[RISCV][VLOPT] Add getOperandInfo for integer and floating point widening reductions (#122176)

Author: michaelmaitland

llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp

@@ -710,6 +710,19 @@ getOperandLog2EEW(const MachineOperand &MO, const MachineRegisterInfo *MRI) {
     return MILog2SEW;
   }
 
+  // Vector Widening Integer Reduction Instructions
+  // The Dest and VS1 read only element 0 for the vector register. Return
+  // 2*EEW for these. VS2 has EEW=SEW and EMUL=LMUL.
+  case RISCV::VWREDSUM_VS:
+  case RISCV::VWREDSUMU_VS:
+  // Vector Widening Floating-Point Reduction Instructions
+  case RISCV::VFWREDOSUM_VS:
+  case RISCV::VFWREDUSUM_VS: {
+    bool TwoTimes = IsMODef || MO.getOperandNo() == 3;
+    unsigned Log2EEW = TwoTimes ? MILog2SEW + 1 : MILog2SEW;
+    return Log2EEW;
+  }
+
   default:
     return std::nullopt;
   }
@@ -729,6 +742,8 @@ getOperandInfo(const MachineOperand &MO, const MachineRegisterInfo *MRI) {
   switch (RVV->BaseInstr) {
   // Vector Reduction Operations
   // Vector Single-Width Integer Reduction Instructions
+  // Vector Widening Integer Reduction Instructions
+  // Vector Widening Floating-Point Reduction Instructions
   // The Dest and VS1 only read element 0 of the vector register. Return just
   // the EEW for these.
   case RISCV::VREDAND_VS:
@@ -739,6 +754,10 @@ getOperandInfo(const MachineOperand &MO, const MachineRegisterInfo *MRI) {
   case RISCV::VREDOR_VS:
   case RISCV::VREDSUM_VS:
   case RISCV::VREDXOR_VS:
+  case RISCV::VWREDSUM_VS:
+  case RISCV::VWREDSUMU_VS:
+  case RISCV::VFWREDOSUM_VS:
+  case RISCV::VFWREDUSUM_VS:
     if (MO.getOperandNo() != 2)
       return OperandInfo(*Log2EEW);
     break;

llvm/test/CodeGen/RISCV/rvv/vl-opt-op-info.mir

@@ -1224,6 +1224,7 @@ body: |
     %x:vr = PseudoVMAND_MM_B1 $noreg, $noreg, -1, 0
     %y:vr = PseudoVIOTA_M_MF2 $noreg, %x, 1, 3 /* e8 */, 0
 ...
+---
 name: vred_vs2
 body: |
   bb.0:
@@ -1337,3 +1338,53 @@ body: |
     %y:vr = PseudoVREDSUM_VS_M1_E8 $noreg, $noreg, %x, 1, 3 /* e8 */, 0
     %z:vr = PseudoVADD_VV_M1 $noreg, %x, $noreg, 2, 3 /* e8 */, 0
 ...
+---
+name: vwred_vs2
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: vwred_vs2
+    ; CHECK: %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, 1, 4 /* e16 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %y:vr = PseudoVWREDSUM_VS_M1_E8 $noreg, $noreg, %x, 1, 3 /* e8 */, 0 /* tu, mu */
+    %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 4 /* e16 */, 0
+    %y:vr = PseudoVWREDSUM_VS_M1_E8 $noreg, $noreg, %x, 1, 3 /* e8 */, 0
+...
+---
+name: vwred_vs1
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: vwred_vs1
+    ; CHECK: %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, 1, 3 /* e8 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %y:vr = PseudoVWREDSUM_VS_M1_E8 $noreg, %x, $noreg, 1, 3 /* e8 */, 0 /* tu, mu */
+    %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 3 /* e8 */, 0
+    %y:vr = PseudoVWREDSUM_VS_M1_E8 $noreg, %x, $noreg, 1, 3 /* e8 */, 0
+...
+---
+name: vwred_vs1_incompatible_eew
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: vwred_vs1_incompatible_eew
+    ; CHECK: %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 4 /* e16 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %y:vr = PseudoVWREDSUM_VS_M1_E8 $noreg, %x, $noreg, 1, 3 /* e8 */, 0 /* tu, mu */
+    %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 4 /* e16 */, 0
+    %y:vr = PseudoVWREDSUM_VS_M1_E8 $noreg, %x, $noreg, 1, 3 /* e8 */, 0
+...
+---
+name: vwred_vs2_incompatible_eew
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: vwred_vs2_incompatible_eew
+    ; CHECK: %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 3 /* e8 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %y:vr = PseudoVWREDSUM_VS_M1_E8 $noreg, $noreg, %x, 1, 3 /* e8 */, 0 /* tu, mu */
+    %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 3 /* e8 */, 0
+    %y:vr = PseudoVWREDSUM_VS_M1_E8 $noreg, $noreg, %x, 1, 3 /* e8 */, 0
+...
+---
+name: vwred_incompatible_emul
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: vwred_incompatible_emul
+    ; CHECK: %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 3 /* e8 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %y:vr = PseudoVWREDSUM_VS_MF2_E8 $noreg, %x, $noreg, 1, 3 /* e8 */, 0 /* tu, mu */
+    %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 3 /* e8 */, 0
+    %y:vr = PseudoVWREDSUM_VS_MF2_E8 $noreg, %x, $noreg, 1, 3 /* e8 */, 0
+...