AMDGPU: Use explicit register size indirect pseudos

[arsenm]

Summary

This stops using an unknown reg class operand.

Currently build_vector selection has a broken looking check
where it tries to use a VGPR reg class and an SGPR one if it
sees an SGPR use.

With the source operand has an explicit VGPR class,
illegal copies will be inserted that SIFixSGPRCopies will take care
of normally later, which will allow removing the weird check
of build_vector users. Without this, when removed v_movrels_b32 would
still be emitted even though all of the values were only stored in
SGPRs.

[• tstellarAMD]

LGTM.

[arsenm]

r249494

lib/Target/AMDGPU/SIInstrInfo.cpp

MachineInstrBuilder SIInstrInfo::buildIndirectRead(
                                    MachineBasicBlock *MBB,
                                    MachineBasicBlock::iterator I,
                                    unsigned ValueReg,
                                    unsigned Address, unsigned OffsetReg) const {
   const DebugLoc &DL = MBB->findDebugLoc(I);
   unsigned IndirectBaseReg = AMDGPU::VGPR_32RegClass.getRegister(
                                       getIndirectIndexBegin(*MBB->getParent()));
 
-  return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
+  return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC_V1))
           .addOperand(I->getOperand(0))
           .addOperand(I->getOperand(1))
           .addReg(IndirectBaseReg)
           .addReg(OffsetReg)
           .addImm(0);
 
 }
 

lib/Target/AMDGPU/SIInstructions.td

 let usesCustomInserter = 1 in {
 def SI_RegisterStorePseudo : SIRegStore<(outs)>;
 } // End usesCustomInserter = 1
 def SI_RegisterStore : SIRegStore<(outs SReg_64:$temp)>;
 
 
 } // End UseNamedOperandTable = 1
 
-def SI_INDIRECT_SRC : InstSI <
+class SI_INDIRECT_SRC<RegisterClass rc> : InstSI <
   (outs VGPR_32:$dst, SReg_64:$temp),
-  (ins unknown:$src, VSrc_32:$idx, i32imm:$off),
+  (ins rc:$src, VSrc_32:$idx, i32imm:$off),
   "si_indirect_src $dst, $temp, $src, $idx, $off",
   []
 >;
 
 class SI_INDIRECT_DST<RegisterClass rc> : InstSI <
   (outs rc:$dst, SReg_64:$temp),
   (ins unknown:$src, VSrc_32:$idx, i32imm:$off, VGPR_32:$val),
   "si_indirect_dst $dst, $temp, $src, $idx, $off, $val",
   []
 > {
   let Constraints = "$src = $dst";
 }
 
+// TODO: We can support indirect SGPR access.
+def SI_INDIRECT_SRC_V1 : SI_INDIRECT_SRC<VGPR_32>;
+def SI_INDIRECT_SRC_V2 : SI_INDIRECT_SRC<VReg_64>;
+def SI_INDIRECT_SRC_V4 : SI_INDIRECT_SRC<VReg_128>;
+def SI_INDIRECT_SRC_V8 : SI_INDIRECT_SRC<VReg_256>;
+def SI_INDIRECT_SRC_V16 : SI_INDIRECT_SRC<VReg_512>;
+
 def SI_INDIRECT_DST_V1 : SI_INDIRECT_DST<VGPR_32>;
 def SI_INDIRECT_DST_V2 : SI_INDIRECT_DST<VReg_64>;
 def SI_INDIRECT_DST_V4 : SI_INDIRECT_DST<VReg_128>;
 def SI_INDIRECT_DST_V8 : SI_INDIRECT_DST<VReg_256>;
 def SI_INDIRECT_DST_V16 : SI_INDIRECT_DST<VReg_512>;
 
 } // Uses = [EXEC,VCC,M0], Defs = [EXEC,VCC,M0]
 

 def : MTBUF_StoreResource <v2i32, 2, TBUFFER_STORE_FORMAT_XY>;
 def : MTBUF_StoreResource <v4i32, 3, TBUFFER_STORE_FORMAT_XYZ>;
 def : MTBUF_StoreResource <v4i32, 4, TBUFFER_STORE_FORMAT_XYZW>;
 
 /********** ====================== **********/
 /**********   Indirect adressing   **********/
 /********** ====================== **********/
 
-multiclass SI_INDIRECT_Pattern <ValueType vt, ValueType eltvt, SI_INDIRECT_DST IndDst> {
+multiclass SI_INDIRECT_Pattern <ValueType vt, ValueType eltvt, string VecSize> {
 
   // 1. Extract with offset
   def : Pat<
     (eltvt (vector_extract vt:$vec, (add i32:$idx, imm:$off))),
-    (SI_INDIRECT_SRC $vec, $idx, imm:$off)
+    (!cast<Instruction>("SI_INDIRECT_SRC_"#VecSize) $vec, $idx, imm:$off)
   >;
 
