[RISCV] Teach vsetvli insertion to remember when predecessors have same AVL and SEW/LMUL ratio if their VTYPEs otherwise mismatch.

[craig.topper]

Summary

Previously we went directly to unknown state on VTYPE mismatch.
If we instead remember the partial match, we can use this to
still use X0, X0 vsetvli in successors if AVL and needed SEW/LMUL
ratio match.

[craig.topper]

Fixed typo in comment

[kito-cheng]

llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

65

Need init SEWLMULRatioOnly here.

[craig.topper]

llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

65

Thanks!

[craig.topper]

Initialize member variable

[frasercrmck]

llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

227

I'm wondering if removing this use of the == operator is a sign we shouldn't have one for these types. Is it confusing to have different meaning of equality? Alternatively perhaps this operator is now unused so can be axed without much discussion.

llvm/test/CodeGen/RISCV/rvv/vsetvli-insert-crossbb.ll

501

Nit but this comment feels like it's missing something in The AVL is the same the SEW/LMUL ratio. A comma, or an and?

[craig.topper]

Address review feedback

[frasercrmck]

LGTM, thanks!

llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

class VSETVLIInfo {
   } State = Uninitialized;
 
   // Fields from VTYPE.
   RISCVII::VLMUL VLMul = RISCVII::LMUL_1;
   uint8_t SEW = 0;
   uint8_t TailAgnostic : 1;
   uint8_t MaskAgnostic : 1;
   uint8_t MaskRegOp : 1;
+  uint8_t SEWLMULRatioOnly : 1;
 
 public:
   VSETVLIInfo()
       : AVLImm(0), TailAgnostic(false), MaskAgnostic(false), MaskRegOp(false) {}
 
   static VSETVLIInfo getUnknown() {
     VSETVLIInfo Info;
     Info.setUnknown();
     return Info;
   }
 
   bool isValid() const { return State != Uninitialized; }
void setVTYPE(RISCVII::VLMUL L, unsigned S, bool TA, bool MA, bool MRO) {
     VLMul = L;
     SEW = S;
     TailAgnostic = TA;
     MaskAgnostic = MA;
     MaskRegOp = MRO;
   }
 
   unsigned encodeVTYPE() const {
-    assert(isValid() && !isUnknown() &&
+    assert(isValid() && !isUnknown() && !SEWLMULRatioOnly &&
            "Can't encode VTYPE for uninitialized or unknown");
     return RISCVVType::encodeVTYPE(VLMul, SEW, TailAgnostic, MaskAgnostic);
   }
 
+  bool hasSEWLMULRatioOnly() const { return SEWLMULRatioOnly; }
+
   bool hasSameVTYPE(const VSETVLIInfo &Other) const {
     assert(isValid() && Other.isValid() &&
            "Can't compare invalid VSETVLIInfos");
     assert(!isUnknown() && !Other.isUnknown() &&
            "Can't compare VTYPE in unknown state");
+    assert(!SEWLMULRatioOnly && !Other.SEWLMULRatioOnly &&
+           "Can't compare when only LMUL/SEW ratio is valid.");
     return std::tie(VLMul, SEW, TailAgnostic, MaskAgnostic) ==
            std::tie(Other.VLMul, Other.SEW, Other.TailAgnostic,
                     Other.MaskAgnostic);
   }
 
   // Convert VLMUL to a fixed point value with 3 bits of fraction.
   unsigned getSEWLMULRatio() const {
     assert(isValid() && !isUnknown() &&
public:
   }
 
   // Determine whether the vector instructions requirements represented by
   // InstrInfo are compatible with the previous vsetvli instruction represented
   // by this.
   bool isCompatible(const VSETVLIInfo &InstrInfo) const {
     assert(isValid() && InstrInfo.isValid() &&
            "Can't compare invalid VSETVLIInfos");
+    assert(!InstrInfo.SEWLMULRatioOnly &&
+           "Expected a valid VTYPE for instruction!");
     // Nothing is compatible with Unknown.
     if (isUnknown() || InstrInfo.isUnknown())
       return false;
 
+    // If only our VLMAX ratio is valid, then this isn't compatible.
+    if (SEWLMULRatioOnly)
+      return false;
+
     // If the instruction doesn't need an AVLReg and the SEW matches, consider
     // it/ compatible.
     if (InstrInfo.hasAVLReg() && InstrInfo.AVLReg == RISCV::NoRegister) {
       if (SEW == InstrInfo.SEW)
         return true;
     }
 
