[RISCV] Support vscale intrinsic and ISDNode.

[HsiangKai]

Summary

In RISC-V, we have VLENB register to know the real vector size at run time. vscale is the run time constant in LLVM to know the factor in the scalable vector types. We could use VLENB to get vscale value in LLVM.

[craig.topper]

llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td

2330

I think int_vscale becomes (vscale 1) in SelectionDAGBuilder so we can probably drop this pattern?

llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td

 //===----------------------------------------------------------------------===//
 // 17.4. Vector Register Gather Instructions
 //===----------------------------------------------------------------------===//
 defm PseudoVRGATHER    : VPseudoBinaryV_VV_VX_VI<uimm5, "@earlyclobber $rd">;
 
 //===----------------------------------------------------------------------===//
 // Patterns.
 //===----------------------------------------------------------------------===//
+def uimmpow2xlen : ImmLeaf<XLenVT, [{
+  if (Subtarget->is64Bit())
+    return isPowerOf2_64(Imm) && isUInt<6>(Imm - 1);
+  else
+    return isPowerOf2_32(Imm) && isUInt<6>(Imm - 1);
+}]>;
+
+def uimm6 : ImmLeaf<XLenVT, [{
+  return isUInt<6>(Imm - 1);
+}]>;
+
+def Log2 : SDNodeXForm<imm,
+[{
+  if (Subtarget->is64Bit())
+    return CurDAG->getTargetConstant(Log2_64(N->getZExtValue()), SDLoc(N),
+        MVT::i64);
+  else
+    return CurDAG->getTargetConstant(Log2_32(N->getZExtValue()), SDLoc(N),
+        MVT::i32);
+}]>;
+
+// Constant 3 for 8 bytes in <vscale x 8-bytes>.
+def read_vscale : OutPatFrag<(ops), (SRLI (PseudoReadVLENB), 3)>;
+
 let Predicates = [HasStdExtV] in {
 
+def : Pat<(vscale 0), (i64 X0)>;
+def : Pat<(vscale 1), (read_vscale)>;
+def : Pat<(vscale uimmpow2xlen:$mul), (SLLI (read_vscale), (Log2 imm:$mul))>;
+// Only support up to (vscale x 64) bytes due to the maximum LMUL is 8.
+def : Pat<(vscale uimm6:$mul), (MUL (read_vscale), (ADDI X0, uimm6:$mul))>;
+
 //===----------------------------------------------------------------------===//
 // 7. Vector Loads and Stores
 //===----------------------------------------------------------------------===//
 
 //===----------------------------------------------------------------------===//
 // 7.4 Vector Unit-Stride Instructions
 //===----------------------------------------------------------------------===//
 

llvm/test/CodeGen/RISCV/rvv/rvv-vscale.i32.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple riscv32 -mattr=+m,+experimental-v < %s \
+; RUN:    | FileCheck %s
+
+define i32 @vscale_zero() nounwind {
+; CHECK-LABEL: vscale_zero:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    mv a0, zero
+; CHECK-NEXT:    ret
+entry:
+  %0 = call i32 @llvm.vscale.i32()
+  %1 = mul i32 %0, 0
+  ret i32 %1
+}
+
+define i32 @vscale_one() nounwind {
+; CHECK-LABEL: vscale_one:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    ret
+entry:
+  %0 = call i32 @llvm.vscale.i32()
+  %1 = mul i32 %0, 1
+  ret i32 %1
+}
+
+define i32 @vscale_uimmpow2xlen() nounwind {
+; CHECK-LABEL: vscale_uimmpow2xlen:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    slli a0, a0, 6
+; CHECK-NEXT:    ret
+entry:
+  %0 = call i32 @llvm.vscale.i32()
+  %1 = mul i32 %0, 64
+  ret i32 %1
+}
+
+define i32 @vscale_non_pow2() nounwind {
+; CHECK-LABEL: vscale_non_pow2:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    addi a1, zero, 24
+; CHECK-NEXT:    mul a0, a0, a1
+; CHECK-NEXT:    ret
+entry:
+  %0 = call i32 @llvm.vscale.i32()
+  %1 = mul i32 %0, 24
+  ret i32 %1
+}
+
+declare i32 @llvm.vscale.i32()

llvm/test/CodeGen/RISCV/rvv/rvv-vscale.i64.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple riscv64 -mattr=+m,+experimental-v < %s \
+; RUN:    | FileCheck %s
+
+define i64 @vscale_zero() nounwind {
+; CHECK-LABEL: vscale_zero:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    mv a0, zero
+; CHECK-NEXT:    ret
+entry:
+  %0 = call i64 @llvm.vscale.i64()
+  %1 = mul i64 %0, 0
+  ret i64 %1
+}
+
+define i64 @vscale_one() nounwind {
+; CHECK-LABEL: vscale_one:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    ret
+entry:
+  %0 = call i64 @llvm.vscale.i64()
+  %1 = mul i64 %0, 1
+  ret i64 %1
+}
+
+define i64 @vscale_uimmpow2xlen() nounwind {
+; CHECK-LABEL: vscale_uimmpow2xlen:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    slli a0, a0, 6
+; CHECK-NEXT:    ret
+entry:
+  %0 = call i64 @llvm.vscale.i64()
+  %1 = mul i64 %0, 64
+  ret i64 %1
+}
+
+define i64 @vscale_non_pow2() nounwind {
+; CHECK-LABEL: vscale_non_pow2:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    addi a1, zero, 24
+; CHECK-NEXT:    mul a0, a0, a1
+; CHECK-NEXT:    ret
+entry:
+  %0 = call i64 @llvm.vscale.i64()
+  %1 = mul i64 %0, 24
+  ret i64 %1
+}
+
+declare i64 @llvm.vscale.i64()