[TableGen][RISCV] Relax a restriction in generating patterns for commutable SDNodes.

[craig.topper]

Summary

Previously, all children would be checked to see if any were an
explicit Register. If anywhere no commutable patterns would be
generated. This patch loosens the restriction to only check the
children that are being commuted.

Digging back through history, this code predates the existence of
commutable intrinsics and commutable SDNodes with more than 2
operands. At that time the loop would count the number of children that
weren't registers and if that was equal to 2 it would allow commuting.
I don't think this loop was re-considered when commutable
intrinsics were added or when we allowed SDNodes with more than 2
operands.

This important for RISCV were our isel patterns have a V0 mask
operand after the commutable operands on some RISCVISD opcodes.

[frasercrmck]

Ah great, I was going to do this today! My local patch wasn't as ambitious and didn't merge the intrinsic/non-intrinsic paths though. Makes sense to me. I didn't get around to doing any diffs on the generated tables - does anything but RISCV change?

llvm/utils/TableGen/CodeGenDAGPatterns.cpp

4648–4649

Several lines here could potentially be simplified if we had a unsigned offset = isCommIntrinsic ? 1 : 0 or somesuch. Then we could assert(N->getNumChilden() >= 2 + offset), i = offset, e = 2 + offset?

[craig.topper]

Simplify code as suggested.

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vadd-vp.ll

; CHECK-NEXT:    ret
   %vb = shufflevector <4 x i8> %elt.head, <4 x i8> poison, <4 x i32> zeroinitializer
   %v = call <4 x i8> @llvm.vp.add.v4i8(<4 x i8> %va, <4 x i8> %vb, <4 x i1> %m, i32 %evl)
   ret <4 x i8> %v
 }
 
 define <4 x i8> @vadd_vx_v4i8_commute(<4 x i8> %va, i8 %b, <4 x i1> %m, i32 zeroext %evl) {
 ; CHECK-LABEL: vadd_vx_v4i8_commute:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetivli zero, 4, e8, mf4, ta, mu
-; CHECK-NEXT:    vmv.v.x v9, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e8, mf4, ta, mu
-; CHECK-NEXT:    vadd.vv v8, v9, v8, v0.t
+; CHECK-NEXT:    vadd.vx v8, v8, a0, v0.t
 ; CHECK-NEXT:    ret
   %elt.head = insertelement <4 x i8> poison, i8 %b, i32 0
   %vb = shufflevector <4 x i8> %elt.head, <4 x i8> poison, <4 x i32> zeroinitializer
   %v = call <4 x i8> @llvm.vp.add.v4i8(<4 x i8> %vb, <4 x i8> %va, <4 x i1> %m, i32 %evl)
   ret <4 x i8> %v
 }
 
 define <4 x i8> @vadd_vx_v4i8_unmasked(<4 x i8> %va, i8 %b, i32 zeroext %evl) {

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vand-vp.ll

; CHECK-NEXT:    ret
   %vb = shufflevector <2 x i8> %elt.head, <2 x i8> poison, <2 x i32> zeroinitializer
   %v = call <2 x i8> @llvm.vp.and.v2i8(<2 x i8> %va, <2 x i8> %vb, <2 x i1> %m, i32 %evl)
   ret <2 x i8> %v
 }
 
 define <2 x i8> @vand_vx_v2i8_commute(<2 x i8> %va, i8 %b, <2 x i1> %m, i32 zeroext %evl) {
 ; CHECK-LABEL: vand_vx_v2i8_commute:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetivli zero, 2, e8, mf8, ta, mu
-; CHECK-NEXT:    vmv.v.x v9, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e8, mf8, ta, mu
-; CHECK-NEXT:    vand.vv v8, v9, v8, v0.t
+; CHECK-NEXT:    vand.vx v8, v8, a0, v0.t
 ; CHECK-NEXT:    ret
   %elt.head = insertelement <2 x i8> poison, i8 %b, i32 0
   %vb = shufflevector <2 x i8> %elt.head, <2 x i8> poison, <2 x i32> zeroinitializer
   %v = call <2 x i8> @llvm.vp.and.v2i8(<2 x i8> %vb, <2 x i8> %va, <2 x i1> %m, i32 %evl)
   ret <2 x i8> %v
 }
 
 define <2 x i8> @vand_vx_v2i8_unmasked(<2 x i8> %va, i8 %b, i32 zeroext %evl) {

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vfadd-vp.ll

; CHECK-NEXT:    ret
   %vb = shufflevector <2 x float> %elt.head, <2 x float> poison, <2 x i32> zeroinitializer
   %v = call <2 x float> @llvm.vp.fadd.v2f32(<2 x float> %va, <2 x float> %vb, <2 x i1> %m, i32 %evl)
   ret <2 x float> %v
 }
 
