
+                // can_prove is stronger than graph_equal, because it doesn't require index expressions to be
+                // exactly the same, but evaluate to the same value. We keep the graph_equal check, because
+                // it's faster and should be executed before the more expensive check.
                 if (loads[i][0]->name == loads[j][0]->name &&
                     next_indices[j].defined() &&
-                    graph_equal(indices[i], next_indices[j]) &&
+                    (graph_equal(indices[i], next_indices[j]) ||
+                     ((indices[i].type() == next_indices[j].type()) && can_prove(indices_csed[i] == next_indices_csed[j]))) &&
                     next_predicates[j].defined() &&
-                    graph_equal(predicates[i], next_predicates[j])) {
+                    (graph_equal(predicates[i], next_predicates[j]) ||
+                     ((predicates[i].type() == next_predicates[j].type()) && can_prove(predicates_csed[i] == next_predicates_csed[j])))) {
                     chains.push_back({j, i});
                     debug(3) << "Found carried value:\n"
                              << i << ":  -> " << Expr(loads[i][0]) << "\n"

test/correctness/CMakeLists.txt

@@ -199,6 +199,7 @@ tests(GROUPS correctness
       likely.cpp
       load_library.cpp
       logical.cpp
+      loop_carry.cpp
       loop_invariant_extern_calls.cpp
       loop_level_generator_param.cpp
       lossless_cast.cpp

test/correctness/loop_carry.cpp

@@ -0,0 +1,69 @@
+#include "Halide.h"
+#include <stdio.h>
+
+// This file demonstrates two example custom lowering passes. The
+// first just makes sure the IR passes some test, and doesn't modify
+// it. The second actually changes the IR in some useful way.
+
+using namespace Halide;
+using namespace Halide::Internal;
+
+// Verify that all floating point divisions by constants have been
+// converted to float multiplication.
+class LoopCarryWrapper : public IRMutator {
+    using IRMutator::visit;
+
+    int register_count_;
+    Stmt mutate(const Stmt &stmt) override {
+        return simplify(loop_carry(stmt, register_count_));
+    }
+
+public:
+    LoopCarryWrapper(int register_count)
+        : register_count_(register_count) {
+    }
+};
+
+int main(int argc, char **argv) {
+    Func input;
+    Func g;
+    Func h;
+    Func f;
+    Var x, y, xo, yo, xi, yi;
+
+    input(x, y) = x + y;
+
+    Expr sum_expr = 0;
+    for (int ix = -100; ix <= 100; ix++) {
+        // Generate two chains of sums, but only one of them will be carried.
+        sum_expr += input(x, y + ix);
+        sum_expr += input(x + 13, y + 2 * ix);
+    }
+    g(x, y) = sum_expr;
+    h(x, y) = g(x, y) + 12;
+    f(x, y) = h(x, y);
+
+    // Make a maximum number of the carried values very large for the purpose
+    // of this test.
+    constexpr int kMaxRegisterCount = 1024;
+    f.add_custom_lowering_pass(new LoopCarryWrapper(kMaxRegisterCount));
+
+    const int size = 128;
+    f.compute_root()
+        .bound(x, 0, size)
+        .bound(y, 0, size);
+
+    h.compute_root()
+        .tile(x, y, xo, yo, xi, yi, 16, 16, TailStrategy::RoundUp);
+
+    g.compute_at(h, xo)
+        .reorder(y, x)
+        .vectorize(x, 4);
+
+    input.compute_root();
+
+    f.realize({size, size});
+
+    printf("Success!\n");
+    return 0;
+}