From 2b045324b2ca3974b8acbcb2867dd39159a75ea3 Mon Sep 17 00:00:00 2001
From: Monk Chiang <monk.chiang@sifive.com>
Date: Wed, 29 Jun 2022 23:17:57 -0700
Subject: [PATCH] [RISCV] Add scheduling resources for vector segment
 instructions.

Add scheduling resources for vector segment instructions

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D128886
---
 llvm/lib/Target/RISCV/RISCVInstrInfoV.td | 76 ++++++++++++++++++------
 llvm/lib/Target/RISCV/RISCVScheduleV.td  | 28 +++++++++
 2 files changed, 85 insertions(+), 19 deletions(-)

diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoV.td b/llvm/lib/Target/RISCV/RISCVInstrInfoV.td
index d466e278cafc..1ad634344c09 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoV.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoV.td
@@ -115,6 +115,35 @@ class VSXSched<int n, string o> :
 class VLFSched<int n> : Sched <[!cast<SchedReadWrite>("WriteVLDFF" # n),
                                 ReadVLDX, ReadVMask]>;
 
+// Unit-Stride Segment Loads and Stores
+class VLSEGSched<int nf, int eew> : Sched<[
+  !cast<SchedReadWrite>("WriteVLSEG" #nf #"e" #eew), ReadVLDX, ReadVMask]>;
+class VSSEGSched<int nf, int eew> : Sched<[
+  !cast<SchedReadWrite>("WriteVSSEG" #nf #"e" #eew),
+  !cast<SchedReadWrite>("ReadVSTE" #eew #"V"), ReadVSTX, ReadVMask]>;
+class VLSEGFFSched<int nf, int eew> : Sched<[
+  !cast<SchedReadWrite>("WriteVLSEGFF" #nf #"e" #eew), ReadVLDX, ReadVMask]>;
+// Strided Segment Loads and Stores
+class VLSSEGSched<int nf, int eew> : Sched<[
+  !cast<SchedReadWrite>("WriteVLSSEG" #nf #"e" #eew), ReadVLDX, ReadVLDSX,
+  ReadVMask]>;
+class VSSSEGSched<int nf, int eew> : Sched<[
+  !cast<SchedReadWrite>("WriteVSSSEG" #nf #"e" #eew),
+  !cast<SchedReadWrite>("ReadVSTS" #eew #"V"), ReadVSTX, ReadVSTSX, ReadVMask]>;
+// Indexed Segment Loads and Stores
+class VLUXSEGSched<int nf, int eew> : Sched<[
+  !cast<SchedReadWrite>("WriteVLUXSEG" #nf #"e" #eew), ReadVLDX, ReadVLDUXV,
+  ReadVMask]>;
+class VLOXSEGSched<int nf, int eew> : Sched<[
+  !cast<SchedReadWrite>("WriteVLOXSEG" #nf #"e" #eew), ReadVLDX, ReadVLDOXV,
+  ReadVMask]>;
+class VSUXSEGSched<int nf, int eew> : Sched<[
+  !cast<SchedReadWrite>("WriteVSUXSEG" #nf #"e" #eew),
+  !cast<SchedReadWrite>("ReadVSTUX" #eew), ReadVSTX, ReadVSTUXV, ReadVMask]>;
+class VSOXSEGSched<int nf, int eew> : Sched<[
+  !cast<SchedReadWrite>("WriteVSOXSEG" #nf #"e" #eew),
+  !cast<SchedReadWrite>("ReadVSTOX" #eew), ReadVSTX, ReadVSTOXV, ReadVMask]>;
+
 //===----------------------------------------------------------------------===//
 // Instruction class templates
 //===----------------------------------------------------------------------===//
@@ -1495,31 +1524,35 @@ let Predicates = [HasVInstructions] in {
       defvar w = !cast<RISCVWidth>("LSWidth"#eew);
 