 define <2 x float> @vfadd_vf_v2f32_commute(<2 x float> %va, float %b, <2 x i1> %m, i32 zeroext %evl) {
 ; CHECK-LABEL: vfadd_vf_v2f32_commute:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetivli zero, 2, e32, mf2, ta, mu
-; CHECK-NEXT:    vfmv.v.f v9, fa0
 ; CHECK-NEXT:    vsetvli zero, a0, e32, mf2, ta, mu
-; CHECK-NEXT:    vfadd.vv v8, v9, v8, v0.t
+; CHECK-NEXT:    vfadd.vf v8, v8, fa0, v0.t
 ; CHECK-NEXT:    ret
   %elt.head = insertelement <2 x float> poison, float %b, i32 0
   %vb = shufflevector <2 x float> %elt.head, <2 x float> poison, <2 x i32> zeroinitializer
   %v = call <2 x float> @llvm.vp.fadd.v2f32(<2 x float> %vb, <2 x float> %va, <2 x i1> %m, i32 %evl)
   ret <2 x float> %v
 }
 
 define <2 x float> @vfadd_vf_v2f32_unmasked(<2 x float> %va, float %b, i32 zeroext %evl) {

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vmul-vp.ll

; CHECK-NEXT:    ret
   %vb = shufflevector <8 x i16> %elt.head, <8 x i16> poison, <8 x i32> zeroinitializer
   %v = call <8 x i16> @llvm.vp.mul.v8i16(<8 x i16> %va, <8 x i16> %vb, <8 x i1> %m, i32 %evl)
   ret <8 x i16> %v
 }
 
 define <8 x i16> @vmul_vx_v8i16_commute(<8 x i16> %va, i16 %b, <8 x i1> %m, i32 zeroext %evl) {
 ; CHECK-LABEL: vmul_vx_v8i16_commute:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetivli zero, 8, e16, m1, ta, mu
-; CHECK-NEXT:    vmv.v.x v9, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e16, m1, ta, mu
-; CHECK-NEXT:    vmul.vv v8, v9, v8, v0.t
+; CHECK-NEXT:    vmul.vx v8, v8, a0, v0.t
 ; CHECK-NEXT:    ret
   %elt.head = insertelement <8 x i16> poison, i16 %b, i32 0
   %vb = shufflevector <8 x i16> %elt.head, <8 x i16> poison, <8 x i32> zeroinitializer
   %v = call <8 x i16> @llvm.vp.mul.v8i16(<8 x i16> %vb, <8 x i16> %va, <8 x i1> %m, i32 %evl)
   ret <8 x i16> %v
 }
 
 define <8 x i16> @vmul_vx_v8i16_unmasked(<8 x i16> %va, i16 %b, i32 zeroext %evl) {

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vor-vp.ll

; CHECK-NEXT:    ret
   %vb = shufflevector <4 x i8> %elt.head, <4 x i8> poison, <4 x i32> zeroinitializer
   %v = call <4 x i8> @llvm.vp.or.v4i8(<4 x i8> %va, <4 x i8> %vb, <4 x i1> %m, i32 %evl)
   ret <4 x i8> %v
 }
 
 define <4 x i8> @vor_vx_v4i8_commute(<4 x i8> %va, i8 %b, <4 x i1> %m, i32 zeroext %evl) {
 ; CHECK-LABEL: vor_vx_v4i8_commute:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetivli zero, 4, e8, mf4, ta, mu
-; CHECK-NEXT:    vmv.v.x v9, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e8, mf4, ta, mu
-; CHECK-NEXT:    vor.vv v8, v9, v8, v0.t
+; CHECK-NEXT:    vor.vx v8, v8, a0, v0.t
 ; CHECK-NEXT:    ret
   %elt.head = insertelement <4 x i8> poison, i8 %b, i32 0
   %vb = shufflevector <4 x i8> %elt.head, <4 x i8> poison, <4 x i32> zeroinitializer
   %v = call <4 x i8> @llvm.vp.or.v4i8(<4 x i8> %vb, <4 x i8> %va, <4 x i1> %m, i32 %evl)
   ret <4 x i8> %v
 }
 
