[RISCV] Clear vill for whole vector register moves in vsetvli insertion

This is an alternative to llvm#117866 that works by demanding a valid vtype instead using a separate pass.

The main advantage of this is that it allows coalesceVSETVLIs to just reuse an existing vsetvli later in the block.

To do this we need to first transfer the vsetvli info to some arbitrary valid state in transferBefore when we encounter a vector copy. Then we add a new vill demanded field that will happily accept any other known vtype, which allows us to coalesce these where possible.

Note we also need to check for vector copies in computeVLVTYPEChanges, otherwise the pass will completely skip over functions that only have vector copies and nothing else.

This is one part of a fix for llvm#114518. We still need to check if there's other cases where vector copies/whole register moves that are inserted after vsetvli insertion.

llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

@@ -195,6 +195,27 @@ static bool hasUndefinedPassthru(const MachineInstr &MI) {
   return UseMO.getReg() == RISCV::NoRegister || UseMO.isUndef();
 }
 
+/// Return true if \p MI is a copy that will be lowered to one or more vmvNr.vs.
+static bool isVecCopy(const MachineInstr &MI) {
+  static const TargetRegisterClass *RVVRegClasses[] = {
+      &RISCV::VRRegClass,     &RISCV::VRM2RegClass,   &RISCV::VRM4RegClass,
+      &RISCV::VRM8RegClass,   &RISCV::VRN2M1RegClass, &RISCV::VRN2M2RegClass,
+      &RISCV::VRN2M4RegClass, &RISCV::VRN3M1RegClass, &RISCV::VRN3M2RegClass,
+      &RISCV::VRN4M1RegClass, &RISCV::VRN4M2RegClass, &RISCV::VRN5M1RegClass,
+      &RISCV::VRN6M1RegClass, &RISCV::VRN7M1RegClass, &RISCV::VRN8M1RegClass};
+  if (!MI.isCopy())
+    return false;
+
+  Register DstReg = MI.getOperand(0).getReg();
+  Register SrcReg = MI.getOperand(1).getReg();
+  for (const auto &RegClass : RVVRegClasses) {
+    if (RegClass->contains(DstReg, SrcReg)) {
+      return true;
+    }
+  }
+  return false;
+}
+
 /// Which subfields of VL or VTYPE have values we need to preserve?
 struct DemandedFields {
   // Some unknown property of VL is used.  If demanded, must preserve entire
@@ -221,10 +242,13 @@ struct DemandedFields {
   bool SEWLMULRatio = false;
   bool TailPolicy = false;
   bool MaskPolicy = false;
+  // If this is true, we demand that VTYPE is set to some legal state, i.e. that
+  // vill is unset.
+  bool VILL = false;
 
   // Return true if any part of VTYPE was used
   bool usedVTYPE() const {
-    return SEW || LMUL || SEWLMULRatio || TailPolicy || MaskPolicy;
+    return SEW || LMUL || SEWLMULRatio || TailPolicy || MaskPolicy || VILL;
   }
 
   // Return true if any property of VL was used
@@ -239,6 +263,7 @@ struct DemandedFields {
     SEWLMULRatio = true;
     TailPolicy = true;
     MaskPolicy = true;
+    VILL = true;
   }
 
   // Mark all VL properties as demanded
@@ -263,6 +288,7 @@ struct DemandedFields {
     SEWLMULRatio |= B.SEWLMULRatio;
     TailPolicy |= B.TailPolicy;
     MaskPolicy |= B.MaskPolicy;
+    VILL |= B.VILL;
   }
 
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
@@ -308,7 +334,8 @@ struct DemandedFields {
     OS << ", ";
     OS << "SEWLMULRatio=" << SEWLMULRatio << ", ";
     OS << "TailPolicy=" << TailPolicy << ", ";
-    OS << "MaskPolicy=" << MaskPolicy;
+    OS << "MaskPolicy=" << MaskPolicy << ", ";
+    OS << "VILL=" << VILL;
     OS << "}";
   }
 #endif
@@ -503,6 +530,16 @@ DemandedFields getDemanded(const MachineInstr &MI, const RISCVSubtarget *ST) {
     }
   }
 