       def VLSEG#nf#E#eew#_V :
-        VUnitStrideSegmentLoad<!add(nf, -1), w, "vlseg"#nf#"e"#eew#".v">;
+        VUnitStrideSegmentLoad<!add(nf, -1), w, "vlseg"#nf#"e"#eew#".v">,
+        VLSEGSched<nf, eew>;
       def VLSEG#nf#E#eew#FF_V :
-        VUnitStrideSegmentLoadFF<!add(nf, -1), w, "vlseg"#nf#"e"#eew#"ff.v">;
+        VUnitStrideSegmentLoadFF<!add(nf, -1), w, "vlseg"#nf#"e"#eew#"ff.v">,
+        VLSEGFFSched<nf, eew>;
       def VSSEG#nf#E#eew#_V :
-        VUnitStrideSegmentStore<!add(nf, -1), w, "vsseg"#nf#"e"#eew#".v">;
-
+        VUnitStrideSegmentStore<!add(nf, -1), w, "vsseg"#nf#"e"#eew#".v">,
+        VSSEGSched<nf, eew>;
       // Vector Strided Instructions
       def VLSSEG#nf#E#eew#_V :
-        VStridedSegmentLoad<!add(nf, -1), w, "vlsseg"#nf#"e"#eew#".v">;
+        VStridedSegmentLoad<!add(nf, -1), w, "vlsseg"#nf#"e"#eew#".v">,
+        VLSSEGSched<nf, eew>;
       def VSSSEG#nf#E#eew#_V :
-        VStridedSegmentStore<!add(nf, -1), w, "vssseg"#nf#"e"#eew#".v">;
+        VStridedSegmentStore<!add(nf, -1), w, "vssseg"#nf#"e"#eew#".v">,
+        VSSSEGSched<nf, eew>;
 
       // Vector Indexed Instructions
       def VLUXSEG#nf#EI#eew#_V :
         VIndexedSegmentLoad<!add(nf, -1), MOPLDIndexedUnord, w,
-                            "vluxseg"#nf#"ei"#eew#".v">;
+                            "vluxseg"#nf#"ei"#eew#".v">, VLUXSEGSched<nf, eew>;
       def VLOXSEG#nf#EI#eew#_V :
         VIndexedSegmentLoad<!add(nf, -1), MOPLDIndexedOrder, w,
-                            "vloxseg"#nf#"ei"#eew#".v">;
+                            "vloxseg"#nf#"ei"#eew#".v">, VLOXSEGSched<nf, eew>;
       def VSUXSEG#nf#EI#eew#_V :
         VIndexedSegmentStore<!add(nf, -1), MOPSTIndexedUnord, w,
-                             "vsuxseg"#nf#"ei"#eew#".v">;
+                             "vsuxseg"#nf#"ei"#eew#".v">, VSUXSEGSched<nf, eew>;
       def VSOXSEG#nf#EI#eew#_V :
         VIndexedSegmentStore<!add(nf, -1), MOPSTIndexedOrder, w,
-                             "vsoxseg"#nf#"ei"#eew#".v">;
+                             "vsoxseg"#nf#"ei"#eew#".v">, VSOXSEGSched<nf, eew>;
     }
   }
 } // Predicates = [HasVInstructions]
@@ -1528,17 +1561,22 @@ let Predicates = [HasVInstructionsI64] in {
   foreach nf=2-8 in {
     // Vector Unit-strided Segment Instructions
     def VLSEG#nf#E64_V :
-      VUnitStrideSegmentLoad<!add(nf, -1), LSWidth64, "vlseg"#nf#"e64.v">;
+      VUnitStrideSegmentLoad<!add(nf, -1), LSWidth64, "vlseg"#nf#"e64.v">,
+      VLSEGSched<nf, 64>;
     def VLSEG#nf#E64FF_V :
-      VUnitStrideSegmentLoadFF<!add(nf, -1), LSWidth64, "vlseg"#nf#"e64ff.v">;
+      VUnitStrideSegmentLoadFF<!add(nf, -1), LSWidth64, "vlseg"#nf#"e64ff.v">,
+      VLSEGFFSched<nf, 64>;
     def VSSEG#nf#E64_V :
-      VUnitStrideSegmentStore<!add(nf, -1), LSWidth64, "vsseg"#nf#"e64.v">;
+      VUnitStrideSegmentStore<!add(nf, -1), LSWidth64, "vsseg"#nf#"e64.v">,
+      VSSEGSched<nf, 64>;
 
