[llvm] [AMDGPU] Improved register pressure estimates for rescheduling (PR #213116)

[Romanov Vlad via llvm-commits]

================
@@ -2216,6 +2214,105 @@ bool MemoryClauseInitialScheduleStage::shouldRevertScheduling(
   return mayCauseSpilling(WavesAfter);
 }
 
+static cl::opt<bool> EnableLiveIntervalRPReschedule(
+    "amdgpu-lirp-reschedule", cl::Hidden,
+    cl::desc("Enable live interval RP reschedule stage"), cl::init(true));
+
+static cl::opt<unsigned> LiveIntervalRPThreshold(
+    "amdgpu-lirp-threshold", cl::Hidden,
+    cl::desc("Percent increase of live interval RP over instant pressure to "
+             "trigger rescheduling"),
+    cl::init(10));
+
+static cl::opt<unsigned> LiveIntervalRPVGPRReduction(
+    "amdgpu-lirp-vgpr-reduction", cl::Hidden,
+    cl::desc(
+        "Reduction factor (percent) for VGPR threshold during live interval RP "
+        "reschedule stage"),
+    cl::init(90));
+
+static cl::opt<unsigned> LiveIntervalRPVGPRReductionEpilogue(
+    "amdgpu-lirp-vgpr-reduction-epilogue", cl::Hidden,
+    cl::desc(
+        "Reduction factor (percent) for VGPR threshold during live interval RP "
+        "reschedule stage for exit blocks"),
+    cl::init(70));
+
+bool LiveIntervalRPStage::initGCNSchedStage() {
+  if (!EnableLiveIntervalRPReschedule)
+    return false;
+
+  if (!GCNSchedStage::initGCNSchedStage())
+    return false;
+
+  if (!S.VGPRThresholdPercent) {
+    LLVM_DEBUG(dbgs() << "LIRP: expected VGPRThresholdPercent to be enabled, "
+                         "not using live interval RP reschedule stage\n");
+    return false;
+  }
+
+  return true;
+}
+
+bool LiveIntervalRPStage::initGCNRegion() {
+  unsigned InstantRP = DAG.Pressure[RegionIdx].getArchVGPRNum();
----------------
romanovvlad wrote:

I believe we can ds_read into agprs.
AFAIK rewritemfma doesn't guarantee that there are no "over the limit agpr uses", it can only move some vgpr pressure to agpr pressure.
So, I think checking agprs and reducing agpr limit here could be useful, but probably can be done later. 

https://github.com/llvm/llvm-project/pull/213116