     // VTypes must match unless the instruction is a mask reg operation, then it
if (!Other.isValid())
       return !isValid();
 
     // Unknown is only equal to another Unknown.
     if (isUnknown())
       return Other.isUnknown();
     if (Other.isUnknown())
       return isUnknown();
 
-    // Otherwise compare the VTYPE and AVL.
-    return hasSameVTYPE(Other) && hasSameAVL(Other);
+    if (!hasSameAVL(Other))
+      return false;
+
+    // If only the VLMAX is valid, check that it is the same.
+    if (SEWLMULRatioOnly && Other.SEWLMULRatioOnly)
+      return hasSameVLMAX(Other);
+
+    // If the full VTYPE is valid, check that it is the same.
+    if (!SEWLMULRatioOnly && !Other.SEWLMULRatioOnly)
+      return hasSameVTYPE(Other);
+
+    // If the VLMaxOnly bits different, then they aren't equal.
+    return false;
   }
 
   // Calculate the VSETVLIInfo visible to a block assuming this and Other are
   // both predecessors.
   VSETVLIInfo intersect(const VSETVLIInfo &Other) const {
     // If the new value isn't valid, ignore it.
     if (!Other.isValid())
       return *this;
 
     // If this value isn't valid, this must be the first predecessor, use it.
     if (!isValid())
       return Other;
 
-    if (*this == Other)
+    // If either is unknown, the result is unknown.
+    if (isUnknown() || Other.isUnknown())
+      return VSETVLIInfo::getUnknown();
+
+    // If the AVL doesn't match, the result is unknown.
+    if (!hasSameAVL(Other))
+      return VSETVLIInfo::getUnknown();
+
+    // If the VTYPEs are valid and match, return this.
+    if (!SEWLMULRatioOnly && !Other.SEWLMULRatioOnly && hasSameVTYPE(Other))
       return *this;
 
-    // If the configurations don't match, assume unknown.
+    // Otherwise see if the SEW/LMUL ratio matches and return a SEW/LMUL ratio
+    // only value.
+    if (hasSameVLMAX(Other)) {
+      VSETVLIInfo MergeInfo = *this;
+      MergeInfo.SEWLMULRatioOnly = true;
+      return MergeInfo;
+    }
+
+    // Otherwise the result is unknown.
     return VSETVLIInfo::getUnknown();
   }
 
   // Calculate the VSETVLIInfo visible at the end of the block assuming this
   // is the predecessor value, and Other is change for this block.
   VSETVLIInfo merge(const VSETVLIInfo &Other) const {
     assert(isValid() && "Can only merge with a valid VSETVLInfo");
 
bool RISCVInsertVSETVLI::needVSETVLI(const VSETVLIInfo &Require,
   if (CurInfo.isCompatible(Require))
     return false;
 
   // We didn't find a compatible value. If our AVL is a virtual register,
   // it might be defined by a VSET(I)VLI. If it has the same VTYPE we need
   // and the last VL/VTYPE we observed is the same, we don't need a
   // VSETVLI here.
   if (!CurInfo.isUnknown() && Require.hasAVLReg() &&
-      Require.getAVLReg().isVirtual() && Require.hasSameVTYPE(CurInfo)) {
+      Require.getAVLReg().isVirtual() && !CurInfo.hasSEWLMULRatioOnly() &&
+      Require.hasSameVTYPE(CurInfo)) {
     if (MachineInstr *DefMI = MRI->getVRegDef(Require.getAVLReg())) {
       if (DefMI->getOpcode() == RISCV::PseudoVSETVLI ||
           DefMI->getOpcode() == RISCV::PseudoVSETIVLI) {
         VSETVLIInfo DefInfo = getInfoForVSETVLI(*DefMI);
         if (DefInfo.hasSameAVL(CurInfo) && DefInfo.hasSameVTYPE(CurInfo))
           return false;
       }
     }

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-unaligned.ll

 ; RV32-NEXT:    slli a3, a4, 8
 ; RV32-NEXT:    or a0, a3, a0
 ; RV32-NEXT:    slli a0, a0, 16
 ; RV32-NEXT:    or a0, a0, a2
 ; RV32-NEXT:    vmv.s.x v26, a0
 ; RV32-NEXT:    vsetvli zero, zero, e32, mf2, tu, mu
 ; RV32-NEXT:    vslideup.vi v25, v26, 1
 ; RV32-NEXT:  .LBB8_4: # %else2
-; RV32-NEXT:    vsetivli zero, 2, e32, mf2, ta, mu
+; RV32-NEXT:    vsetvli zero, zero, e32, mf2, ta, mu
 ; RV32-NEXT:    vse32.v v25, (a1)
 ; RV32-NEXT:    addi sp, sp, 16
 ; RV32-NEXT:    ret
 ;
 ; RV64-LABEL: masked_load_v2i32_align1:
 ; RV64:       # %bb.0:
 ; RV64-NEXT:    addi sp, sp, -16
 ; RV64-NEXT:    vsetivli zero, 2, e32, mf2, ta, mu

