The e500 SPE floating-point emulation code is called from
SPEFloatingPointException and SPEFloatingPointRoundException in
arch/powerpc/kernel/traps.c. Those functions have support for
generating SIGFPE, but do_spe_mathemu and speround_handler don't
generate a return value to indicate that this should be done. Such a
return value should depend on whether an exception is raised that has
been set via prctl to generate SIGFPE. This patch adds the relevant
logic in these functions so that SIGFPE is generated as expected by
the glibc testsuite.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Signed-off-by: Scott Wood <scottwood@freescale.com>
The e500 SPE floating-point emulation code has several problems in how
it handles conversions to integer and fixed-point fractional types.
There are the following 20 relevant instructions. These can convert
to signed or unsigned 32-bit integers, either rounding towards zero
(as correct for C casts from floating-point to integer) or according
to the current rounding mode, or to signed or unsigned 32-bit
fixed-point values (values in the range [-1, 1) or [0, 1)). For
conversion from double precision there are also instructions to
convert to 64-bit integers, rounding towards zero, although as far as
I know those instructions are completely theoretical (they are only
defined for implementations that support both SPE and classic 64-bit,
and I'm not aware of any such hardware even though the architecture
definition permits that combination).
#define EFSCTUI 0x2d4
#define EFSCTSI 0x2d5
#define EFSCTUF 0x2d6
#define EFSCTSF 0x2d7
#define EFSCTUIZ 0x2d8
#define EFSCTSIZ 0x2da
#define EVFSCTUI 0x294
#define EVFSCTSI 0x295
#define EVFSCTUF 0x296
#define EVFSCTSF 0x297
#define EVFSCTUIZ 0x298
#define EVFSCTSIZ 0x29a
#define EFDCTUIDZ 0x2ea
#define EFDCTSIDZ 0x2eb
#define EFDCTUI 0x2f4
#define EFDCTSI 0x2f5
#define EFDCTUF 0x2f6
#define EFDCTSF 0x2f7
#define EFDCTUIZ 0x2f8
#define EFDCTSIZ 0x2fa
The emulation code, for the instructions that come in variants
rounding either towards zero or according to the current rounding
direction, uses "if (func & 0x4)" as a condition for using _FP_ROUND
(otherwise _FP_ROUND_ZERO is used). The condition is correct, but the
code it controls isn't. Whether _FP_ROUND or _FP_ROUND_ZERO is used
makes no difference, as the effect of those soft-fp macros is to round
an intermediate floating-point result using the low three bits (the
last one sticky) of the working format. As these operations are
dealing with a freshly unpacked floating-point input, those low bits
are zero and no rounding occurs. The emulation code then uses the
FP_TO_INT_* macros for the actual integer conversion, with the effect
of always rounding towards zero; for rounding according to the current
rounding direction, it should be using FP_TO_INT_ROUND_*.
The instructions in question have semantics defined (in the Power ISA
documents) for out-of-range values and NaNs: out-of-range values
saturate and NaNs are converted to zero. The emulation does nothing
to follow those semantics for NaNs (the soft-fp handling is to treat
them as infinities), and messes up the saturation semantics. For
single-precision conversion to integers, (((func & 0x3) != 0) || SB_s)
is the condition used for doing a signed conversion. The first part
is correct, but the second isn't: negative numbers should result in
saturation to 0 when converted to unsigned. Double-precision
conversion to 64-bit integers correctly uses ((func & 0x1) == 0).
Double-precision conversion to 32-bit integers uses (((func & 0x3) !=
0) || DB_s), with correct first part and incorrect second part. And
vector float conversion to integers uses (((func & 0x3) != 0) ||
SB0_s) (and similar for the other vector element), where the sign bit
check is again wrong.
The incorrect handling of negative numbers converted to unsigned was
introduced in commit afc0a07d4a. The
rationale given there was a C testcase with cast from float to
unsigned int. Conversion of out-of-range floating-point numbers to
integer types in C is undefined behavior in the base standard, defined
in Annex F to produce an unspecified value. That is, the C testcase
used to justify that patch is incorrect - there is no ISO C
requirement for a particular value resulting from this conversion -
and in any case, the correct semantics for such emulation are the
semantics for the instruction (unsigned saturation, which is what it
does in hardware when the emulation is disabled).
