[llvm] [RISCV] Take SEW/LMUL into account for value tracking of vsetvli[max] (PR #82163)

[Luke Lau via llvm-commits]

================
@@ -1576,12 +1577,31 @@ static void computeKnownBitsFromOperator(const Operator *I,
         Known.Zero.setBitsFrom(32);
         break;
       case Intrinsic::riscv_vsetvli:
-      case Intrinsic::riscv_vsetvlimax:
-        // Assume that VL output is <= 65536.
-        // TODO: Take SEW and LMUL into account.
-        if (BitWidth > 17)
-          Known.Zero.setBitsFrom(17);
+      case Intrinsic::riscv_vsetvlimax: {
+        bool HasAVL = II->getIntrinsicID() == Intrinsic::riscv_vsetvli;
+        const ConstantRange &Range =
+            getVScaleRange(II->getFunction(), BitWidth);
+        uint64_t VSEW =
+            cast<ConstantInt>(II->getArgOperand(HasAVL))->getZExtValue();
+        uint64_t SEW = 1 << (VSEW + 3);
+        uint64_t VLMUL =
+            cast<ConstantInt>(II->getArgOperand(1 + HasAVL))->getZExtValue();
+        bool Fractional = VLMUL > 4;
+        uint64_t LMUL = Fractional ? (1 << (8 - VLMUL)) : (1 << VLMUL);
+        uint64_t MaxVL =
+            Range.getUpper().getZExtValue() * RISCV::RVVBitsPerBlock / SEW;
+        MaxVL = Fractional ? MaxVL / LMUL : MaxVL * LMUL;
----------------
lukel97 wrote:

Would using `RISCVVType::decodeVLMUL` help here? Something like
```c++
        auto [Fractional, LMUL] = RISCVVType::decodeVLMUL(VLMUL);
```

Or alternatively you can just compute VLMAX as MaxVLEN / SEWLMULRatio

```suggestion
        uint64_t MaxVLEN = Range.getUpper().getZExtValue() * RISCV::RVVBitsPerBlock;
        uint64_t MaxVL =  MaxVLEN / RISCVII:getSEWLMULRatio(SEW, VLMUL);
```

https://github.com/llvm/llvm-project/pull/82163