+  // In §32.16.6, whole vector register moves have a dependency on SEW. At the
+  // MIR level though we don't encode the element type, and it gives the same
+  // result whatever the SEW may be.
+  //
+  // However it does need valid SEW, i.e. vill must be cleared. The entry to a
+  // function, calls and inline assembly may all set it, so make sure we clear
+  // it for whole register copies.
+  if (isVecCopy(MI))
+    Res.VILL = true;
+
   return Res;
 }
 
@@ -1208,6 +1245,17 @@ static VSETVLIInfo adjustIncoming(VSETVLIInfo PrevInfo, VSETVLIInfo NewInfo,
 // legal for MI, but may not be the state requested by MI.
 void RISCVInsertVSETVLI::transferBefore(VSETVLIInfo &Info,
                                         const MachineInstr &MI) const {
+  if (isVecCopy(MI) &&
+      (Info.isUnknown() || !Info.isValid() || Info.hasSEWLMULRatioOnly())) {
+    // Use an arbitrary but valid AVL and VTYPE so vill will be cleared. It may
+    // be coalesced into another vsetvli since we won't demand any fields.
+    VSETVLIInfo NewInfo; // Need a new VSETVLIInfo to clear SEWLMULRatioOnly
+    NewInfo.setAVLImm(0);
+    NewInfo.setVTYPE(RISCVII::VLMUL::LMUL_1, 8, true, true);
+    Info = NewInfo;
+    return;
+  }
+
   if (!RISCVII::hasSEWOp(MI.getDesc().TSFlags))
     return;
 
@@ -1296,7 +1344,8 @@ bool RISCVInsertVSETVLI::computeVLVTYPEChanges(const MachineBasicBlock &MBB,
   for (const MachineInstr &MI : MBB) {
     transferBefore(Info, MI);
 
-    if (isVectorConfigInstr(MI) || RISCVII::hasSEWOp(MI.getDesc().TSFlags))
+    if (isVectorConfigInstr(MI) || RISCVII::hasSEWOp(MI.getDesc().TSFlags) ||
+        isVecCopy(MI))
       HadVectorOp = true;
 
     transferAfter(Info, MI);
@@ -1426,6 +1475,12 @@ void RISCVInsertVSETVLI::emitVSETVLIs(MachineBasicBlock &MBB) {
       PrefixTransparent = false;
     }
 
+    if (isVecCopy(MI) &&
+        !PrevInfo.isCompatible(DemandedFields::all(), CurInfo, LIS)) {
+      insertVSETVLI(MBB, MI, MI.getDebugLoc(), CurInfo, PrevInfo);
+      PrefixTransparent = false;
+    }
+
     uint64_t TSFlags = MI.getDesc().TSFlags;
     if (RISCVII::hasSEWOp(TSFlags)) {
       if (!PrevInfo.isCompatible(DemandedFields::all(), CurInfo, LIS)) {

llvm/test/CodeGen/RISCV/inline-asm-v-constraint.ll

@@ -45,6 +45,7 @@ define <vscale x 1 x i8> @constraint_vd(<vscale x 1 x i8> %0, <vscale x 1 x i8>
 define <vscale x 1 x i1> @constraint_vm(<vscale x 1 x i1> %0, <vscale x 1 x i1> %1) nounwind {
 ; RV32I-LABEL: constraint_vm:
 ; RV32I:       # %bb.0:
+; RV32I-NEXT:    vsetivli zero, 0, e8, m1, ta, ma
 ; RV32I-NEXT:    vmv1r.v v9, v0
 ; RV32I-NEXT:    vmv1r.v v0, v8
 ; RV32I-NEXT:    #APP
@@ -54,6 +55,7 @@ define <vscale x 1 x i1> @constraint_vm(<vscale x 1 x i1> %0, <vscale x 1 x i1>
 ;
 ; RV64I-LABEL: constraint_vm:
 ; RV64I:       # %bb.0:
+; RV64I-NEXT:    vsetivli zero, 0, e8, m1, ta, ma
 ; RV64I-NEXT:    vmv1r.v v9, v0
 ; RV64I-NEXT:    vmv1r.v v0, v8
 ; RV64I-NEXT:    #APP

