[RISCV] Allow vsetvlis with same register AVL in doLocalPostpass (llv…

…m#76801)
XinWang10 · Jan 11, 2024 · e879002 · e879002
1 parent be76f16
commit e879002
Show file tree

Hide file tree

Showing 3 changed files with 17 additions and 29 deletions.
diff --git a/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp b/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
@@ -1464,20 +1464,6 @@ static void doUnion(DemandedFields &A, DemandedFields B) {
   A.MaskPolicy |= B.MaskPolicy;
 }
 
-static bool isNonZeroAVL(const MachineOperand &MO,
-                         const MachineRegisterInfo &MRI) {
-  if (MO.isReg()) {
-    if (MO.getReg() == RISCV::X0)
-      return true;
-    if (MachineInstr *MI = MRI.getVRegDef(MO.getReg());
-        MI && isNonZeroLoadImmediate(*MI))
-      return true;
-    return false;
-  }
-  assert(MO.isImm());
-  return 0 != MO.getImm();
-}
-
 // Return true if we can mutate PrevMI to match MI without changing any the
 // fields which would be observed.
 static bool canMutatePriorConfig(const MachineInstr &PrevMI,
@@ -1491,21 +1477,26 @@ static bool canMutatePriorConfig(const MachineInstr &PrevMI,
     if (Used.VLAny)
       return false;
 
-    // We don't bother to handle the equally zero case here as it's largely
-    // uninteresting.
     if (Used.VLZeroness) {
       if (isVLPreservingConfig(PrevMI))
         return false;
-      if (!isNonZeroAVL(MI.getOperand(1), MRI) ||
-          !isNonZeroAVL(PrevMI.getOperand(1), MRI))
+      if (!getInfoForVSETVLI(PrevMI).hasEquallyZeroAVL(getInfoForVSETVLI(MI),
+                                                       MRI))
         return false;
     }
 
-    // TODO: Track whether the register is defined between
-    // PrevMI and MI.
-    if (MI.getOperand(1).isReg() &&
-        RISCV::X0 != MI.getOperand(1).getReg())
-      return false;
+    auto &AVL = MI.getOperand(1);
+    auto &PrevAVL = PrevMI.getOperand(1);
+    assert(MRI.isSSA());
+
+    // If the AVL is a register, we need to make sure MI's AVL dominates PrevMI.
+    // For now just check that PrevMI uses the same virtual register.
+    if (AVL.isReg() && AVL.getReg() != RISCV::X0) {
+      if (AVL.getReg().isPhysical())
+        return false;
+      if (!PrevAVL.isReg() || PrevAVL.getReg() != AVL.getReg())
+        return false;
+    }
   }
 
   if (!PrevMI.getOperand(2).isImm() || !MI.getOperand(2).isImm())

diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-insert.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-insert.ll
@@ -63,9 +63,8 @@ define <32 x i32> @insertelt_v32i32_31(<32 x i32> %a, i32 %y) {
 ; CHECK-LABEL: insertelt_v32i32_31:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    li a1, 32
-; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, ma
-; CHECK-NEXT:    vmv.s.x v16, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e32, m8, ta, ma
+; CHECK-NEXT:    vmv.s.x v16, a0
 ; CHECK-NEXT:    vslideup.vi v8, v16, 31
 ; CHECK-NEXT:    ret
   %b = insertelement <32 x i32> %a, i32 %y, i32 31
@@ -101,9 +100,8 @@ define <64 x i32> @insertelt_v64i32_63(<64 x i32> %a, i32 %y) {
 ; CHECK-LABEL: insertelt_v64i32_63:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    li a1, 32
-; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, ma
-; CHECK-NEXT:    vmv.s.x v24, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e32, m8, ta, ma
+; CHECK-NEXT:    vmv.s.x v24, a0
 ; CHECK-NEXT:    vslideup.vi v16, v24, 31
 ; CHECK-NEXT:    ret
   %b = insertelement <64 x i32> %a, i32 %y, i32 63

diff --git a/llvm/test/CodeGen/RISCV/rvv/vsetvli-insert.ll b/llvm/test/CodeGen/RISCV/rvv/vsetvli-insert.ll
@@ -643,9 +643,8 @@ define <vscale x 2 x float> @fp_reduction_vfmv_s_f(float %0, <vscale x 8 x float
 define dso_local <vscale x 2 x i32> @int_reduction_vmv_s_x(i32 signext %0, <vscale x 8 x i32> %1, i64 %2) {
 ; CHECK-LABEL: int_reduction_vmv_s_x:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, ma
-; CHECK-NEXT:    vmv.s.x v12, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e32, m4, ta, ma
+; CHECK-NEXT:    vmv.s.x v12, a0
 ; CHECK-NEXT:    vredsum.vs v8, v8, v12
 ; CHECK-NEXT:    ret
   %4 = tail call <vscale x 8 x i32> @llvm.riscv.vmv.s.x.nxv8i32.i64(<vscale x 8 x i32> poison, i32 %0, i64 %2)