From 6f41394457c5b3fb629dbf452c4942f14f00eaa9 Mon Sep 17 00:00:00 2001
From: Nuno Lopes <nuno.lopes@tecnico.ulisboa.pt>
Date: Sat, 7 Dec 2024 17:07:52 +0000
Subject: [PATCH] memory: optimize enconding when alignment of globals doesn't
 change This is the common case

---
 ir/globals.cpp      | 1 +
 ir/globals.h        | 3 +++
 ir/memory.cpp       | 3 +++
 ir/pointer.cpp      | 9 +++++++--
 tools/transform.cpp | 4 ++++
 5 files changed, 18 insertions(+), 2 deletions(-)

diff --git a/ir/globals.cpp b/ir/globals.cpp
index 62bb7c31c..f4a3e75d1 100644
--- a/ir/globals.cpp
+++ b/ir/globals.cpp
@@ -42,6 +42,7 @@ bool has_ptr_arg = true;
 bool has_initializes_attr = true;
 bool has_null_block = true;
 bool null_is_dereferenceable = false;
+bool has_globals_diff_align = true;
 bool does_int_mem_access = true;
 bool does_ptr_mem_access = true;
 bool does_ptr_store = true;
diff --git a/ir/globals.h b/ir/globals.h
index 478318f06..5c66219c3 100644
--- a/ir/globals.h
+++ b/ir/globals.h
@@ -87,6 +87,9 @@ extern bool has_null_block;
 
 extern bool null_is_dereferenceable;
 
+/// Whether there is at least one global with different alignment in src/tgt
+extern bool has_globals_diff_align;
+
 /// Whether the programs do memory accesses that load/store int/ptrs
 extern bool does_int_mem_access;
 extern bool does_ptr_mem_access;
diff --git a/ir/memory.cpp b/ir/memory.cpp
index 181960b32..326e9012f 100644
--- a/ir/memory.cpp
+++ b/ir/memory.cpp
@@ -2126,6 +2126,9 @@ Memory::alloc(const expr *size, uint64_t align, BlockKind blockKind,
   expr nooverflow = true;
   if (size) {
     size_zext  = size->zextOrTrunc(bits_size_t);
+    // we round up the size statically instead of creating a large expr later
+    if (!has_globals_diff_align)
+      size_zext = size_zext.round_up(expr::mkUInt(align, bits_size_t));
     nooverflow = size->bits() <= bits_size_t ? true :
                    size->extract(size->bits()-1, bits_size_t) == 0;
   }
diff --git a/ir/pointer.cpp b/ir/pointer.cpp
index 595d8f8f1..93f80a16b 100644
--- a/ir/pointer.cpp
+++ b/ir/pointer.cpp
@@ -331,7 +331,11 @@ expr Pointer::blockSizeOffsetT() const {
 }
 
 expr Pointer::blockSizeAligned() const {
-  return blockSize().round_up_bits(blockAlignment().zextOrTrunc(bits_size_t));
+  auto size = blockSize();
+  // programs can't observe whether the size was increased up to alignment
+  if (!has_globals_diff_align)
+    return size;
+  return size.round_up_bits(blockAlignment().zextOrTrunc(bits_size_t));
 }
 
 expr Pointer::blockSizeAlignedOffsetT() const {
@@ -456,7 +460,8 @@ expr Pointer::inbounds(bool simplify_ptr) {
 
 expr Pointer::blockAlignment() const {
   return getValue("blk_align", m.local_blk_align, m.non_local_blk_align,
-                   expr::mkUInt(0, Memory::bitsAlignmentInfo()), true);
+                   expr::mkUInt(0, Memory::bitsAlignmentInfo()),
+                   has_globals_diff_align);
 }
 
 expr Pointer::isBlockAligned(uint64_t align, bool exact) const {
diff --git a/tools/transform.cpp b/tools/transform.cpp
index 62236402e..ad45c6511 100644
--- a/tools/transform.cpp
+++ b/tools/transform.cpp
@@ -960,6 +960,7 @@ static void calculateAndInitConstants(Transform &t) {
   uint64_t glb_alloc_aligned_size = 0;
 
   num_consts_src = 0;
+  has_globals_diff_align = false;
 
   for (auto GV : globals_src) {
     if (GV->isConst())
@@ -974,7 +975,10 @@ static void calculateAndInitConstants(Transform &t) {
       [GVT](auto *GV) -> bool { return GVT->getName() == GV->getName(); });
     if (I == globals_src.end()) {
       ++num_globals;
+    } else {
+      has_globals_diff_align |= GVT->getAlignment() != (*I)->getAlignment();
     }
+
     glb_alloc_aligned_size
       = add_saturate(glb_alloc_aligned_size,
                      aligned_alloc_size(GVT->size(), GVT->getAlignment()));