llvm/test/CodeGen/RISCV/rvv/abs-vp.ll

@@ -567,6 +567,7 @@ define <vscale x 16 x i64> @vp_abs_nxv16i64(<vscale x 16 x i64> %va, <vscale x 1
 ; CHECK-NEXT:    slli a1, a1, 4
 ; CHECK-NEXT:    sub sp, sp, a1
 ; CHECK-NEXT:    .cfi_escape 0x0f, 0x0d, 0x72, 0x00, 0x11, 0x10, 0x22, 0x11, 0x10, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 16 + 16 * vlenb
+; CHECK-NEXT:    vsetvli a1, zero, e8, mf4, ta, ma
 ; CHECK-NEXT:    vmv1r.v v24, v0
 ; CHECK-NEXT:    csrr a1, vlenb
 ; CHECK-NEXT:    slli a1, a1, 3
@@ -576,7 +577,6 @@ define <vscale x 16 x i64> @vp_abs_nxv16i64(<vscale x 16 x i64> %va, <vscale x 1
 ; CHECK-NEXT:    csrr a1, vlenb
 ; CHECK-NEXT:    srli a2, a1, 3
 ; CHECK-NEXT:    sub a3, a0, a1
-; CHECK-NEXT:    vsetvli a4, zero, e8, mf4, ta, ma
 ; CHECK-NEXT:    vslidedown.vx v0, v0, a2
 ; CHECK-NEXT:    sltu a2, a0, a3
 ; CHECK-NEXT:    addi a2, a2, -1

llvm/test/CodeGen/RISCV/rvv/bitreverse-vp.ll

@@ -3075,6 +3075,7 @@ define <vscale x 64 x i16> @vp_bitreverse_nxv64i16(<vscale x 64 x i16> %va, <vsc
 ; CHECK-NEXT:    slli a1, a1, 4
 ; CHECK-NEXT:    sub sp, sp, a1
 ; CHECK-NEXT:    .cfi_escape 0x0f, 0x0d, 0x72, 0x00, 0x11, 0x10, 0x22, 0x11, 0x10, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 16 + 16 * vlenb
+; CHECK-NEXT:    vsetvli a1, zero, e8, m1, ta, ma
 ; CHECK-NEXT:    vmv1r.v v24, v0
 ; CHECK-NEXT:    csrr a1, vlenb
 ; CHECK-NEXT:    slli a1, a1, 3
@@ -3086,7 +3087,6 @@ define <vscale x 64 x i16> @vp_bitreverse_nxv64i16(<vscale x 64 x i16> %va, <vsc
 ; CHECK-NEXT:    lui a2, 3
 ; CHECK-NEXT:    srli a4, a3, 1
 ; CHECK-NEXT:    slli a3, a3, 2
-; CHECK-NEXT:    vsetvli a5, zero, e8, m1, ta, ma
 ; CHECK-NEXT:    vslidedown.vx v0, v0, a4
 ; CHECK-NEXT:    sub a4, a0, a3
 ; CHECK-NEXT:    sltu a5, a0, a4
@@ -3158,11 +3158,11 @@ define <vscale x 64 x i16> @vp_bitreverse_nxv64i16(<vscale x 64 x i16> %va, <vsc
 ;
 ; CHECK-ZVBB-LABEL: vp_bitreverse_nxv64i16:
 ; CHECK-ZVBB:       # %bb.0:
+; CHECK-ZVBB-NEXT:    vsetvli a1, zero, e8, m1, ta, ma
 ; CHECK-ZVBB-NEXT:    vmv1r.v v24, v0
 ; CHECK-ZVBB-NEXT:    csrr a1, vlenb
 ; CHECK-ZVBB-NEXT:    srli a2, a1, 1
 ; CHECK-ZVBB-NEXT:    slli a1, a1, 2
-; CHECK-ZVBB-NEXT:    vsetvli a3, zero, e8, m1, ta, ma
 ; CHECK-ZVBB-NEXT:    vslidedown.vx v0, v0, a2
 ; CHECK-ZVBB-NEXT:    sub a2, a0, a1
 ; CHECK-ZVBB-NEXT:    sltu a3, a0, a2