 ; RV64-NEXT:    slli a3, a4, 8
 ; RV64-NEXT:    or a0, a3, a0
 ; RV64-NEXT:    slli a0, a0, 16
 ; RV64-NEXT:    or a0, a0, a2
 ; RV64-NEXT:    vmv.s.x v26, a0
 ; RV64-NEXT:    vsetvli zero, zero, e32, mf2, tu, mu
 ; RV64-NEXT:    vslideup.vi v25, v26, 1
 ; RV64-NEXT:  .LBB8_4: # %else2
-; RV64-NEXT:    vsetivli zero, 2, e32, mf2, ta, mu
+; RV64-NEXT:    vsetvli zero, zero, e32, mf2, ta, mu
 ; RV64-NEXT:    vse32.v v25, (a1)
 ; RV64-NEXT:    addi sp, sp, 16
 ; RV64-NEXT:    ret
   %mask = icmp eq <2 x i32> %m, zeroinitializer
   %load = call <2 x i32> @llvm.masked.load.v2i32(<2 x i32>* %a, i32 1, <2 x i1> %mask, <2 x i32> undef)
   store <2 x i32> %load, <2 x i32>* %res_ptr
   ret void
 }

llvm/test/CodeGen/RISCV/rvv/vsetvli-insert-crossbb.ll

 for.end:                                          ; preds = %for.body, %entry
   ret void
 }
 
