Two fixes for the exception status that qemu-alpha reports to code
built with gcc -mieee. I found both running glibc's math tests under
linux-user.
Patch 1 fixes 471d493047, which was meant to record masked exceptions
in the FPCR but had no effect for insns with software completion.
Without it FE_INEXACT is never set for -mieee code.
Patch 2 makes underflow match the hardware for a result that rounds up
to the smallest normal number. Alpha raises no underflow there and
softfloat does.
Patch 1 does not depend on patch 2.
Tested with linux-user only; system mode is untested.
glibc's math tests (make check subdirs=math), run under linux-user:
PASS FAIL XFAIL
master 1400 110 1
series 1510 0 1
Nearly all of the failures on master are 'Exception "Inexact" not
set'. The XFAIL is test-float32x-float64-div. On master it fails for
missing inexact; with the series it fails only the four
div_{upward,downward}_double cases that glibc expects to fail on
Alpha.
The tcg-alpha-linux-user suite has six failures (float_convs,
float_convd, float_madds and a plugin variant of each). They fail the
same way on master.
Signed-off-by: Matt Turner <[email protected]>
---
Matt Turner (2):
target/alpha: Record masked exceptions in FPCR for /S insns
target/alpha: Determine tininess from the delivered result
target/alpha/fpu_helper.c | 46 ++++++++++++++++++++++++++++++++++++----------
1 file changed, 36 insertions(+), 10 deletions(-)
---
base-commit: 615ece3c406b262996e569f29ea6eed0d81c4d8d
change-id: 20261012-alpha-fpcr-status-s-5b5c42702cfc
Best regards,
--
Matt Turner <[email protected]>