 define <4 x i8> @vor_vx_v4i8_unmasked(<4 x i8> %va, i8 %b, i32 zeroext %evl) {

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vxor-vp.ll

; CHECK-NEXT:    ret
   %vb = shufflevector <2 x i8> %elt.head, <2 x i8> poison, <2 x i32> zeroinitializer
   %v = call <2 x i8> @llvm.vp.xor.v2i8(<2 x i8> %va, <2 x i8> %vb, <2 x i1> %m, i32 %evl)
   ret <2 x i8> %v
 }
 
 define <2 x i8> @vxor_vx_v2i8_commute(<2 x i8> %va, i8 %b, <2 x i1> %m, i32 zeroext %evl) {
 ; CHECK-LABEL: vxor_vx_v2i8_commute:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetivli zero, 2, e8, mf8, ta, mu
-; CHECK-NEXT:    vmv.v.x v9, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e8, mf8, ta, mu
-; CHECK-NEXT:    vxor.vv v8, v9, v8, v0.t
+; CHECK-NEXT:    vxor.vx v8, v8, a0, v0.t
 ; CHECK-NEXT:    ret
   %elt.head = insertelement <2 x i8> poison, i8 %b, i32 0
   %vb = shufflevector <2 x i8> %elt.head, <2 x i8> poison, <2 x i32> zeroinitializer
   %v = call <2 x i8> @llvm.vp.xor.v2i8(<2 x i8> %vb, <2 x i8> %va, <2 x i1> %m, i32 %evl)
   ret <2 x i8> %v
 }
 
 define <2 x i8> @vxor_vx_v2i8_unmasked(<2 x i8> %va, i8 %b, i32 zeroext %evl) {

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

; CHECK-NEXT:    ret
   %vb = shufflevector <vscale x 1 x i8> %elt.head, <vscale x 1 x i8> poison, <vscale x 1 x i32> zeroinitializer
   %v = call <vscale x 1 x i8> @llvm.vp.add.nxv1i8(<vscale x 1 x i8> %va, <vscale x 1 x i8> %vb, <vscale x 1 x i1> %m, i32 %evl)
   ret <vscale x 1 x i8> %v
 }
 
 define <vscale x 1 x i8> @vadd_vx_nxv1i8_commute(<vscale x 1 x i8> %va, i8 %b, <vscale x 1 x i1> %m, i32 zeroext %evl) {
 ; CHECK-LABEL: vadd_vx_nxv1i8_commute:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a2, zero, e8, mf8, ta, mu
-; CHECK-NEXT:    vmv.v.x v9, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e8, mf8, ta, mu
-; CHECK-NEXT:    vadd.vv v8, v9, v8, v0.t
+; CHECK-NEXT:    vadd.vx v8, v8, a0, v0.t
 ; CHECK-NEXT:    ret
   %elt.head = insertelement <vscale x 1 x i8> poison, i8 %b, i32 0
   %vb = shufflevector <vscale x 1 x i8> %elt.head, <vscale x 1 x i8> poison, <vscale x 1 x i32> zeroinitializer
   %v = call <vscale x 1 x i8> @llvm.vp.add.nxv1i8(<vscale x 1 x i8> %vb, <vscale x 1 x i8> %va, <vscale x 1 x i1> %m, i32 %evl)
   ret <vscale x 1 x i8> %v
 }
 
 define <vscale x 1 x i8> @vadd_vx_nxv1i8_unmasked(<vscale x 1 x i8> %va, i8 %b, i32 zeroext %evl) {

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

; CHECK-NEXT:    ret
   %vb = shufflevector <vscale x 32 x i16> %elt.head, <vscale x 32 x i16> poison, <vscale x 32 x i32> zeroinitializer
   %v = call <vscale x 32 x i16> @llvm.vp.and.nxv32i16(<vscale x 32 x i16> %va, <vscale x 32 x i16> %vb, <vscale x 32 x i1> %m, i32 %evl)
   ret <vscale x 32 x i16> %v
 }
 
 define <vscale x 32 x i16> @vand_vx_nxv32i16_commute(<vscale x 32 x i16> %va, i16 %b, <vscale x 32 x i1> %m, i32 zeroext %evl) {
 ; CHECK-LABEL: vand_vx_nxv32i16_commute:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a2, zero, e16, m8, ta, mu
-; CHECK-NEXT:    vmv.v.x v16, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e16, m8, ta, mu
-; CHECK-NEXT:    vand.vv v8, v16, v8, v0.t
+; CHECK-NEXT:    vand.vx v8, v8, a0, v0.t
 ; CHECK-NEXT:    ret
   %elt.head = insertelement <vscale x 32 x i16> poison, i16 %b, i32 0
   %vb = shufflevector <vscale x 32 x i16> %elt.head, <vscale x 32 x i16> poison, <vscale x 32 x i32> zeroinitializer
   %v = call <vscale x 32 x i16> @llvm.vp.and.nxv32i16(<vscale x 32 x i16> %vb, <vscale x 32 x i16> %va, <vscale x 32 x i1> %m, i32 %evl)
   ret <vscale x 32 x i16> %v
 }
 
 define <vscale x 32 x i16> @vand_vx_nxv32i16_unmasked(<vscale x 32 x i16> %va, i16 %b, i32 zeroext %evl) {

