From e3547af7bedf5cf01cbfd38f18e7a37dda430e7f Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Sun, 25 Mar 2018 10:21:19 +0000
Subject: [PATCH] [X86] Add the ability to override memory folding latency to
 schedules and add 1uop for memory folds for Intel models

The Intel models need an extra 1uop for memory folded instructions, plus a lot of instructions take a non-default memory latency which should allow us to use the multiclass a lot more to tidy things up.

Differential Revision: https://reviews.llvm.org/D44840

llvm-svn: 328446
---
 llvm/lib/Target/X86/X86SchedBroadwell.td     | 11 ++++++-----
 llvm/lib/Target/X86/X86SchedHaswell.td       | 11 ++++++-----
 llvm/lib/Target/X86/X86SchedSandyBridge.td   | 11 ++++++-----
 llvm/lib/Target/X86/X86SchedSkylakeClient.td | 11 ++++++-----
 llvm/lib/Target/X86/X86SchedSkylakeServer.td | 11 ++++++-----
 llvm/lib/Target/X86/X86ScheduleSLM.td        |  9 +++++----
 6 files changed, 35 insertions(+), 29 deletions(-)

diff --git a/llvm/lib/Target/X86/X86SchedBroadwell.td b/llvm/lib/Target/X86/X86SchedBroadwell.td
index b575d98..0aa0700 100755
--- a/llvm/lib/Target/X86/X86SchedBroadwell.td
+++ b/llvm/lib/Target/X86/X86SchedBroadwell.td
@@ -80,7 +80,8 @@ def : ReadAdvance<ReadAfterLd, 5>;
 // folded loads.
 multiclass BWWriteResPair<X86FoldableSchedWrite SchedRW,
                           list<ProcResourceKind> ExePorts,
-                          int Lat, list<int> Res = [1], int UOps = 1> {
+                          int Lat, list<int> Res = [1], int UOps = 1,
+                          int LoadLat = 5> {
   // Register variant is using a single cycle on ExePort.
   def : WriteRes<SchedRW, ExePorts> {
     let Latency = Lat;
@@ -88,12 +89,12 @@ multiclass BWWriteResPair<X86FoldableSchedWrite SchedRW,
     let NumMicroOps = UOps;
   }
 
-  // Memory variant also uses a cycle on port 2/3 and adds 5 cycles to the
-  // latency.
+  // Memory variant also uses a cycle on port 2/3 and adds LoadLat cycles to
+  // the latency (default = 5).
   def : WriteRes<SchedRW.Folded, !listconcat([BWPort23], ExePorts)> {
-    let Latency = !add(Lat, 5);
+    let Latency = !add(Lat, LoadLat);
     let ResourceCycles = !listconcat([1], Res);
-    let NumMicroOps = UOps;
+    let NumMicroOps = !add(UOps, 1);
   }
 }
 
diff --git a/llvm/lib/Target/X86/X86SchedHaswell.td b/llvm/lib/Target/X86/X86SchedHaswell.td
index c15498d..8ae1fbf 100644
--- a/llvm/lib/Target/X86/X86SchedHaswell.td
+++ b/llvm/lib/Target/X86/X86SchedHaswell.td
@@ -81,7 +81,8 @@ def : ReadAdvance<ReadAfterLd, 5>;
 // folded loads.
 multiclass HWWriteResPair<X86FoldableSchedWrite SchedRW,
                           list<ProcResourceKind> ExePorts,
-                          int Lat, list<int> Res = [1], int UOps = 1> {
+                          int Lat, list<int> Res = [1], int UOps = 1,
+                          int LoadLat = 5> {
   // Register variant is using a single cycle on ExePort.
   def : WriteRes<SchedRW, ExePorts> {
     let Latency = Lat;
@@ -89,12 +90,12 @@ multiclass HWWriteResPair<X86FoldableSchedWrite SchedRW,
     let NumMicroOps = UOps;
   }
 
-  // Memory variant also uses a cycle on port 2/3 and adds 5 cycles to the
-  // latency.
+  // Memory variant also uses a cycle on port 2/3 and adds LoadLat cycles to
+  // the latency (default = 5).
   def : WriteRes<SchedRW.Folded, !listconcat([HWPort23], ExePorts)> {
-    let Latency = !add(Lat, 5);
+    let Latency = !add(Lat, LoadLat);
     let ResourceCycles = !listconcat([1], Res);
-    let NumMicroOps = UOps;
+    let NumMicroOps = !add(UOps, 1);
   }
 }
 
diff --git a/llvm/lib/Target/X86/X86SchedSandyBridge.td b/llvm/lib/Target/X86/X86SchedSandyBridge.td
index a23c732..86cfce1 100644
--- a/llvm/lib/Target/X86/X86SchedSandyBridge.td
+++ b/llvm/lib/Target/X86/X86SchedSandyBridge.td
@@ -72,7 +72,8 @@ def : ReadAdvance<ReadAfterLd, 4>;
 // folded loads.
 multiclass SBWriteResPair<X86FoldableSchedWrite SchedRW,
                           list<ProcResourceKind> ExePorts,
-                          int Lat, list<int> Res = [1], int UOps = 1> {
+                          int Lat, list<int> Res = [1], int UOps = 1,
+                          int LoadLat = 4> {
   // Register variant is using a single cycle on ExePort.
   def : WriteRes<SchedRW, ExePorts> {
     let Latency = Lat;
@@ -80,12 +81,12 @@ multiclass SBWriteResPair<X86FoldableSchedWrite SchedRW,
     let NumMicroOps = UOps;
   }
 
