[libc] Fix printf long double bugs

[michaelrj-google]

Found in testing against abseil. Two bugs were found:

1. SHIFT_AMOUNT in float_converter would sometimes be
   negative causing an underflow when passed as the amount to left shift
   by for BigInt.
2. is_lowest_block had an off-by-one because it was adding 1 to the
   block index.

Both are fixed and there are new tests to catch regressions.

[llvmbot]

@llvm/pr-subscribers-libc

Author: Michael Jones (michaelrj-google)

Changes

Found in testing against abseil. Two bugs were found:

1. SHIFT_AMOUNT in float_converter<long double> would sometimes be
   negative causing an underflow when passed as the amount to left shift
   by for BigInt.
2. is_lowest_block had an off-by-one because it was adding 1 to the
   block index.

Both are fixed and there are new tests to catch regressions.

---

Full diff: https://github.com/llvm/llvm-project/pull/166474.diff

2 Files Affected:

• (modified) libc/src/__support/float_to_string.h (+6-2)
• (modified) libc/test/src/stdio/sprintf_test.cpp (+12)

diff --git a/libc/src/__support/float_to_string.h b/libc/src/__support/float_to_string.h
index cab146a5b8698..9115ed2856881 100644
--- a/libc/src/__support/float_to_string.h
+++ b/libc/src/__support/float_to_string.h
@@ -700,7 +700,11 @@ template <> class FloatToString<long double> {
 
       const int SHIFT_AMOUNT = FLOAT_AS_INT_WIDTH + exponent;
       static_assert(EXTRA_INT_WIDTH >= sizeof(long double) * 8);
-      float_as_fixed <<= SHIFT_AMOUNT;
+      if (SHIFT_AMOUNT > 0) {
+        float_as_fixed <<= SHIFT_AMOUNT;
+      } else {
+        float_as_fixed >>= -SHIFT_AMOUNT;
+      }
 
       // If there are still digits above the decimal point, handle those.
       if (cpp::countl_zero(float_as_fixed) <
@@ -769,7 +773,7 @@ template <> class FloatToString<long double> {
     // The decimal representation of 2**(-i) will have exactly i digits after
     // the decimal point.
     const int num_requested_digits =
-        static_cast<int>((negative_block_index + 1) * BLOCK_SIZE);
+        static_cast<int>(negative_block_index * BLOCK_SIZE);
 
     return num_requested_digits > -exponent;
   }
diff --git a/libc/test/src/stdio/sprintf_test.cpp b/libc/test/src/stdio/sprintf_test.cpp
index f1b545ba546f9..42fdd59cf4d9c 100644
--- a/libc/test/src/stdio/sprintf_test.cpp
+++ b/libc/test/src/stdio/sprintf_test.cpp
@@ -1537,6 +1537,14 @@ TEST(LlvmLibcSPrintfTest, FloatDecimalLongDoubleConv) {
 #if defined(LIBC_TYPES_LONG_DOUBLE_IS_X86_FLOAT80)
 
 #ifndef LIBC_COPT_FLOAT_TO_STR_REDUCED_PRECISION
+  written = LIBC_NAMESPACE::sprintf(
+      buff, "%.75Lf",
+      0.0833333333333333333355920878593448009041821933351457118988037109375L);
+  ASSERT_STREQ_LEN(written, buff,
+                   "0."
+                   "08333333333333333333559208785934480090418219333514571189880"
+                   "3710937500000000");
+
   written = LIBC_NAMESPACE::sprintf(buff, "%Lf", 1e100L);
   ASSERT_STREQ_LEN(written, buff,
                    "99999999999999999996693535322073426194986990198284960792713"
@@ -2976,6 +2984,10 @@ TEST(LlvmLibcSPrintfTest, FloatAutoLongDoubleConv) {
   written = LIBC_NAMESPACE::sprintf(buff, "%Lg", 0xf.fffffffffffffffp+16380L);
   ASSERT_STREQ_LEN(written, buff, "1.18973e+4932");
 
+  // Minimum normal
+  written = LIBC_NAMESPACE::sprintf(buff, "%Lg", 3.36210314311209350626E-4932L);
+  ASSERT_STREQ_LEN(written, buff, "3.3621e-4932");
+
   written = LIBC_NAMESPACE::sprintf(buff, "%Lg", 0xa.aaaaaaaaaaaaaabp-7L);
   ASSERT_STREQ_LEN(written, buff, "0.0833333");