 declare i64 @llvm.riscv.vsetvli.i64(i64, i64 immarg, i64 immarg)
 declare <vscale x 16 x float> @llvm.riscv.vle.nxv16f32.i64(<vscale x 16 x float>* nocapture, i64)
 declare <vscale x 16 x float> @llvm.riscv.vfmacc.nxv16f32.f32.i64(<vscale x 16 x float>, float, <vscale x 16 x float>, i64)
 declare void @llvm.riscv.vse.nxv16f32.i64(<vscale x 16 x float>, <vscale x 16 x float>* nocapture, i64)
+
+; We need a vsetvli in the last block because the predecessors have different
+; VTYPEs. The AVL is the same the SEW/LMUL ratio implies the same VLMAX so
+; we don't need to read AVL and can keep VL unchanged.
+define <vscale x 2 x i32> @test_vsetvli_x0_x0(<vscale x 2 x i32>* %x, <vscale x 2 x i16>* %y, <vscale x 2 x i32> %z, i64 %vl, i1 %cond) nounwind {
+; CHECK-LABEL: test_vsetvli_x0_x0:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a2, e32, m1, ta, mu
+; CHECK-NEXT:    vle32.v v25, (a0)
+; CHECK-NEXT:    andi a0, a3, 1
+; CHECK-NEXT:    beqz a0, .LBB9_2
+; CHECK-NEXT:  # %bb.1: # %if
+; CHECK-NEXT:    vsetvli zero, zero, e16, mf2, ta, mu
+; CHECK-NEXT:    vle16.v v26, (a1)
+; CHECK-NEXT:    vwadd.vx v8, v26, zero
+; CHECK-NEXT:  .LBB9_2: # %if.end
+; CHECK-NEXT:    vsetvli zero, zero, e32, m1, ta, mu
+; CHECK-NEXT:    vadd.vv v8, v25, v8
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 2 x i32> @llvm.riscv.vle.nxv2i32(<vscale x 2 x i32>* %x, i64 %vl)
+  br i1 %cond, label %if, label %if.end
+
+if:
+  %b = call <vscale x 2 x i16> @llvm.riscv.vle.nxv2i16(<vscale x 2 x i16>* %y, i64 %vl)
+  %c = call <vscale x 2 x i32> @llvm.riscv.vwadd.nxv2i32(<vscale x 2 x i16> %b, i16 0, i64 %vl)
+  br label %if.end
+
+if.end:
+  %d = phi <vscale x 2 x i32> [ %z, %entry ], [ %c, %if ]
+  %e = call <vscale x 2 x i32> @llvm.riscv.vadd.nxv2i32(<vscale x 2 x i32> %a, <vscale x 2 x i32> %d, i64 %vl)
+  ret <vscale x 2 x i32> %e
+}
+declare <vscale x 2 x i32> @llvm.riscv.vle.nxv2i32(<vscale x 2 x i32>*, i64)
+declare <vscale x 2 x i16> @llvm.riscv.vle.nxv2i16(<vscale x 2 x i16>*, i64)
+declare <vscale x 2 x i32> @llvm.riscv.vwadd.nxv2i32(<vscale x 2 x i16>, i16, i64)
+declare <vscale x 2 x i32> @llvm.riscv.vadd.nxv2i32(<vscale x 2 x i32>, <vscale x 2 x i32>, i64)
+
+; We can use X0, X0 vsetvli in if2 and if2.end. The merge point as if.end will
+; see two different vtypes with the same SEW/LMUL ratio. At if2.end we will only
+; know the SEW/LMUL ratio for the if.end predecessor and the full vtype for
+; the if2 predecessor. This makes sure we can merge a SEW/LMUL predecessor with
+; a predecessor we know the vtype for.
+define <vscale x 2 x i32> @test_vsetvli_x0_x0_2(<vscale x 2 x i32>* %x, <vscale x 2 x i16>* %y, <vscale x 2 x i16>* %z, i64 %vl, i1 %cond, i1 %cond2, <vscale x 2 x i32> %w) nounwind {
+; CHECK-LABEL: test_vsetvli_x0_x0_2:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a3, e32, m1, ta, mu
+; CHECK-NEXT:    vle32.v v25, (a0)
+; CHECK-NEXT:    andi a0, a4, 1
+; CHECK-NEXT:    beqz a0, .LBB10_2
+; CHECK-NEXT:  # %bb.1: # %if
+; CHECK-NEXT:    vsetvli zero, zero, e16, mf2, ta, mu
+; CHECK-NEXT:    vle16.v v26, (a1)
+; CHECK-NEXT:    vwadd.wv v25, v25, v26
+; CHECK-NEXT:  .LBB10_2: # %if.end
+; CHECK-NEXT:    andi a0, a5, 1
+; CHECK-NEXT:    beqz a0, .LBB10_4
+; CHECK-NEXT:  # %bb.3: # %if2
+; CHECK-NEXT:    vsetvli zero, zero, e16, mf2, ta, mu
+; CHECK-NEXT:    vle16.v v26, (a2)
+; CHECK-NEXT:    vwadd.wv v25, v25, v26
+; CHECK-NEXT:  .LBB10_4: # %if2.end
+; CHECK-NEXT:    vsetvli zero, zero, e32, m1, ta, mu
+; CHECK-NEXT:    vadd.vv v8, v25, v8
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 2 x i32> @llvm.riscv.vle.nxv2i32(<vscale x 2 x i32>* %x, i64 %vl)
+  br i1 %cond, label %if, label %if.end
+
+if:
+  %b = call <vscale x 2 x i16> @llvm.riscv.vle.nxv2i16(<vscale x 2 x i16>* %y, i64 %vl)
+  %c = call <vscale x 2 x i32> @llvm.riscv.vwadd.w.nxv2i32.nxv2i16(<vscale x 2 x i32> %a, <vscale x 2 x i16> %b, i64 %vl)
+  br label %if.end
+
+if.end:
+  %d = phi <vscale x 2 x i32> [ %a, %entry ], [ %c, %if ]
+  br i1 %cond2, label %if2, label %if2.end
+
+if2:
+  %e = call <vscale x 2 x i16> @llvm.riscv.vle.nxv2i16(<vscale x 2 x i16>* %z, i64 %vl)
+  %f = call <vscale x 2 x i32> @llvm.riscv.vwadd.w.nxv2i32.nxv2i16(<vscale x 2 x i32> %d, <vscale x 2 x i16> %e, i64 %vl)
+  br label %if2.end
+
+if2.end:
+  %g = phi <vscale x 2 x i32> [ %d, %if.end ], [ %f, %if2 ]
+  %h = call <vscale x 2 x i32> @llvm.riscv.vadd.nxv2i32(<vscale x 2 x i32> %g, <vscale x 2 x i32> %w, i64 %vl)
+  ret <vscale x 2 x i32> %h
+}
+declare <vscale x 2 x i32> @llvm.riscv.vwadd.w.nxv2i32.nxv2i16(<vscale x 2 x i32>, <vscale x 2 x i16>, i64)