[RISCV][NFC] Use NFList in NFSet

[wangpc-pp]

These two subroutines are almost the same except the LMUL encoding.

We can use NFList in NFSet to remove duplicated code.

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Changes

These two subroutines are almost the same except the LMUL encoding.

We can use NFList in NFSet to remove duplicated code.

---

Full diff: https://github.com/llvm/llvm-project/pull/67517.diff

2 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td (+2-4)
• (modified) llvm/lib/Target/RISCV/RISCVRegisterInfo.td (+5-4)

diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
index 0c76e36daf72033..e0d75ef77b76a53 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
@@ -219,10 +219,8 @@ defvar FPListW = [SCALAR_F16, SCALAR_F32];
 defvar BFPListW = [SCALAR_BF16];
 
 class NFSet<LMULInfo m> {
-  list<int> L = !cond(!eq(m.value, V_M8.value): [],
-                      !eq(m.value, V_M4.value): [2],
-                      !eq(m.value, V_M2.value): [2, 3, 4],
-                      true: [2, 3, 4, 5, 6, 7, 8]);
+  defvar lmul = !shl(1, m.value);
+  list<int> L = NFList<lmul>.L;
 }
 
 class octuple_to_str<int octuple> {
diff --git a/llvm/lib/Target/RISCV/RISCVRegisterInfo.td b/llvm/lib/Target/RISCV/RISCVRegisterInfo.td
index 1a6145f92908134..ab0d354967b34c7 100644
--- a/llvm/lib/Target/RISCV/RISCVRegisterInfo.td
+++ b/llvm/lib/Target/RISCV/RISCVRegisterInfo.td
@@ -337,14 +337,15 @@ defvar LMULList = [1, 2, 4, 8];
 // Utility classes for segment load/store.
 //===----------------------------------------------------------------------===//
 // The set of legal NF for LMUL = lmul.
-// LMUL == 1, NF = 2, 3, 4, 5, 6, 7, 8
+// LMUL <= 1, NF = 2, 3, 4, 5, 6, 7, 8
 // LMUL == 2, NF = 2, 3, 4
 // LMUL == 4, NF = 2
+// LMUL == 8, no legal NF
 class NFList<int lmul> {
-  list<int> L = !cond(!eq(lmul, 1): [2, 3, 4, 5, 6, 7, 8],
-                      !eq(lmul, 2): [2, 3, 4],
+  list<int> L = !cond(!eq(lmul, 8): [],
                       !eq(lmul, 4): [2],
-                      !eq(lmul, 8): []);
+                      !eq(lmul, 2): [2, 3, 4],
+                      true: [2, 3, 4, 5, 6, 7, 8]);
 }
 
 // Generate [start, end) SubRegIndex list.

[michaelmaitland]

LGTM.

[michaelmaitland]

For the fractional LMULs, this encoding gives:

>>> 1 << 0b101
32
>>> 1 << 0b110
64
>>> 1 << 0b111
128

NFList was never used with fractional LMUL in the past. This encoding ends up working out since fractional LMUL will hit the true branch of NFList, but it does diverge from RISCVII::VLMUL encoding which caught my eye. I wanted to say something so we're aware of it, but I am not requesting a change since I think it is fine as this is an implementation detail.

[michaelmaitland]

Actually, I think fractional will just take on 0b000 since it’s a 3 bit value. Still fine with me.

Edit: sorry it’s not a 3 bit value I take that back. Please Ignore this comment.