[RISCV] Discard the false operand in vmerge.vvm -> vmv.v.v peephole

[lukel97]

vmerge.vvm needs to have an all ones mask, so nothing is taken from the false operand. So instead of checking that the passthru is the same as false, just use the passthru directly for the tail elements.

This supersedes the convertVMergeToVMv part of #105788, as noted in https://github.com/llvm/llvm-project/pull/105788/files#r1731683971

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Author: Luke Lau (lukel97)

Changes

vmerge.vvm needs to have an all ones mask, so nothing is taken from the false operand. So instead of checking that the passthru is the same as false, just use the passthru directly for the tail elements.

This supersedes the convertVMergeToVMv part of #105788, as noted in https://github.com/llvm/llvm-project/pull/105788/files#r1731683971

---

Full diff: https://github.com/llvm/llvm-project/pull/106688.diff

2 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVVectorPeephole.cpp (+5-13)
• (modified) llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-to-vmv.mir (+5-5)

diff --git a/llvm/lib/Target/RISCV/RISCVVectorPeephole.cpp b/llvm/lib/Target/RISCV/RISCVVectorPeephole.cpp
index 34e5d9224f7150..6e5e9297bf29db 100644
--- a/llvm/lib/Target/RISCV/RISCVVectorPeephole.cpp
+++ b/llvm/lib/Target/RISCV/RISCVVectorPeephole.cpp
@@ -329,8 +329,8 @@ bool RISCVVectorPeephole::convertToWholeRegister(MachineInstr &MI) const {
   return true;
 }
 
-// Transform (VMERGE_VVM_<LMUL> false, false, true, allones, vl, sew) to
-// (VMV_V_V_<LMUL> false, true, vl, sew). It may decrease uses of VMSET.
+// Transform (VMERGE_VVM_<LMUL> pt, false, true, allones, vl, sew) to
+// (VMV_V_V_<LMUL> pt, true, vl, sew). It may decrease uses of VMSET.
 bool RISCVVectorPeephole::convertVMergeToVMv(MachineInstr &MI) const {
 #define CASE_VMERGE_TO_VMV(lmul)                                               \
   case RISCV::PseudoVMERGE_VVM_##lmul:                                         \
@@ -349,21 +349,12 @@ bool RISCVVectorPeephole::convertVMergeToVMv(MachineInstr &MI) const {
     CASE_VMERGE_TO_VMV(M8)
   }
 
-  Register PassthruReg = MI.getOperand(1).getReg();
-  Register FalseReg = MI.getOperand(2).getReg();
-  // Check passthru == false (or passthru == undef)
-  if (PassthruReg != RISCV::NoRegister &&
-      TRI->lookThruCopyLike(PassthruReg, MRI) !=
-          TRI->lookThruCopyLike(FalseReg, MRI))
-    return false;
-
   assert(MI.getOperand(4).isReg() && MI.getOperand(4).getReg() == RISCV::V0);
   if (!isAllOnesMask(V0Defs.lookup(&MI)))
     return false;
 
   MI.setDesc(TII->get(NewOpc));
-  MI.removeOperand(1);  // Passthru operand
-  MI.tieOperands(0, 1); // Tie false to dest
+  MI.removeOperand(2);  // Passthru operand
   MI.removeOperand(3);  // Mask operand
   MI.addOperand(
       MachineOperand::CreateImm(RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED));
@@ -371,7 +362,8 @@ bool RISCVVectorPeephole::convertVMergeToVMv(MachineInstr &MI) const {
   // vmv.v.v doesn't have a mask operand, so we may be able to inflate the
   // register class for the destination and passthru operands e.g. VRNoV0 -> VR
   MRI->recomputeRegClass(MI.getOperand(0).getReg());
-  MRI->recomputeRegClass(MI.getOperand(1).getReg());
+  if (MI.getOperand(1).getReg() != RISCV::NoRegister)
+    MRI->recomputeRegClass(MI.getOperand(1).getReg());
   return true;
 }
 
diff --git a/llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-to-vmv.mir b/llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-to-vmv.mir
index 01fff3de0aa8bd..1419eede6ca9d1 100644
--- a/llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-to-vmv.mir
+++ b/llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-to-vmv.mir
@@ -15,7 +15,7 @@ body: |
     ; CHECK-NEXT: %avl:gprnox0 = COPY $x1
     ; CHECK-NEXT: %mask:vmv0 = PseudoVMSET_M_B8 %avl, 5 /* e32 */
     ; CHECK-NEXT: $v0 = COPY %mask
-    ; CHECK-NEXT: %x:vr = PseudoVMV_V_V_M1 %false, %true, %avl, 5 /* e32 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %x:vr = PseudoVMV_V_V_M1 $noreg, %true, %avl, 5 /* e32 */, 0 /* tu, mu */
     %false:vr = COPY $v8
     %true:vr = COPY $v9
     %avl:gprnox0 = COPY $x1
@@ -31,13 +31,13 @@ body: |
     ; CHECK-LABEL: name: undef_false
     ; CHECK: liveins: $x1, $v8, $v9
     ; CHECK-NEXT: {{  $}}
-    ; CHECK-NEXT: %pt:vrnov0 = COPY $v8
+    ; CHECK-NEXT: %pt:vr = COPY $v8
     ; CHECK-NEXT: %false:vr = COPY $noreg
     ; CHECK-NEXT: %true:vr = COPY $v9
     ; CHECK-NEXT: %avl:gprnox0 = COPY $x1
     ; CHECK-NEXT: %mask:vmv0 = PseudoVMSET_M_B8 %avl, 5 /* e32 */
     ; CHECK-NEXT: $v0 = COPY %mask
-    ; CHECK-NEXT: %x:vrnov0 = PseudoVMERGE_VVM_M1 %pt, %false, %true, $v0, %avl, 5 /* e32 */
+    ; CHECK-NEXT: %x:vr = PseudoVMV_V_V_M1 %pt, %true, %avl, 5 /* e32 */, 0 /* tu, mu */
     %pt:vrnov0 = COPY $v8
     %false:vr = COPY $noreg
     %true:vr = COPY $v9
@@ -55,12 +55,12 @@ body: |
     ; CHECK: liveins: $x1, $v8, $v9
     ; CHECK-NEXT: {{  $}}
     ; CHECK-NEXT: %false:vr = COPY $v8
-    ; CHECK-NEXT: %pt:vrnov0 = COPY $v8
+    ; CHECK-NEXT: %pt:vr = COPY $v8
     ; CHECK-NEXT: %true:vr = COPY $v9
     ; CHECK-NEXT: %avl:gprnox0 = COPY $x1
     ; CHECK-NEXT: %mask:vmv0 = PseudoVMSET_M_B8 %avl, 5 /* e32 */
     ; CHECK-NEXT: $v0 = COPY %mask
-    ; CHECK-NEXT: %x:vr = PseudoVMV_V_V_M1 %false, %true, %avl, 5 /* e32 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %x:vr = PseudoVMV_V_V_M1 %pt, %true, %avl, 5 /* e32 */, 0 /* tu, mu */
     %false:vr = COPY $v8
     %pt:vrnov0 = COPY $v8
     %true:vr = COPY $v9

[topperc]

LGTM