From b4019892c97dcfa365e1420ec2a8b06b89a142b3 Mon Sep 17 00:00:00 2001
From: Ioannis Assiouras <Ioannis.Assiouras@amd.com>
Date: Wed, 18 Dec 2024 23:16:52 +0000
Subject: [PATCH] SWDEV-505504 - Disable vectorization in GetHipDispatchTable

Change-Id: Id33144623555a5d25e029ca644f6274610dcd0ad


[ROCm/clr commit: 158b6a29e0468d36a092ad8d204ce04b70dd4eda]
---
 projects/clr/hipamd/src/hip_api_trace.cpp | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)

diff --git a/projects/clr/hipamd/src/hip_api_trace.cpp b/projects/clr/hipamd/src/hip_api_trace.cpp
index eefe147e46..c9d8303a1b 100644
--- a/projects/clr/hipamd/src/hip_api_trace.cpp
+++ b/projects/clr/hipamd/src/hip_api_trace.cpp
@@ -1366,11 +1366,22 @@ template <typename Tp> Tp& GetDispatchTableImpl() {
 }
 }  // namespace
 
-const HipDispatchTable* GetHipDispatchTable() {
+// At the -O3 optimization level, these functions are vectorized (gcc 11.4.1),
+// resulting in extra code to preload the dispatch table onto the stack using vector registers.
+// This preloading occurs before verifying if the table has been initialized.
+// Since the table is typically already initialized, this adds unnecessary overhead to the critical
+// path.
+#if defined(__GNUC__) || defined(__clang__)
+#define NO_VECTORIZE __attribute__((optimize("no-tree-vectorize")))
+#else
+#define NO_VECTORIZE
+#endif
+NO_VECTORIZE const HipDispatchTable* GetHipDispatchTable() {
   static auto* _v = &GetDispatchTableImpl<HipDispatchTable>();
   return _v;
 }
-const HipCompilerDispatchTable* GetHipCompilerDispatchTable() {
+NO_VECTORIZE const HipCompilerDispatchTable*
+GetHipCompilerDispatchTable() {
   static auto* _v = &GetDispatchTableImpl<HipCompilerDispatchTable>();
   return _v;
 }