llvm/test/CodeGen/RISCV/rvv/bswap-vp.ll

@@ -1584,6 +1584,7 @@ define <vscale x 64 x i16> @vp_bswap_nxv64i16(<vscale x 64 x i16> %va, <vscale x
 ; CHECK-NEXT:    slli a1, a1, 4
 ; CHECK-NEXT:    sub sp, sp, a1
 ; CHECK-NEXT:    .cfi_escape 0x0f, 0x0d, 0x72, 0x00, 0x11, 0x10, 0x22, 0x11, 0x10, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 16 + 16 * vlenb
+; CHECK-NEXT:    vsetvli a1, zero, e8, m1, ta, ma
 ; CHECK-NEXT:    vmv1r.v v24, v0
 ; CHECK-NEXT:    csrr a1, vlenb
 ; CHECK-NEXT:    slli a1, a1, 3
@@ -1593,7 +1594,6 @@ define <vscale x 64 x i16> @vp_bswap_nxv64i16(<vscale x 64 x i16> %va, <vscale x
 ; CHECK-NEXT:    csrr a1, vlenb
 ; CHECK-NEXT:    srli a2, a1, 1
 ; CHECK-NEXT:    slli a1, a1, 2
-; CHECK-NEXT:    vsetvli a3, zero, e8, m1, ta, ma
 ; CHECK-NEXT:    vslidedown.vx v0, v0, a2
 ; CHECK-NEXT:    sub a2, a0, a1
 ; CHECK-NEXT:    sltu a3, a0, a2
@@ -1631,11 +1631,11 @@ define <vscale x 64 x i16> @vp_bswap_nxv64i16(<vscale x 64 x i16> %va, <vscale x
 ;
 ; CHECK-ZVKB-LABEL: vp_bswap_nxv64i16:
 ; CHECK-ZVKB:       # %bb.0:
+; CHECK-ZVKB-NEXT:    vsetvli a1, zero, e8, m1, ta, ma
 ; CHECK-ZVKB-NEXT:    vmv1r.v v24, v0
 ; CHECK-ZVKB-NEXT:    csrr a1, vlenb
 ; CHECK-ZVKB-NEXT:    srli a2, a1, 1
 ; CHECK-ZVKB-NEXT:    slli a1, a1, 2
-; CHECK-ZVKB-NEXT:    vsetvli a3, zero, e8, m1, ta, ma
 ; CHECK-ZVKB-NEXT:    vslidedown.vx v0, v0, a2
 ; CHECK-ZVKB-NEXT:    sub a2, a0, a1
 ; CHECK-ZVKB-NEXT:    sltu a3, a0, a2

