From bad4e7b748f615d19ea0b3b5beebd69decd24be3 Mon Sep 17 00:00:00 2001
From: Tom Stellard <thomas.stellard@amd.com>
Date: Fri, 21 Nov 2014 22:31:46 +0000
Subject: [PATCH] R600/SI: Add an s_mov_b32 to patterns which use the
 M0RegClass

We need to use a s_mov_b32 rather than a copy, so that CSE will
eliminate redundant moves to the m0 register.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222584 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/R600/SIInstrInfo.cpp   | 20 --------------------
 lib/Target/R600/SIInstructions.td | 12 ++++++++----
 2 files changed, 8 insertions(+), 24 deletions(-)

diff --git a/lib/Target/R600/SIInstrInfo.cpp b/lib/Target/R600/SIInstrInfo.cpp
index 8343362d116..1a0010c03dc 100644
--- a/lib/Target/R600/SIInstrInfo.cpp
+++ b/lib/Target/R600/SIInstrInfo.cpp
@@ -326,26 +326,6 @@ SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
   unsigned Opcode;
   const int16_t *SubIndices;
 
-  if (AMDGPU::M0 == DestReg) {
-    // Check if M0 isn't already set to this value
-    for (MachineBasicBlock::reverse_iterator E = MBB.rend(),
-      I = MachineBasicBlock::reverse_iterator(MI); I != E; ++I) {
-
-      if (!I->definesRegister(AMDGPU::M0))
-        continue;
-
-      unsigned Opc = I->getOpcode();
-      if (Opc != TargetOpcode::COPY && Opc != AMDGPU::S_MOV_B32)
-        break;
-
-      if (!I->readsRegister(SrcReg))
-        break;
-
-      // The copy isn't necessary
-      return;
-    }
-  }
-
   if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
     assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
     BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index e1eb95580ac..00ce9bfcc26 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -2499,17 +2499,21 @@ def : Pat <
 /********** Interpolation Paterns **********/
 /********** ===================== **********/
 
+// The value of $params is constant through out the entire kernel.
+// We need to use S_MOV_B32 $params, because CSE ignores copies, so
+// without it we end up with a lot of redundant moves.
+
 def : Pat <
   (int_SI_fs_constant imm:$attr_chan, imm:$attr, i32:$params),
-  (V_INTERP_MOV_F32 INTERP.P0, imm:$attr_chan, imm:$attr, $params)
+  (V_INTERP_MOV_F32 INTERP.P0, imm:$attr_chan, imm:$attr, (S_MOV_B32 $params))
 >;
 
 def : Pat <
-  (int_SI_fs_interp imm:$attr_chan, imm:$attr, M0Reg:$params, v2i32:$ij),
+  (int_SI_fs_interp imm:$attr_chan, imm:$attr, i32:$params, v2i32:$ij),
   (V_INTERP_P2_F32 (V_INTERP_P1_F32 (EXTRACT_SUBREG v2i32:$ij, sub0),
-                                    imm:$attr_chan, imm:$attr, i32:$params),
+                                    imm:$attr_chan, imm:$attr, (S_MOV_B32 $params)),
                    (EXTRACT_SUBREG $ij, sub1),
-                   imm:$attr_chan, imm:$attr, $params)
+                   imm:$attr_chan, imm:$attr, (S_MOV_B32 $params))
 >;
 
 /********** ================== **********/
-- 
2.34.1