From cda15ce9eae1e60bb5b579a90f72d30a49bc6354 Mon Sep 17 00:00:00 2001
From: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Date: Thu, 12 Jun 2025 15:36:14 -0400
Subject: [PATCH] RegisterAllocationPass: optimize long divsion

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
---
 FEXCore/Source/Interface/IR/IR.h              |  5 +++++
 .../IR/Passes/RegisterAllocationPass.cpp      | 22 +++++++++++++++++++
 2 files changed, 27 insertions(+)

diff --git a/FEXCore/Source/Interface/IR/IR.h b/FEXCore/Source/Interface/IR/IR.h
index 5177a1bcc..b4a1cb5a5 100644
--- a/FEXCore/Source/Interface/IR/IR.h
+++ b/FEXCore/Source/Interface/IR/IR.h
@@ -165,6 +165,11 @@ struct FEX_PACKED NodeWrapperBase final {
     LOGMAN_THROW_A_FMT(IsPointer(), "Offsets are within 2GiB range");
   }
 
+  void SetInvalid() {
+    NodeOffset = 0;
+    LOGMAN_THROW_A_FMT(IsInvalid(), "Zero state");
+  }
+
   void SetImmediate(uint32_t Immediate) {
     LOGMAN_THROW_A_FMT(Immediate < (1u << 31), "Bounded");
     NodeOffset = Immediate | (1u << 31);
diff --git a/FEXCore/Source/Interface/IR/Passes/RegisterAllocationPass.cpp b/FEXCore/Source/Interface/IR/Passes/RegisterAllocationPass.cpp
index b64d9f17c..6d11c80be 100644
--- a/FEXCore/Source/Interface/IR/Passes/RegisterAllocationPass.cpp
+++ b/FEXCore/Source/Interface/IR/Passes/RegisterAllocationPass.cpp
@@ -382,6 +382,19 @@ inline bool KillMove(IROp_Header* LastOp, IROp_Header* IROp, Ref LastNode, Ref C
   return LastOp->Op == OP_STOREREGISTER;
 }
 
+inline bool IsSignext(const IROp_Header* IROp, OrderedNodeWrapper Src, OpSize Size) {
+  if (IROp->Op == OP_SBFE) {
+    auto Sbfe = IROp->C<IR::IROp_Sbfe>();
+    return Sbfe->Width == 1 && Sbfe->lsb == (IR::OpSizeAsBits(Size) - 1) && Sbfe->Src == Src;
+  } else {
+    return false;
+  }
+}
+
+inline bool IsZero(const IROp_Header* IROp) {
+  return IROp->Op == OP_CONSTANT && IROp->C<IROp_Constant>()->Constant == 0;
+}
+
 bool ConstrainedRAPass::TryPostRAMerge(Ref LastNode, Ref CodeNode, IROp_Header* IROp) {
   auto LastOp = IR->GetOp<IROp_Header>(LastNode);
 
@@ -408,6 +421,15 @@ bool ConstrainedRAPass::TryPostRAMerge(Ref LastNode, Ref CodeNode, IROp_Header*
       IREmit->RemovePostRA(CodeNode);
       return true;
     }
+  } else if ((IROp->Op == OP_DIV || IROp->Op == OP_UDIV) && IROp->Size >= OpSize::i32Bit) {
+    // If Upper came from a sign/zero extension, we only need a 64-bit division.
+    auto Op = IROp->CW<IR::IROp_Div>();
+    if (!Op->Upper.IsInvalid() && PhysicalRegister(Op->Upper) == PhysicalRegister(LastNode)) {
+      if (IROp->Op == OP_DIV ? IsSignext(LastOp, Op->Lower, IROp->Size) : IsZero(LastOp)) {
+        Op->Upper.SetInvalid();
+        return PhysicalRegister(LastNode) == PhysicalRegister(Op->OutRemainder);
+      }
+    }
   }
 
   // Merge moves that are immediately consumed.