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

; CHECK-NEXT:    ret
   %vb = shufflevector <vscale x 16 x i32> %elt.head, <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer
   %v = call <vscale x 16 x i32> @llvm.vp.mul.nxv16i32(<vscale x 16 x i32> %va, <vscale x 16 x i32> %vb, <vscale x 16 x i1> %m, i32 %evl)
   ret <vscale x 16 x i32> %v
 }
 
 define <vscale x 16 x i32> @vmul_vx_nxv16i32_commute(<vscale x 16 x i32> %va, i32 %b, <vscale x 16 x i1> %m, i32 zeroext %evl) {
 ; CHECK-LABEL: vmul_vx_nxv16i32_commute:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a2, zero, e32, m8, ta, mu
-; CHECK-NEXT:    vmv.v.x v16, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e32, m8, ta, mu
-; CHECK-NEXT:    vmul.vv v8, v16, v8, v0.t
+; CHECK-NEXT:    vmul.vx v8, v8, a0, v0.t
 ; CHECK-NEXT:    ret
   %elt.head = insertelement <vscale x 16 x i32> poison, i32 %b, i32 0
   %vb = shufflevector <vscale x 16 x i32> %elt.head, <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer
   %v = call <vscale x 16 x i32> @llvm.vp.mul.nxv16i32(<vscale x 16 x i32> %vb, <vscale x 16 x i32> %va, <vscale x 16 x i1> %m, i32 %evl)
   ret <vscale x 16 x i32> %v
 }
 
 define <vscale x 16 x i32> @vmul_vx_nxv16i32_unmasked(<vscale x 16 x i32> %va, i32 %b, i32 zeroext %evl) {

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

; CHECK-NEXT:    ret
   %vb = shufflevector <vscale x 2 x i32> %elt.head, <vscale x 2 x i32> poison, <vscale x 2 x i32> zeroinitializer
   %v = call <vscale x 2 x i32> @llvm.vp.or.nxv2i32(<vscale x 2 x i32> %va, <vscale x 2 x i32> %vb, <vscale x 2 x i1> %m, i32 %evl)
   ret <vscale x 2 x i32> %v
 }
 
 define <vscale x 2 x i32> @vor_vx_nxv2i32_commute(<vscale x 2 x i32> %va, i32 %b, <vscale x 2 x i1> %m, i32 zeroext %evl) {
 ; CHECK-LABEL: vor_vx_nxv2i32_commute:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a2, zero, e32, m1, ta, mu
-; CHECK-NEXT:    vmv.v.x v9, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, mu
-; CHECK-NEXT:    vor.vv v8, v9, v8, v0.t
+; CHECK-NEXT:    vor.vx v8, v8, a0, v0.t
 ; CHECK-NEXT:    ret
   %elt.head = insertelement <vscale x 2 x i32> poison, i32 %b, i32 0
   %vb = shufflevector <vscale x 2 x i32> %elt.head, <vscale x 2 x i32> poison, <vscale x 2 x i32> zeroinitializer
   %v = call <vscale x 2 x i32> @llvm.vp.or.nxv2i32(<vscale x 2 x i32> %vb, <vscale x 2 x i32> %va, <vscale x 2 x i1> %m, i32 %evl)
   ret <vscale x 2 x i32> %v
 }
 
 define <vscale x 2 x i32> @vor_vx_nxv2i32_unmasked(<vscale x 2 x i32> %va, i32 %b, i32 zeroext %evl) {