   // 2. Extract without offset
   def : Pat<
     (eltvt (vector_extract vt:$vec, i32:$idx)),
-    (SI_INDIRECT_SRC $vec, $idx, 0)
+    (!cast<Instruction>("SI_INDIRECT_SRC_"#VecSize) $vec, $idx, 0)
   >;
 
   // 3. Insert with offset
   def : Pat<
     (vector_insert vt:$vec, eltvt:$val, (add i32:$idx, imm:$off)),
-    (IndDst $vec, $idx, imm:$off, $val)
+    (!cast<Instruction>("SI_INDIRECT_DST_"#VecSize) $vec, $idx, imm:$off, $val)
   >;
 
   // 4. Insert without offset
   def : Pat<
     (vector_insert vt:$vec, eltvt:$val, i32:$idx),
-    (IndDst $vec, $idx, 0, $val)
+    (!cast<Instruction>("SI_INDIRECT_DST_"#VecSize) $vec, $idx, 0, $val)
   >;
 }
 
-defm : SI_INDIRECT_Pattern <v2f32, f32, SI_INDIRECT_DST_V2>;
-defm : SI_INDIRECT_Pattern <v4f32, f32, SI_INDIRECT_DST_V4>;
-defm : SI_INDIRECT_Pattern <v8f32, f32, SI_INDIRECT_DST_V8>;
-defm : SI_INDIRECT_Pattern <v16f32, f32, SI_INDIRECT_DST_V16>;
+defm : SI_INDIRECT_Pattern <v2f32, f32, "V2">;
+defm : SI_INDIRECT_Pattern <v4f32, f32, "V4">;
+defm : SI_INDIRECT_Pattern <v8f32, f32, "V8">;
+defm : SI_INDIRECT_Pattern <v16f32, f32, "V16">;
 
-defm : SI_INDIRECT_Pattern <v2i32, i32, SI_INDIRECT_DST_V2>;
-defm : SI_INDIRECT_Pattern <v4i32, i32, SI_INDIRECT_DST_V4>;
-defm : SI_INDIRECT_Pattern <v8i32, i32, SI_INDIRECT_DST_V8>;
-defm : SI_INDIRECT_Pattern <v16i32, i32, SI_INDIRECT_DST_V16>;
+defm : SI_INDIRECT_Pattern <v2i32, i32, "V2">;
+defm : SI_INDIRECT_Pattern <v4i32, i32, "V4">;
+defm : SI_INDIRECT_Pattern <v8i32, i32, "V8">;
+defm : SI_INDIRECT_Pattern <v16i32, i32, "V16">;
 
 //===----------------------------------------------------------------------===//
 // Conversion Patterns
 //===----------------------------------------------------------------------===//
 
 def : Pat<(i32 (sext_inreg i32:$src, i1)),
   (S_BFE_I32 i32:$src, 65536)>; // 0 | 1 << 16
 

lib/Target/AMDGPU/SILowerControlFlow.cpp

for (I = MBB.begin(); I != MBB.end(); I = Next) {
             HaveKill = true;
           Kill(MI);
           break;
 
         case AMDGPU::S_BRANCH:
           Branch(MI);
           break;
 
-        case AMDGPU::SI_INDIRECT_SRC:
+        case AMDGPU::SI_INDIRECT_SRC_V1:
+        case AMDGPU::SI_INDIRECT_SRC_V2:
+        case AMDGPU::SI_INDIRECT_SRC_V4:
+        case AMDGPU::SI_INDIRECT_SRC_V8:
+        case AMDGPU::SI_INDIRECT_SRC_V16:
           IndirectSrc(MI);
           break;
 
         case AMDGPU::SI_INDIRECT_DST_V1:
         case AMDGPU::SI_INDIRECT_DST_V2:
         case AMDGPU::SI_INDIRECT_DST_V4:
         case AMDGPU::SI_INDIRECT_DST_V8:
         case AMDGPU::SI_INDIRECT_DST_V16:

test/CodeGen/AMDGPU/indirect-addressing-si.ll

 ; RUN: llc < %s -march=amdgcn -mcpu=SI -verify-machineinstrs | FileCheck %s
 ; RUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck %s
 
 ; Tests for indirect addressing on SI, which is implemented using dynamic
 ; indexing of vectors.
 
 ; CHECK-LABEL: {{^}}extract_w_offset:
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 4.0
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 0x40400000
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 2.0
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 1.0
 ; CHECK: s_mov_b32 m0
 ; CHECK-NEXT: v_movrels_b32_e32
 define void @extract_w_offset(float addrspace(1)* %out, i32 %in) {
 entry:
-  %0 = add i32 %in, 1
-  %1 = extractelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, i32 %0
-  store float %1, float addrspace(1)* %out
+  %idx = add i32 %in, 1
+  %elt = extractelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, i32 %idx
+  store float %elt, float addrspace(1)* %out
+  ret void
+}
+
+; XXX: Could do v_or_b32 directly
+; CHECK-LABEL: {{^}}extract_w_offset_salu_use_vector:
+; CHECK-DAG: s_or_b32
+; CHECK-DAG: s_or_b32
+; CHECK-DAG: s_or_b32
+; CHECK-DAG: s_or_b32
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}}
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}}
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}}
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}}
+; CHECK: s_mov_b32 m0
+; CHECK-NEXT: v_movrels_b32_e32
+define void @extract_w_offset_salu_use_vector(i32 addrspace(1)* %out, i32 %in, <4 x i32> %or.val) {
+entry:
+  %idx = add i32 %in, 1
+  %vec = or <4 x i32> %or.val, <i32 1, i32 2, i32 3, i32 4>
+  %elt = extractelement <4 x i32> %vec, i32 %idx
+  store i32 %elt, i32 addrspace(1)* %out
   ret void
 }
 
