[llvm] [AMDGPU] Rewrite GFX12 SGPR hazard handling to dedicated pass (PR #118750)

[Carl Ritson via llvm-commits]

================
@@ -0,0 +1,502 @@
+//===- AMDGPUWaitSGPRHazards.cpp - Insert waits for SGPR read hazards -----===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// Insert s_wait_alu instructions to mitigate SGPR read hazards on GFX12.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "GCNSubtarget.h"
+#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "SIInstrInfo.h"
+#include "llvm/ADT/SetVector.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "amdgpu-wait-sgpr-hazards"
+
+static cl::opt<bool> GlobalEnableSGPRHazardWaits(
+    "amdgpu-sgpr-hazard-wait", cl::init(true), cl::Hidden,
+    cl::desc("Enable required s_wait_alu on SGPR hazards"));
+
+static cl::opt<bool> GlobalCullSGPRHazardsOnFunctionBoundary(
+    "amdgpu-sgpr-hazard-boundary-cull", cl::init(false), cl::Hidden,
+    cl::desc("Cull hazards on function boundaries"));
----------------
perlfu wrote:

Hazard culling introduces a more expensive sequence of DS_NOP instruction to clear all hazards from the VALU.  It can be more efficient to do this in some cases, but otherwise adds slow down to other case.  Hence why it is an option.  Correctness is not effected with this option.

https://github.com/llvm/llvm-project/pull/118750