     // Vector Strided Segment Instructions
     def VLSSEG#nf#E64_V :
-      VStridedSegmentLoad<!add(nf, -1), LSWidth64, "vlsseg"#nf#"e64.v">;
+      VStridedSegmentLoad<!add(nf, -1), LSWidth64, "vlsseg"#nf#"e64.v">,
+      VLSSEGSched<nf, 64>;
     def VSSSEG#nf#E64_V :
-      VStridedSegmentStore<!add(nf, -1), LSWidth64, "vssseg"#nf#"e64.v">;
+      VStridedSegmentStore<!add(nf, -1), LSWidth64, "vssseg"#nf#"e64.v">,
+      VSSSEGSched<nf, 64>;
   }
 } // Predicates = [HasVInstructionsI64]
 let Predicates = [HasVInstructionsI64, IsRV64] in {
@@ -1546,16 +1584,16 @@ let Predicates = [HasVInstructionsI64, IsRV64] in {
     // Vector Indexed Segment Instructions
     def VLUXSEG#nf#EI64_V :
       VIndexedSegmentLoad<!add(nf, -1), MOPLDIndexedUnord, LSWidth64,
-                          "vluxseg"#nf#"ei64.v">;
+                          "vluxseg"#nf#"ei64.v">, VLUXSEGSched<nf, 64>;
     def VLOXSEG#nf#EI64_V :
       VIndexedSegmentLoad<!add(nf, -1), MOPLDIndexedOrder, LSWidth64,
-                          "vloxseg"#nf#"ei64.v">;
+                          "vloxseg"#nf#"ei64.v">, VLOXSEGSched<nf, 64>;
     def VSUXSEG#nf#EI64_V :
       VIndexedSegmentStore<!add(nf, -1), MOPSTIndexedUnord, LSWidth64,
-                           "vsuxseg"#nf#"ei64.v">;
+                           "vsuxseg"#nf#"ei64.v">, VSUXSEGSched<nf, 64>;
     def VSOXSEG#nf#EI64_V :
       VIndexedSegmentStore<!add(nf, -1), MOPSTIndexedOrder, LSWidth64,
-                           "vsoxseg"#nf#"ei64.v">;
+                           "vsoxseg"#nf#"ei64.v">, VSOXSEGSched<nf, 64>;
   }
 } // Predicates = [HasVInstructionsI64, IsRV64]
 
diff --git a/llvm/lib/Target/RISCV/RISCVScheduleV.td b/llvm/lib/Target/RISCV/RISCVScheduleV.td
index 43af1802d706..bafcf47b82e4 100644
--- a/llvm/lib/Target/RISCV/RISCVScheduleV.td
+++ b/llvm/lib/Target/RISCV/RISCVScheduleV.td
@@ -53,6 +53,20 @@ def WriteVLDFF8       : SchedWrite;
 def WriteVLDFF16      : SchedWrite;
 def WriteVLDFF32      : SchedWrite;
 def WriteVLDFF64      : SchedWrite;
+// 7.8. Vector Segment Instructions
+foreach nf=2-8 in {
+  foreach eew = [8, 16, 32, 64] in {
+    def WriteVLSEG # nf # e # eew : SchedWrite;
+    def WriteVSSEG # nf # e # eew : SchedWrite;
+    def WriteVLSEGFF # nf # e # eew : SchedWrite;
+    def WriteVLSSEG # nf # e # eew : SchedWrite;
+    def WriteVSSSEG # nf # e # eew : SchedWrite;
+    def WriteVLUXSEG # nf # e # eew : SchedWrite;
+    def WriteVLOXSEG # nf # e # eew : SchedWrite;
+    def WriteVSUXSEG # nf # e # eew : SchedWrite;
+    def WriteVSOXSEG # nf # e # eew : SchedWrite;
+  }
+}
 // 7.9. Vector Whole Register Instructions
 def WriteVLD1R8       : SchedWrite;
 def WriteVLD1R16      : SchedWrite;
@@ -538,6 +552,20 @@ def : WriteRes<WriteVST1R, []>;
 def : WriteRes<WriteVST2R, []>;
 def : WriteRes<WriteVST4R, []>;
 def : WriteRes<WriteVST8R, []>;
+// Vector Segment Loads and Stores
+foreach nf=2-8 in {
+  foreach eew = [8, 16, 32, 64] in {
+    def : WriteRes <!cast<SchedWrite>("WriteVLSEG" # nf # "e" # eew), []>;
+    def : WriteRes <!cast<SchedWrite>("WriteVLSEGFF" # nf # "e" # eew), []>;
+    def : WriteRes <!cast<SchedWrite>("WriteVSSEG" # nf # "e" # eew), []>;
+    def : WriteRes <!cast<SchedWrite>("WriteVLSSEG" # nf # "e" # eew), []>;
+    def : WriteRes <!cast<SchedWrite>("WriteVSSSEG" # nf # "e" # eew), []>;
+    def : WriteRes <!cast<SchedWrite>("WriteVLUXSEG" # nf # "e" # eew), []>;
+    def : WriteRes <!cast<SchedWrite>("WriteVLOXSEG" # nf # "e" # eew), []>;
+    def : WriteRes <!cast<SchedWrite>("WriteVSUXSEG" # nf # "e" # eew), []>;
+    def : WriteRes <!cast<SchedWrite>("WriteVSOXSEG" # nf # "e" # eew), []>;
+  }
+}
 
 // 12. Vector Integer Arithmetic Instructions
 def : WriteRes<WriteVIALUV, []>;