The conversion to fixed-point values has its own problems. That code
doesn't try to do a full emulation; it relies on the trap handler only
being called for arguments that are infinities, NaNs, subnormal or out
of range. That's fine, but the logic ((vb.wp[1] >> 23) == 0xff &&
((vb.wp[1] & 0x7fffff) > 0)) for NaN detection won't detect negative
NaNs as being NaNs (the same applies for the double-precision case),
and subnormals are mapped to 0 rather than respecting the rounding
mode; the code should also explicitly raise the "invalid" exception.
The code for vectors works by executing the scalar float instruction
with the trapping disabled, meaning at least subnormals won't be
handled correctly.
As well as all those problems in the main emulation code, the rounding
handler - used to emulate rounding upward and downward when not
supported in hardware and when no higher priority exception occurred -
has its own problems.
* It gets called in some cases even for the instructions rounding to
zero, and then acts according to the current rounding mode when it
should just leave alone the truncated result provided by hardware.
* It presumes that the result is a single-precision, double-precision
or single-precision vector as appropriate for the instruction type,
determines the sign of the result accordingly, and then adjusts the
result based on that sign and the rounding mode.
- In the single-precision cases at least the sign determination for
an integer result is the same as for a floating-point result; in
the double-precision case, converted to 32-bit integer or fixed
point, the sign of a double-precision value is in the high part of
the register but it's the low part of the register that has the
result of the conversion.
- If the result is unsigned fixed-point, its sign may be wrongly
determined as negative (does not actually cause problems, because
inexact unsigned fixed-point results with the high bit set can
only appear when converting from double, in which case the sign
determination is instead wrongly using the high part of the
register).
- If the sign of the result is correctly determined as negative, any
adjustment required to change the truncated result to one correct
for the rounding mode should be in the opposite direction for
two's-complement integers as for sign-magnitude floating-point
values.
- And if the integer result is zero, the correct sign can only be
determined by examining the original operand, and not at all (as
far as I can tell) if the operand and result are the same
register.
This patch fixes all these problems (as far as possible, given the
inability to determine the correct sign in the rounding handler when
the truncated result is 0, the conversion is to a signed type and the
truncated result has overwritten the original operand). Conversion to
fixed-point now uses full emulation, and does not use "asm" in the
vector case; the semantics are exactly those of converting to integer
according to the current rounding direction, once the exponent has
been adjusted, so the code makes such an adjustment then uses the
FP_TO_INT_ROUND macros.
The testcase I used for verifying that the instructions (other than
the theoretical conversions to 64-bit integers) produce the correct
results is at <http://lkml.org/lkml/2013/10/8/708>.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Signed-off-by: Scott Wood <scottwood@freescale.com>
The e500 SPE floating-point emulation code for the rounding modes
rounding to positive or negative infinity (which may not be
implemented in hardware) tries to avoid emulating rounding if the
result was inexact. However, it tests inexactness using the sticky
bit with the cumulative result of previous operations, rather than
with the non-sticky bits relating to the operation that generated the
interrupt. Furthermore, when a vector operation generates the
interrupt, it's possible that only one of the low and high parts is
inexact, and so only that part should have rounding emulated. This
results in incorrect rounding of exact results in these modes when the
sticky bit is set from a previous operation.
(I'm not sure why the rounding interrupts are generated at all when
the result is exact, but empirically the hardware does generate them.)
This patch checks for inexactness using the correct bits of SPEFSCR,
and ensures that rounding only occurs when the relevant part of the
result was actually inexact.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Signed-off-by: Scott Wood <scottwood@freescale.com>
The e500 SPE floating-point emulation code clears existing exceptions
(__FPU_FPSCR &= ~FP_EX_MASK;) before ORing in the exceptions from the
emulated operation. However, these exception bits are the "sticky",
cumulative exception bits, and should only be cleared by the user
program setting SPEFSCR, not implicitly by any floating-point
instruction (whether executed purely by the hardware or emulated).
The spurious clearing of these bits shows up as missing exceptions in
glibc testing.