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

; CHECK-NEXT:    ret
   %vb = shufflevector <vscale x 1 x i16> %elt.head, <vscale x 1 x i16> poison, <vscale x 1 x i32> zeroinitializer
   %v = call <vscale x 1 x i16> @llvm.vp.xor.nxv1i16(<vscale x 1 x i16> %va, <vscale x 1 x i16> %vb, <vscale x 1 x i1> %m, i32 %evl)
   ret <vscale x 1 x i16> %v
 }
 
 define <vscale x 1 x i16> @vxor_vx_nxv1i16_commute(<vscale x 1 x i16> %va, i16 %b, <vscale x 1 x i1> %m, i32 zeroext %evl) {
 ; CHECK-LABEL: vxor_vx_nxv1i16_commute:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a2, zero, e16, mf4, ta, mu
-; CHECK-NEXT:    vmv.v.x v9, a0
 ; CHECK-NEXT:    vsetvli zero, a1, e16, mf4, ta, mu
-; CHECK-NEXT:    vxor.vv v8, v9, v8, v0.t
+; CHECK-NEXT:    vxor.vx v8, v8, a0, v0.t
 ; CHECK-NEXT:    ret
   %elt.head = insertelement <vscale x 1 x i16> poison, i16 %b, i32 0
   %vb = shufflevector <vscale x 1 x i16> %elt.head, <vscale x 1 x i16> poison, <vscale x 1 x i32> zeroinitializer
   %v = call <vscale x 1 x i16> @llvm.vp.xor.nxv1i16(<vscale x 1 x i16> %vb, <vscale x 1 x i16> %va, <vscale x 1 x i1> %m, i32 %evl)
   ret <vscale x 1 x i16> %v
 }
 
 define <vscale x 1 x i16> @vxor_vx_nxv1i16_unmasked(<vscale x 1 x i16> %va, i16 %b, i32 zeroext %evl) {

llvm/utils/TableGen/CodeGenDAGPatterns.cpp

for (unsigned i = 0, e = N->getNumChildren(); i != e; ++i)
     GenerateVariantsOf(N->getChildShared(i), ChildVariants[i], CDP, DepVars);
 
   // Build all permutations based on how the children were formed.
   CombineChildVariants(N, ChildVariants, OutVariants, CDP, DepVars);
 
   // If this node is commutative, consider the commuted order.
   bool isCommIntrinsic = N->isCommutativeIntrinsic(CDP);
   if (NodeInfo.hasProperty(SDNPCommutative) || isCommIntrinsic) {
-    assert((N->getNumChildren()>=2 || isCommIntrinsic) &&
+    unsigned Skip = isCommIntrinsic ? 1 : 0; // First operand is intrinsic id.
+    assert(N->getNumChildren() >= (2 + Skip) &&
            "Commutative but doesn't have 2 children!");
-    // Don't count children which are actually register references.
-    unsigned NC = 0;
-    for (unsigned i = 0, e = N->getNumChildren(); i != e; ++i) {
+    // Don't allow commuting children which are actually register references.
+    bool NoRegisters = true;
+    unsigned i = 0 + Skip;
+    unsigned e = 2 + Skip;
+    for (; i != e; ++i) {
       TreePatternNode *Child = N->getChild(i);
       if (Child->isLeaf())
         if (DefInit *DI = dyn_cast<DefInit>(Child->getLeafValue())) {
           Record *RR = DI->getDef();
           if (RR->isSubClassOf("Register"))
-            continue;
+            NoRegisters = false;
         }
-      NC++;
     }
     // Consider the commuted order.
-    if (isCommIntrinsic) {
-      // Commutative intrinsic. First operand is the intrinsic id, 2nd and 3rd
-      // operands are the commutative operands, and there might be more operands
-      // after those.
-      assert(NC >= 3 &&
-             "Commutative intrinsic should have at least 3 children!");
+    if (NoRegisters) {
       std::vector<std::vector<TreePatternNodePtr>> Variants;
-      Variants.push_back(std::move(ChildVariants[0])); // Intrinsic id.
-      Variants.push_back(std::move(ChildVariants[2]));
-      Variants.push_back(std::move(ChildVariants[1]));
-      for (unsigned i = 3; i != NC; ++i)
-        Variants.push_back(std::move(ChildVariants[i]));
-      CombineChildVariants(N, Variants, OutVariants, CDP, DepVars);
-    } else if (NC == N->getNumChildren()) {
-      std::vector<std::vector<TreePatternNodePtr>> Variants;
-      Variants.push_back(std::move(ChildVariants[1]));
-      Variants.push_back(std::move(ChildVariants[0]));
-      for (unsigned i = 2; i != NC; ++i)
+      unsigned i = 0;
+      if (isCommIntrinsic)
+        Variants.push_back(std::move(ChildVariants[i++])); // Intrinsic id.
+      Variants.push_back(std::move(ChildVariants[i + 1]));
+      i += 2;
+      // Remaining operands are not commuted.
+      for (; i != N->getNumChildren(); ++i)
         Variants.push_back(std::move(ChildVariants[i]));
       CombineChildVariants(N, Variants, OutVariants, CDP, DepVars);
     }
   }
 }
 
 
 // GenerateVariants - Generate variants.  For example, commutative patterns can