[libc-commits] [PATCH] D158118: [libc] Fix HPD on extremely long numbers

[Michael Jones via Phabricator via libc-commits]

michaelrj updated this revision to Diff 551285.
michaelrj added a comment.

I was wrong. I've added tests. They don't test the behavior directly, but they're the best I can do without 100k of zeroes. Also I increased the maximum exponent HPD will store to 2^30, which is ~1 billion. If we discover this limit is too low I will be shocked.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D158118/new/

https://reviews.llvm.org/D158118

Files:
  libc/src/__support/high_precision_decimal.h
  libc/test/src/__support/high_precision_decimal_test.cpp


Index: libc/test/src/__support/high_precision_decimal_test.cpp
===================================================================
--- libc/test/src/__support/high_precision_decimal_test.cpp
+++ libc/test/src/__support/high_precision_decimal_test.cpp
@@ -391,3 +391,18 @@
 
   EXPECT_EQ(hpd.round_to_integer_type<UInt128>(), result);
 }
+
+TEST(LlvmLibcHighPrecisionDecimalTest, BigExpTest) {
+  __llvm_libc::internal::HighPrecisionDecimal big_hpd =
+      __llvm_libc::internal::HighPrecisionDecimal("1e123456789");
+
+  // We need to add one to handle the digit before the decimal point in our
+  // number.
+  EXPECT_EQ(big_hpd.get_decimal_point(), 123456789 + 1);
+
+  __llvm_libc::internal::HighPrecisionDecimal big_negative_hpd =
+      __llvm_libc::internal::HighPrecisionDecimal("1e-123456789");
+
+  // Same, but since the number is negative the net result is -123456788
+  EXPECT_EQ(big_negative_hpd.get_decimal_point(), -123456789 + 1);
+}
Index: libc/src/__support/high_precision_decimal.h
===================================================================
--- libc/src/__support/high_precision_decimal.h
+++ libc/src/__support/high_precision_decimal.h
@@ -334,13 +334,24 @@
     if ((*numString | 32) == 'e') {
       ++numString;
       if (isdigit(*numString) || *numString == '+' || *numString == '-') {
-        int32_t add_to_exp = strtointeger<int32_t>(numString, 10);
-        if (add_to_exp > 100000) {
-          add_to_exp = 100000;
-        } else if (add_to_exp < -100000) {
-          add_to_exp = -100000;
+        auto result = strtointeger<int32_t>(numString, 10);
+        if (result.has_error()) {
+          // TODO: handle error
         }
-        this->decimal_point += add_to_exp;
+        int32_t add_to_exponent = result.value;
+
+        // Here we do this operation as int64 to avoid overflow.
+        int64_t temp_exponent = static_cast<int64_t>(this->decimal_point) +
+                                static_cast<int64_t>(add_to_exponent);
+
+        // Theoretically these numbers should be MAX_EXPONENT for long double,
+        // but that should be ~16,000 which is much less than 1 << 30.
+        if (temp_exponent > (1 << 30)) {
+          temp_exponent = (1 << 30);
+        } else if (temp_exponent < -(1 << 30)) {
+          temp_exponent = -(1 << 30);
+        }
+        this->decimal_point = static_cast<int32_t>(temp_exponent);
       }
     }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D158118.551285.patch
Type: text/x-patch
Size: 2404 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/libc-commits/attachments/20230817/56b2d304/attachment.bin>