[Mesa-dev] [PATCH 1/6] radeon/llvm: interp instructions emits native outputs

[Vincent Lejeune]

---
 lib/Target/AMDGPU/R600ExpandSpecialInstrs.cpp | 30 +++++++++++----------------
 lib/Target/AMDGPU/R600Instructions.td         | 25 +++++-----------------
 lib/Target/AMDGPU/R600RegisterInfo.td         | 10 +++++++++
 3 files changed, 27 insertions(+), 38 deletions(-)

diff --git a/lib/Target/AMDGPU/R600ExpandSpecialInstrs.cpp b/lib/Target/AMDGPU/R600ExpandSpecialInstrs.cpp
index fabdb4d..f9fd65d 100644
--- a/lib/Target/AMDGPU/R600ExpandSpecialInstrs.cpp
+++ b/lib/Target/AMDGPU/R600ExpandSpecialInstrs.cpp
@@ -95,8 +95,9 @@ bool R600ExpandSpecialInstrsPass::ExpandInputPerspective(MachineInstr &MI)
   for (unsigned i = 0; i < 8; i++) {
     unsigned IJIndex = AMDGPU::R600_TReg32RegClass.getRegister(
         2 * IJIndexBase + ((i + 1) % 2));
-    unsigned ReadReg = AMDGPU::R600_TReg32RegClass.getRegister(
-        4 * MI.getOperand(2).getImm());
+    unsigned ReadReg = AMDGPU::R600_ArrayBaseRegClass.getRegister(
+        MI.getOperand(2).getImm());
+
 
     unsigned Sel;
     switch (i % 4) {
@@ -109,16 +110,11 @@ bool R600ExpandSpecialInstrsPass::ExpandInputPerspective(MachineInstr &MI)
 
     unsigned Res = TRI.getSubReg(DstReg, Sel);
 
-    const MCInstrDesc &Opcode = (i < 4)?
-        TII->get(AMDGPU::INTERP_ZW):
-        TII->get(AMDGPU::INTERP_XY);
+    unsigned Opcode = (i < 4)?AMDGPU::INTERP_ZW:AMDGPU::INTERP_XY;
 
-    MachineInstr *NewMI = BuildMI(*(MI.getParent()),
-        I, MI.getParent()->findDebugLoc(I),
-        Opcode, Res)
-        .addReg(IJIndex)
-        .addReg(ReadReg)
-        .addImm(0);
+    MachineBasicBlock &MBB = *(MI.getParent());
+    MachineInstr *NewMI =
+        TII->buildDefaultInstruction(MBB, I, Opcode, Res, IJIndex, ReadReg);
 
     if (!(i> 1 && i < 6)) {
       TII->addFlag(NewMI, 0, MO_FLAG_MASK);
@@ -143,8 +139,8 @@ bool R600ExpandSpecialInstrsPass::ExpandInputConstant(MachineInstr &MI)
   unsigned DstReg = MI.getOperand(0).getReg();
 
   for (unsigned i = 0; i < 4; i++) {
-    unsigned ReadReg = AMDGPU::R600_TReg32RegClass.getRegister(
-        4 * MI.getOperand(1).getImm() + i);
+    unsigned ReadReg = AMDGPU::R600_ArrayBaseRegClass.getRegister(
+        MI.getOperand(1).getImm());
 
     unsigned Sel;
     switch (i % 4) {
@@ -157,11 +153,9 @@ bool R600ExpandSpecialInstrsPass::ExpandInputConstant(MachineInstr &MI)
 
     unsigned Res = TRI.getSubReg(DstReg, Sel);
 
-    MachineInstr *NewMI = BuildMI(*(MI.getParent()),
-        I, MI.getParent()->findDebugLoc(I),
-        TII->get(AMDGPU::INTERP_LOAD_P0), Res)
-        .addReg(ReadReg)
-        .addImm(0);
+    MachineBasicBlock &MBB = *(MI.getParent());
+    MachineInstr *NewMI = TII->buildDefaultInstruction(
+        MBB, I, AMDGPU::INTERP_LOAD_P0, Res, ReadReg);
 
     if (i % 4 !=  3)
       TII->addFlag(NewMI, 0, MO_FLAG_NOT_LAST);
diff --git a/lib/Target/AMDGPU/R600Instructions.td b/lib/Target/AMDGPU/R600Instructions.td
index 7cc74e8..b97b094 100644
--- a/lib/Target/AMDGPU/R600Instructions.td
+++ b/lib/Target/AMDGPU/R600Instructions.td
@@ -444,32 +444,17 @@ def input_constant :  AMDGPUShaderInst <
 
 
 
-def INTERP_XY : InstR600 <0xD6,
-  (outs R600_Reg32:$dst),
-  (ins R600_Reg32:$src0, R600_Reg32:$src1, i32imm:$flags),
-  "INTERP_XY dst",
-  [], AnyALU>
+def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []>
 {
-  let FlagOperandIdx = 3;
+  let bank_swizzle = 5;
 }
 
-def INTERP_ZW : InstR600 <0xD7,
-  (outs R600_Reg32:$dst),
-  (ins R600_Reg32:$src0, R600_Reg32:$src1, i32imm:$flags),
-  "INTERP_ZW dst",
-  [], AnyALU>
+def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []>
 {
-  let FlagOperandIdx = 3;
+  let bank_swizzle = 5;
 }
 
-def INTERP_LOAD_P0 : InstR600 <0xE0,
-  (outs R600_Reg32:$dst),
-  (ins R600_Reg32:$src, i32imm:$flags),
-  "INTERP_LOAD_P0 dst",
-  [], AnyALU>
-{
-  let FlagOperandIdx = 2;
-}
+def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
 
 let Predicates = [isR600toCayman] in { 
 
diff --git a/lib/Target/AMDGPU/R600RegisterInfo.td b/lib/Target/AMDGPU/R600RegisterInfo.td
index c682f2b..d3d6d25 100644
--- a/lib/Target/AMDGPU/R600RegisterInfo.td
+++ b/lib/Target/AMDGPU/R600RegisterInfo.td
@@ -41,6 +41,12 @@ foreach Index = 0-127 in {
                                    Index>;
 }
 
+// Array Base Register holding input in FS
+foreach Index = 448-464 in {
+  def ArrayBase#Index :  R600Reg<"ARRAY_BASE", Index>;
+}
+
+
 // Special Registers
 
 def ZERO : R600Reg<"0.0", 248>;
@@ -56,6 +62,9 @@ def PRED_SEL_OFF: R600Reg<"Pred_sel_off", 0>;
 def PRED_SEL_ZERO : R600Reg<"Pred_sel_zero", 2>;
 def PRED_SEL_ONE : R600Reg<"Pred_sel_one", 3>;
 
+def R600_ArrayBase : RegisterClass <"AMDGPU", [f32, i32], 32,
+                          (add (sequence "ArrayBase%u", 448, 464))>;
+
 def R600_CReg32 : RegisterClass <"AMDGPU", [f32, i32], 32,
                           (add (interleave
                                   (interleave (sequence "C%u_X", 0, 127),
@@ -83,6 +92,7 @@ def R600_TReg32 : RegisterClass <"AMDGPU", [f32, i32], 32,
 def R600_Reg32 : RegisterClass <"AMDGPU", [f32, i32], 32, (add
     R600_TReg32,
     R600_CReg32,
+    R600_ArrayBase,
     ZERO, HALF, ONE, ONE_INT, PV_X, ALU_LITERAL_X, NEG_ONE, NEG_HALF)>;
 
 def R600_Predicate : RegisterClass <"AMDGPU", [i32], 32, (add
-- 
1.7.11.7