Fixing this, however, is not as simple as just not clearing the bits,
because while the bits may be from previous floating-point operations
(in which case they should not be cleared), the processor can also set
the sticky bits itself before the interrupt for an exception occurs,
and this can happen in cases when IEEE 754 semantics are that the
sticky bit should not be set. Specifically, the "invalid" sticky bit
is set in various cases with non-finite operands, where IEEE 754
semantics do not involve raising such an exception, and the
"underflow" sticky bit is set in cases of exact underflow, whereas
IEEE 754 semantics are that this flag is set only for inexact
underflow. Thus, for correct emulation the kernel needs to know the
setting of these two sticky bits before the instruction being
emulated.
When a floating-point operation raises an exception, the kernel can
note the state of the sticky bits immediately afterwards. Some
<fenv.h> functions that affect the state of these bits, such as
fesetenv and feholdexcept, need to use prctl with PR_GET_FPEXC and
PR_SET_FPEXC anyway, and so it is natural to record the state of those
bits during that call into the kernel and so avoid any need for a
separate call into the kernel to inform it of a change to those bits.
Thus, the interface I chose to use (in this patch and the glibc port)
is that one of those prctl calls must be made after any userspace
change to those sticky bits, other than through a floating-point
operation that traps into the kernel anyway. feclearexcept and
fesetexceptflag duly make those calls, which would not be required
were it not for this issue.
The previous EGLIBC port, and the uClibc code copied from it, is
fundamentally broken as regards any use of prctl for floating-point
exceptions because it didn't use the PR_FP_EXC_SW_ENABLE bit in its
prctl calls (and did various worse things, such as passing a pointer
when prctl expected an integer). If you avoid anything where prctl is
used, the clearing of sticky bits still means it will never give
anything approximating correct exception semantics with existing
kernels. I don't believe the patch makes things any worse for
existing code that doesn't try to inform the kernel of changes to
sticky bits - such code may get incorrect exceptions in some cases,
but it would have done so anyway in other cases.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Signed-off-by: Scott Wood <scottwood@freescale.com>
We already have cpu a005 errata handler when instruction cannot be
recognized. Before we lookup the inst, there's type checking, and we also
need to handle it in errata handler when the type checking failed.
Signed-off-by: Liu Yu <yu.liu@freescale.com>
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
Conversion from float to integer should based on both the instruction
encoding and the sign of the operand.
A simple testcase to show the issue:
static float fm;
static signed int si_min = (-2147483647 - 1);
static unsigned int ui;
int main()
{
fm = (float) si_min; ;
ui = (unsigned int)fm;
printf("ui=%d, should be %d\n", ui, si_min);
return 0;
}
Result: ui=-1, should be -2147483648
Signed-off-by: Shan Hai <shan.hai@windriver.com>
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
This erratum can occur if a single-precision floating-point,
double-precision floating-point or vector floating-point instruction on a
mispredicted branch path signals one of the floating-point data interrupts
which are enabled by the SPEFSCR (FINVE, FDBZE, FUNFE or FOVFE bits). This
interrupt must be recorded in a one-cycle window when the misprediction is
resolved. If this extremely rare event should occur, the result could be:
The SPE Data Exception from the mispredicted path may be reported
erroneously if a single-precision floating-point, double-precision
floating-point or vector floating-point instruction is the second
instruction on the correct branch path.
According to errata description, some efp instructions which are not
supposed to trigger SPE exceptions can trigger the exceptions in this case.
However, as we haven't emulated these instructions here, a signal will
send to userspace, and userspace application would exit.
This patch re-issue the efp instruction that we haven't emulated,
so that hardware can properly execute it again if this case happen.
Signed-off-by: Liu Yu <yu.liu@freescale.com>
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
This patch add the handlers of SPE/EFP exceptions.
The code is used to emulate float point arithmetic,
when MSR(SPE) is enabled and receive EFP data interrupt or EFP round interrupt.
This patch has no conflict with or dependence on FP math-emu.
The code has been tested by TestFloat.
Now the code doesn't support SPE/EFP instructions emulation
(it won't be called when receive program interrupt),
but it could be easily added.
Signed-off-by: Liu Yu <yu.liu@freescale.com>
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
Joseph Myers