 ; CHECK-LABEL: {{^}}extract_wo_offset:
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 4.0
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 0x40400000
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 2.0
+; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 1.0
 ; CHECK: s_mov_b32 m0
 ; CHECK-NEXT: v_movrels_b32_e32
 define void @extract_wo_offset(float addrspace(1)* %out, i32 %in) {
 entry:
-  %0 = extractelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, i32 %in
-  store float %0, float addrspace(1)* %out
+  %elt = extractelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, i32 %in
+  store float %elt, float addrspace(1)* %out
   ret void
 }
 
 ; CHECK-LABEL: {{^}}extract_neg_offset_sgpr:
 ; The offset depends on the register that holds the first element of the vector.
 ; CHECK: s_add_i32 m0, s{{[0-9]+}}, 0xfffffe{{[0-9a-z]+}}
 ; CHECK: v_movrels_b32_e32 v{{[0-9]}}, v0
 define void @extract_neg_offset_sgpr(i32 addrspace(1)* %out, i32 %offset) {
 entry:
   %index = add i32 %offset, -512
   %value = extractelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 %index
   store i32 %value, i32 addrspace(1)* %out
   ret void
 }
 
+; CHECK-LABEL: {{^}}extract_neg_offset_sgpr_loaded:
+; The offset depends on the register that holds the first element of the vector.
+; CHECK: s_add_i32 m0, s{{[0-9]+}}, 0xfffffe{{[0-9a-z]+}}
+; CHECK: v_movrels_b32_e32 v{{[0-9]}}, v0
+define void @extract_neg_offset_sgpr_loaded(i32 addrspace(1)* %out, <4 x i32> %vec0, <4 x i32> %vec1, i32 %offset) {
+entry:
+  %index = add i32 %offset, -512
+  %or = or <4 x i32> %vec0, %vec1
+  %value = extractelement <4 x i32> %or, i32 %index
+  store i32 %value, i32 addrspace(1)* %out
+  ret void
+}
+
 ; CHECK-LABEL: {{^}}extract_neg_offset_vgpr:
 ; The offset depends on the register that holds the first element of the vector.
 ; CHECK: v_readfirstlane_b32
 ; CHECK: s_add_i32 m0, m0, 0xfffffe{{[0-9a-z]+}}
 ; CHECK-NEXT: v_movrels_b32_e32 v{{[0-9]}}, v0
 ; CHECK: s_cbranch_execnz
 define void @extract_neg_offset_vgpr(i32 addrspace(1)* %out) {
 entry:

 define void @insert_neg_offset_sgpr(i32 addrspace(1)* %in, <4 x i32> addrspace(1)* %out, i32 %offset) {
 entry:
   %index = add i32 %offset, -512
   %value = insertelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 5, i32 %index
   store <4 x i32> %value, <4 x i32> addrspace(1)* %out
   ret void
 }
 
+; The vector indexed into is originally loaded into an SGPR rather
+; than built with a reg_sequence
+
+; CHECK-LABEL: {{^}}insert_neg_offset_sgpr_loadreg:
+; The offset depends on the register that holds the first element of the vector.
+; CHECK: s_add_i32 m0, s{{[0-9]+}}, 0xfffffe{{[0-9a-z]+}}
+; CHECK: v_movreld_b32_e32 v0, v{{[0-9]}}
+define void @insert_neg_offset_sgpr_loadreg(i32 addrspace(1)* %in, <4 x i32> addrspace(1)* %out, <4 x i32> %vec, i32 %offset) {
+entry:
+  %index = add i32 %offset, -512
+  %value = insertelement <4 x i32> %vec, i32 5, i32 %index
+  store <4 x i32> %value, <4 x i32> addrspace(1)* %out
+  ret void
+}
+
 ; CHECK-LABEL: {{^}}insert_neg_offset_vgpr:
 ; The offset depends on the register that holds the first element of the vector.
 ; CHECK: v_readfirstlane_b32
 ; CHECK: s_add_i32 m0, m0, 0xfffffe{{[0-9a-z]+}}
 ; CHECK-NEXT: v_movreld_b32_e32 v0, v{{[0-9]}}
 ; CHECK: s_cbranch_execnz
 define void @insert_neg_offset_vgpr(i32 addrspace(1)* %in, <4 x i32> addrspace(1)* %out) {
 entry: