On 5/24/26 12:38 AM, H.J. Lu wrote:
Note: The following floating-point exceptions are signalling:
IEEE_INVALID_FLAG IEEE_UNDERFLOW_FLAG IEEE_DENORMAL
STOP 3
FAIL: gfortran.dg/EXformat_1.F90 -O0 execution test
I fed the above and the CI report I received to Claude and it
arrived at the attached patch rather quickly. Can you try this and confirm
whether it fixes it or not?
Regards,
Jerry
---------------
From 837ca2b4c3d1069ddb938a5a7c259d34a85479fc Mon Sep 17 00:00:00 2001
From: Jerry DeLisle <[email protected]>
Date: Sun, 24 May 2026 11:51:49 -0700
Subject: [PATCH] Fortran: [PR93727] Fix EX format kind=8 output on ILP32
targets
On 32-bit targets such as ARM where unsigned long is 32 bits, the
kind=8 case in get_float_hex_string used unsigned long for frac_part.
The kind=8 mantissa requires 52 bits (13 hex digits), so the cast
silently truncated the upper bits, producing wrong hex output. In
addition, converting a ~4.5e15 double value to a 32-bit unsigned long
is out of range, which raised IEEE_INVALID_FLAG on ARM hardware.
The kind=10 case already used unsigned long long for the same reason.
Fix kind=8 to match: change the type, the cast, and the two %lX
format specifiers to %llX.
PR fortran/93727
libgfortran/ChangeLog:
* io/write_float.def (get_float_hex_string): Change frac_part
for kind=8 from unsigned long to unsigned long long and update
format specifiers from %13.13lX to %13.13llX to fix truncated
hex output and IEEE_INVALID_FLAG on ILP32 targets.
Assisted by: Claude Sonnet 4.6
From 837ca2b4c3d1069ddb938a5a7c259d34a85479fc Mon Sep 17 00:00:00 2001
From: Jerry DeLisle <[email protected]>
Date: Sun, 24 May 2026 11:51:49 -0700
Subject: [PATCH] Fortran: [PR93727] Fix EX format kind=8 output on ILP32
targets
On 32-bit targets such as ARM where unsigned long is 32 bits, the
kind=8 case in get_float_hex_string used unsigned long for frac_part.
The kind=8 mantissa requires 52 bits (13 hex digits), so the cast
silently truncated the upper bits, producing wrong hex output. In
addition, converting a ~4.5e15 double value to a 32-bit unsigned long
is out of range, which raised IEEE_INVALID_FLAG on ARM hardware.
The kind=10 case already used unsigned long long for the same reason.
Fix kind=8 to match: change the type, the cast, and the two %lX
format specifiers to %llX.
PR fortran/93727
libgfortran/ChangeLog:
* io/write_float.def (get_float_hex_string): Change frac_part
for kind=8 from unsigned long to unsigned long long and update
format specifiers from %13.13lX to %13.13llX to fix truncated
hex output and IEEE_INVALID_FLAG on ILP32 targets.
Assisted by: Claude Sonnet 4.6
---
libgfortran/io/write_float.def | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/libgfortran/io/write_float.def b/libgfortran/io/write_float.def
index 608106af7c7..66fd7cfe69a 100644
--- a/libgfortran/io/write_float.def
+++ b/libgfortran/io/write_float.def
@@ -128,7 +128,7 @@ get_float_hex_string (const void *source, int kind, char *buffer,
double mant;
int expon;
int int_part;
- unsigned long frac_part;
+ unsigned long long frac_part;
val = *(const GFC_REAL_8 *) source;
is_negative = signbit (val);
@@ -172,11 +172,11 @@ get_float_hex_string (const void *source, int kind, char *buffer,
int_part = (int) mant;
/* 56 is the nearest integer divisible by 4 that is >= 53 (mantissa bits
for kind=8). (56-4)/4 = 13 hex digits for the fractional part. */
- frac_part = (unsigned long) scalbn (mant - (double) int_part, 56 - 4);
+ frac_part = (unsigned long long) scalbn (mant - (double) int_part, 56 - 4);
if (is_negative)
- result = snprintf (buffer, 25, "-0X%X.%13.13lXP%+d", int_part, frac_part, expon);
+ result = snprintf (buffer, 25, "-0X%X.%13.13llXP%+d", int_part, frac_part, expon);
else
- result = snprintf (buffer, 25, "0X%X.%13.13lXP%+d", int_part, frac_part, expon);
+ result = snprintf (buffer, 25, "0X%X.%13.13llXP%+d", int_part, frac_part, expon);
}
break;
#ifdef HAVE_GFC_REAL_10
--
2.54.0