-  // Memory variant also uses a cycle on port 2/3 and adds 4 cycles to the
-  // latency.
+  // Memory variant also uses a cycle on port 2/3 and adds LoadLat cycles to
+  // the latency (default = 4).
   def : WriteRes<SchedRW.Folded, !listconcat([SBPort23], ExePorts)> {
-    let Latency = !add(Lat, 4);
+    let Latency = !add(Lat, LoadLat);
     let ResourceCycles = !listconcat([1], Res);
-    let NumMicroOps = UOps;
+    let NumMicroOps = !add(UOps, 1);
   }
 }
 
diff --git a/llvm/lib/Target/X86/X86SchedSkylakeClient.td b/llvm/lib/Target/X86/X86SchedSkylakeClient.td
index 8ee7046..8909cce 100644
--- a/llvm/lib/Target/X86/X86SchedSkylakeClient.td
+++ b/llvm/lib/Target/X86/X86SchedSkylakeClient.td
@@ -78,7 +78,8 @@ def : ReadAdvance<ReadAfterLd, 5>;
 // folded loads.
 multiclass SKLWriteResPair<X86FoldableSchedWrite SchedRW,
                           list<ProcResourceKind> ExePorts,
-                          int Lat, list<int> Res = [1], int UOps = 1> {
+                          int Lat, list<int> Res = [1], int UOps = 1,
+                          int LoadLat = 5> {
   // Register variant is using a single cycle on ExePort.
   def : WriteRes<SchedRW, ExePorts> {
     let Latency = Lat;
@@ -86,12 +87,12 @@ multiclass SKLWriteResPair<X86FoldableSchedWrite SchedRW,
     let NumMicroOps = UOps;
   }
 
-  // Memory variant also uses a cycle on port 2/3 and adds 5 cycles to the
-  // latency.
+  // Memory variant also uses a cycle on port 2/3 and adds LoadLat cycles to
+  // the latency (default = 5).
   def : WriteRes<SchedRW.Folded, !listconcat([SKLPort23], ExePorts)> {
-    let Latency = !add(Lat, 5);
+    let Latency = !add(Lat, LoadLat);
     let ResourceCycles = !listconcat([1], Res);
-    let NumMicroOps = UOps;
+    let NumMicroOps = !add(UOps, 1);
   }
 }
 
diff --git a/llvm/lib/Target/X86/X86SchedSkylakeServer.td b/llvm/lib/Target/X86/X86SchedSkylakeServer.td
index a5b8d68..5ecfd3e 100755
--- a/llvm/lib/Target/X86/X86SchedSkylakeServer.td
+++ b/llvm/lib/Target/X86/X86SchedSkylakeServer.td
@@ -78,7 +78,8 @@ def : ReadAdvance<ReadAfterLd, 5>;
 // folded loads.
 multiclass SKXWriteResPair<X86FoldableSchedWrite SchedRW,
                           list<ProcResourceKind> ExePorts,
-                          int Lat, list<int> Res = [1], int UOps = 1> {
+                          int Lat, list<int> Res = [1], int UOps = 1,
+                          int LoadLat = 5> {
   // Register variant is using a single cycle on ExePort.
   def : WriteRes<SchedRW, ExePorts> {
     let Latency = Lat;
@@ -86,12 +87,12 @@ multiclass SKXWriteResPair<X86FoldableSchedWrite SchedRW,
     let NumMicroOps = UOps;
   }
 
-  // Memory variant also uses a cycle on port 2/3 and adds 5 cycles to the
-  // latency.
+  // Memory variant also uses a cycle on port 2/3 and adds LoadLat cycles to
+  // the latency (default = 5).
   def : WriteRes<SchedRW.Folded, !listconcat([SKXPort23], ExePorts)> {
-    let Latency = !add(Lat, 5);
+    let Latency = !add(Lat, LoadLat);
     let ResourceCycles = !listconcat([1], Res);
-    let NumMicroOps = UOps;
+    let NumMicroOps = !add(UOps, 1);
   }
 }
 
diff --git a/llvm/lib/Target/X86/X86ScheduleSLM.td b/llvm/lib/Target/X86/X86ScheduleSLM.td
index 6aa4523..851d925 100644
--- a/llvm/lib/Target/X86/X86ScheduleSLM.td
+++ b/llvm/lib/Target/X86/X86ScheduleSLM.td
@@ -57,7 +57,8 @@ def : ReadAdvance<ReadAfterLd, 3>;
 // folded loads.
 multiclass SLMWriteResPair<X86FoldableSchedWrite SchedRW,
                            list<ProcResourceKind> ExePorts,
-                           int Lat, list<int> Res = [1], int UOps = 1> {
+                           int Lat, list<int> Res = [1], int UOps = 1,
+                           int LoadLat = 3> {
   // Register variant is using a single cycle on ExePort.
   def : WriteRes<SchedRW, ExePorts> {
     let Latency = Lat;
@@ -65,10 +66,10 @@ multiclass SLMWriteResPair<X86FoldableSchedWrite SchedRW,
     let NumMicroOps = UOps;
   }
 
-  // Memory variant also uses a cycle on MEC_RSV and adds 3 cycles to the
-  // latency.
+  // Memory variant also uses a cycle on MEC_RSV and adds LoadLat cycles to
+  // the latency (default = 3).
   def : WriteRes<SchedRW.Folded, !listconcat([SLM_MEC_RSV], ExePorts)> {
-    let Latency = !add(Lat, 3);
+    let Latency = !add(Lat, LoadLat);
     let ResourceCycles = !listconcat([1], Res);
     let NumMicroOps = UOps;
   }
-- 
2.7.4