Reland [ValueTracking] Improve the coverage of isKnownToBeAPowerOfTwo for vscale

[Allen]

Summary

The previous commit 9c837b7d0e2e don't check the type width,
so the logic is incorrect because a shift left of vscale can be zero
if all bits get shifted out.

This PR add condition to check above issue.

[nikic]

Given the lack of test changes, do we need this? Presumably https://github.com/llvm/llvm-project/commit/5de89b4d99c913e3c7c2735886f2519a2ed39d8a already covered the motivating cases.

Edit: To clarify, we do need to remove the incorrect code, the question is more about whether we need the new code this introduces.

[Allen]

rebase after revert of f6c726472df1

[Allen]

In my local test, there is some failure tests after the revert of f6c726472df1, and this PR can fix them

LLVM :: Transforms/InstCombine/rem-mul-shl.ll
LLVM :: Transforms/LoopVectorize/AArch64/sve-interleaved-accesses.ll
LLVM :: Transforms/LoopVectorize/AArch64/sve-widen-phi.ll

[nikic]

I don't get any failures for check-llvm on current main, which includes the revert.

[goldstein.w.n]

In D157292#4566153, @nikic wrote:

I don't get any failures for check-llvm on current main, which includes the revert.

Likewise.

[goldstein.w.n]

llvm/lib/Analysis/ValueTracking.cpp

2028

More robust would be to match shift as a value and use ComputeKnownBits to get max shift cnt.

2031

If we have OrZero don't need the extra logic.

2035

Keep these as APInt to avoid overflow / handle abnormal types like i128

[Allen]

oh, I find the commit 5de89b4d99c9 already fixed, thanks @goldstein.w.n and @nikic, so this is not necessary

llvm/lib/Analysis/ValueTracking.cpp

   if (match(V, m_Power2()))
       return true;
   if (Q.CxtI && match(V, m_VScale())) {
     const Function *F = Q.CxtI->getFunction();
     // The vscale_range indicates vscale is a power-of-two.
     return F->hasFnAttribute(Attribute::VScaleRange);
   }
 
+  const APInt *Shift;
+  if (Q.CxtI && match(V, m_Shl(m_VScale(), m_APInt(Shift)))) {
+    const Function *F = Q.CxtI->getFunction();
+    // The vscale_range indicates vscale is a power-of-two.
+    if (F->hasFnAttribute(Attribute::VScaleRange)) {
+      Attribute Attr = F->getFnAttribute(Attribute::VScaleRange);
+      if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax()) {
+        unsigned Bitwidth = V->getType()->getScalarSizeInBits();
+        if (Log2_32(*MaxVScale) + Shift->getZExtValue() < Bitwidth)
+          return true;
+      }
+    }
+  }
+
   // 1 << X is clearly a power of two if the one is not shifted off the end.  If
   // it is shifted off the end then the result is undefined.
   if (match(V, m_Shl(m_One(), m_Value())))
     return true;
 
   // (signmask) >>l X is clearly a power of two if the one is not shifted off
   // the bottom.  If it is shifted off the bottom then the result is undefined.
   if (match(V, m_LShr(m_SignMask(), m_Value())))