llvm/test/CodeGen/RISCV/rvv/calling-conv-fastcc.ll

@@ -336,6 +336,7 @@ define fastcc <vscale x 32 x i32> @ret_nxv32i32_call_nxv32i32_nxv32i32_i32(<vsca
 ; RV32-NEXT:    add a1, a3, a1
 ; RV32-NEXT:    li a3, 2
 ; RV32-NEXT:    vs8r.v v16, (a1)
+; RV32-NEXT:    vsetivli zero, 0, e8, m1, ta, ma
 ; RV32-NEXT:    vmv8r.v v8, v0
 ; RV32-NEXT:    vmv8r.v v16, v24
 ; RV32-NEXT:    call ext2
@@ -374,6 +375,7 @@ define fastcc <vscale x 32 x i32> @ret_nxv32i32_call_nxv32i32_nxv32i32_i32(<vsca
 ; RV64-NEXT:    add a1, a3, a1
 ; RV64-NEXT:    li a3, 2
 ; RV64-NEXT:    vs8r.v v16, (a1)
+; RV64-NEXT:    vsetivli zero, 0, e8, m1, ta, ma
 ; RV64-NEXT:    vmv8r.v v8, v0
 ; RV64-NEXT:    vmv8r.v v16, v24
 ; RV64-NEXT:    call ext2
@@ -451,6 +453,7 @@ define fastcc <vscale x 32 x i32> @ret_nxv32i32_call_nxv32i32_nxv32i32_nxv32i32_
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 128
 ; RV32-NEXT:    vl8r.v v8, (a1) # Unknown-size Folded Reload
+; RV32-NEXT:    vsetivli zero, 0, e8, m1, ta, ma
 ; RV32-NEXT:    vmv8r.v v16, v0
 ; RV32-NEXT:    call ext3
 ; RV32-NEXT:    addi sp, s0, -144
@@ -523,6 +526,7 @@ define fastcc <vscale x 32 x i32> @ret_nxv32i32_call_nxv32i32_nxv32i32_nxv32i32_
 ; RV64-NEXT:    add a1, sp, a1
 ; RV64-NEXT:    addi a1, a1, 128
 ; RV64-NEXT:    vl8r.v v8, (a1) # Unknown-size Folded Reload
+; RV64-NEXT:    vsetivli zero, 0, e8, m1, ta, ma
 ; RV64-NEXT:    vmv8r.v v16, v0
 ; RV64-NEXT:    call ext3
 ; RV64-NEXT:    addi sp, s0, -144

llvm/test/CodeGen/RISCV/rvv/calling-conv.ll

@@ -103,6 +103,7 @@ define target("riscv.vector.tuple", <vscale x 16 x i8>, 2) @caller_tuple_return(
 ; RV32-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
 ; RV32-NEXT:    .cfi_offset ra, -4
 ; RV32-NEXT:    call callee_tuple_return
+; RV32-NEXT:    vsetivli zero, 0, e8, m1, ta, ma
 ; RV32-NEXT:    vmv2r.v v6, v8
 ; RV32-NEXT:    vmv2r.v v8, v10
 ; RV32-NEXT:    vmv2r.v v10, v6
@@ -119,6 +120,7 @@ define target("riscv.vector.tuple", <vscale x 16 x i8>, 2) @caller_tuple_return(
 ; RV64-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
 ; RV64-NEXT:    .cfi_offset ra, -8
 ; RV64-NEXT:    call callee_tuple_return
+; RV64-NEXT:    vsetivli zero, 0, e8, m1, ta, ma
 ; RV64-NEXT:    vmv2r.v v6, v8
 ; RV64-NEXT:    vmv2r.v v8, v10
 ; RV64-NEXT:    vmv2r.v v10, v6
@@ -144,6 +146,7 @@ define void @caller_tuple_argument(target("riscv.vector.tuple", <vscale x 16 x i
 ; RV32-NEXT:    .cfi_def_cfa_offset 16
 ; RV32-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
 ; RV32-NEXT:    .cfi_offset ra, -4
+; RV32-NEXT:    vsetivli zero, 0, e8, m1, ta, ma
 ; RV32-NEXT:    vmv2r.v v6, v8
 ; RV32-NEXT:    vmv2r.v v8, v10
 ; RV32-NEXT:    vmv2r.v v10, v6
@@ -160,6 +163,7 @@ define void @caller_tuple_argument(target("riscv.vector.tuple", <vscale x 16 x i
 ; RV64-NEXT:    .cfi_def_cfa_offset 16
 ; RV64-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
 ; RV64-NEXT:    .cfi_offset ra, -8
+; RV64-NEXT:    vsetivli zero, 0, e8, m1, ta, ma
 ; RV64-NEXT:    vmv2r.v v6, v8
 ; RV64-NEXT:    vmv2r.v v8, v10
 ; RV64-NEXT:    vmv2r.v v10, v6