[llvm] [llvm] produce a more consistent estimates of bit width needed when parsing an integer (PR #205947)

[Eli Friedman via llvm-commits]

================
@@ -522,16 +522,17 @@ bool StringRef::consumeInteger(unsigned Radix, APInt &Result) {
     return false;
   }
 
-  // (Over-)estimate the required number of bits.
   unsigned Log2Radix = 0;
   while ((1U << Log2Radix) < Radix) Log2Radix++;
   bool IsPowerOf2Radix = ((1U << Log2Radix) == Radix);
 
-  unsigned BitWidth = Log2Radix * Str.size();
-  if (BitWidth < Result.getBitWidth())
-    BitWidth = Result.getBitWidth(); // don't shrink the result
-  else if (BitWidth > Result.getBitWidth())
+  // Initialize Result to a reasonable starting width (at least 64 bits),
+  // but do not shrink it if it was already larger.
+  unsigned BitWidth = std::max(64U, Result.getBitWidth());
+
+  if (Result.getBitWidth() < BitWidth) {
     Result = Result.zext(BitWidth);
+  }
 
   APInt RadixAP, CharAP; // unused unless !IsPowerOf2Radix
----------------
efriedma-quic wrote:

There isn't really any reason for RadixAP and CharAP to exist; we have add/mul operators which take a uint64_t as the RHS.  But I guess that can be a followup.

https://github.com/llvm/llvm-project/pull/205947