From ba4da8e0d32a4a4fc0e4df8744fef6039f9b1975 Mon Sep 17 00:00:00 2001
From: Erik Eckstein <eeckstein@apple.com>
Date: Wed, 13 May 2020 16:09:30 +0200
Subject: [PATCH] LICM: enable more stores to moved out of a loop

Even if a store is not dominating the loop exits, it makes sense to move it out of the loop if the pre-header also as a store to the same memory location.
When this is done, dead-store-elimination can then most likely remove the store in the pre-header.
---
 lib/SILOptimizer/LoopTransforms/LICM.cpp | 35 ++++++++++++++++++++
 test/SILOptimizer/licm.sil               | 41 ++++++++++++++++++++++++
 2 files changed, 76 insertions(+)

diff --git a/lib/SILOptimizer/LoopTransforms/LICM.cpp b/lib/SILOptimizer/LoopTransforms/LICM.cpp
index 171a10abd863..a70f93d764c2 100644
--- a/lib/SILOptimizer/LoopTransforms/LICM.cpp
+++ b/lib/SILOptimizer/LoopTransforms/LICM.cpp
@@ -963,6 +963,41 @@ static bool storesCommonlyDominateLoopExits(SILValue addr, SILLoop *loop,
   if (stores.count(header) != 0)
     return true;
 
+  // Also a store in the pre-header dominates all exists. Although the situation
+  // is a bit different here: the store in the pre-header remains - it's not
+  // (re)moved by the LICM transformation.
+  // But even if the loop-stores are not dominating the loop exits, it
+  // makes sense to move them out of the loop if this case. When this is done,
+  // dead-store-elimination can then most likely eliminate the store in the
+  // pre-header.
+  //
+  //   pre_header:
+  //     store %v1 to %addr
+  //   header:
+  //     cond_br %cond, then, tail
+  //   then:
+  //     store %v2 to %addr    // a conditional store in the loop
+  //     br tail
+  //   tail:
+  //     cond_br %loop_cond, header, exit
+  //   exit:
+  //
+  //  will be transformed to
+  //
+  //   pre_header:
+  //     store %v1 to %addr    // <- can be removed by DSE afterwards
+  //   header:
+  //     cond_br %cond, then, tail
+  //   then:
+  //     br tail
+  //   tail(%phi):
+  //     cond_br %loop_cond, header, exit
+  //   exit:
+  //     store %phi to %addr
+  //
+  if (stores.count(loop->getLoopPreheader()) != 0)
+    return true;
+
   // Propagate the store-is-not-alive flag through the control flow in the loop,
   // starting at the header.
   SmallPtrSet<SILBasicBlock *, 16> storesNotAlive;
diff --git a/test/SILOptimizer/licm.sil b/test/SILOptimizer/licm.sil
index 2a6452d43630..153834ffc0f6 100644
--- a/test/SILOptimizer/licm.sil
+++ b/test/SILOptimizer/licm.sil
@@ -502,6 +502,47 @@ bb6:
   return %12 : $()
 }
 
+// CHECK-LABEL: sil @hoist_when_store_is_in_preheader
+// CHECK: bb0(%0 : $*Int32, %1 : $Int32):
+// CHECK:   store
+// CHECK:   load
+// CHECK: bb1(%{{[0-9]+}} : $Int32):
+// CHECK-NOT: load
+// CHECK-NOT: store
+// CHECK: bb4([[P:%[0-9]+]] : $Int32):
+// CHECK: bb6:
+// CHECK:   store [[P]] to %0
+// CHECK: } // end sil function 'hoist_when_store_is_in_preheader'
+sil @hoist_when_store_is_in_preheader : $@convention(thin) (@inout Int32, Int32) -> () {
+bb0(%0 : $*Int32, %1 : $Int32):
+  %8 = struct_element_addr %0 : $*Int32, #Int32._value
+  %9 = struct_extract %1 : $Int32, #Int32._value
+  %10 = integer_literal $Builtin.Int1, 0
+  store %1 to %0 : $*Int32
+  br bb1
+
+bb1:
+  %17 = load %8 : $*Builtin.Int32
+  %18 = builtin "sadd_with_overflow_Int64"(%17 : $Builtin.Int32, %9 : $Builtin.Int32, %10 : $Builtin.Int1) : $(Builtin.Int32, Builtin.Int1)
+  %19 = tuple_extract %18 : $(Builtin.Int32, Builtin.Int1), 0
+  %20 = struct $Int32 (%19 : $Builtin.Int32)
+  cond_br undef, bb2, bb3
+bb2:
+  br bb4
+bb3:
+  store %20 to %0 : $*Int32
+  br bb4
+bb4:
+  cond_br undef, bb5, bb6
+
+bb5:
+  br bb1
+
+bb6:
+  %12 = tuple ()
+  return %12 : $()
+}
+
 // CHECK-LABEL: sil @hoist_loads_and_stores_multiple_exits
 // CHECK:   [[V1:%[0-9]+]] = load %0
 // CHECK:   br bb1([[V1]] : $Int32)