linux/arch/powerpc/math-emu/math_efp.c

928 lines
20 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-or-later
/*
* arch/powerpc/math-emu/math_efp.c
*
* Copyright (C) 2006-2008, 2010 Freescale Semiconductor, Inc.
*
* Author: Ebony Zhu, <ebony.zhu@freescale.com>
* Yu Liu, <yu.liu@freescale.com>
*
* Derived from arch/alpha/math-emu/math.c
* arch/powerpc/math-emu/math.c
*
* Description:
* This file is the exception handler to make E500 SPE instructions
* fully comply with IEEE-754 floating point standard.
*/
#include <linux/types.h>
#include <linux/prctl.h>
#include <linux/uaccess.h>
#include <asm/reg.h>
#define FP_EX_BOOKE_E500_SPE
#include <asm/sfp-machine.h>
#include <math-emu/soft-fp.h>
#include <math-emu/single.h>
#include <math-emu/double.h>
#define EFAPU 0x4
#define VCT 0x4
#define SPFP 0x6
#define DPFP 0x7
#define EFSADD 0x2c0
#define EFSSUB 0x2c1
#define EFSABS 0x2c4
#define EFSNABS 0x2c5
#define EFSNEG 0x2c6
#define EFSMUL 0x2c8
#define EFSDIV 0x2c9
#define EFSCMPGT 0x2cc
#define EFSCMPLT 0x2cd
#define EFSCMPEQ 0x2ce
#define EFSCFD 0x2cf
#define EFSCFSI 0x2d1
#define EFSCTUI 0x2d4
#define EFSCTSI 0x2d5
#define EFSCTUF 0x2d6
#define EFSCTSF 0x2d7
#define EFSCTUIZ 0x2d8
#define EFSCTSIZ 0x2da
#define EVFSADD 0x280
#define EVFSSUB 0x281
#define EVFSABS 0x284
#define EVFSNABS 0x285
#define EVFSNEG 0x286
#define EVFSMUL 0x288
#define EVFSDIV 0x289
#define EVFSCMPGT 0x28c
#define EVFSCMPLT 0x28d
#define EVFSCMPEQ 0x28e
#define EVFSCTUI 0x294
#define EVFSCTSI 0x295
#define EVFSCTUF 0x296
#define EVFSCTSF 0x297
#define EVFSCTUIZ 0x298
#define EVFSCTSIZ 0x29a
#define EFDADD 0x2e0
#define EFDSUB 0x2e1
#define EFDABS 0x2e4
#define EFDNABS 0x2e5
#define EFDNEG 0x2e6
#define EFDMUL 0x2e8
#define EFDDIV 0x2e9
#define EFDCTUIDZ 0x2ea
#define EFDCTSIDZ 0x2eb
#define EFDCMPGT 0x2ec
#define EFDCMPLT 0x2ed
#define EFDCMPEQ 0x2ee
#define EFDCFS 0x2ef
#define EFDCTUI 0x2f4
#define EFDCTSI 0x2f5
#define EFDCTUF 0x2f6
#define EFDCTSF 0x2f7
#define EFDCTUIZ 0x2f8
#define EFDCTSIZ 0x2fa
#define AB 2
#define XA 3
#define XB 4
#define XCR 5
#define NOTYPE 0
#define SIGN_BIT_S (1UL << 31)
#define SIGN_BIT_D (1ULL << 63)
#define FP_EX_MASK (FP_EX_INEXACT | FP_EX_INVALID | FP_EX_DIVZERO | \
FP_EX_UNDERFLOW | FP_EX_OVERFLOW)
static int have_e500_cpu_a005_erratum;
union dw_union {
u64 dp[1];
u32 wp[2];
};
static unsigned long insn_type(unsigned long speinsn)
{
unsigned long ret = NOTYPE;
switch (speinsn & 0x7ff) {
case EFSABS: ret = XA; break;
case EFSADD: ret = AB; break;
case EFSCFD: ret = XB; break;
case EFSCMPEQ: ret = XCR; break;
case EFSCMPGT: ret = XCR; break;
case EFSCMPLT: ret = XCR; break;
case EFSCTSF: ret = XB; break;
case EFSCTSI: ret = XB; break;
case EFSCTSIZ: ret = XB; break;
case EFSCTUF: ret = XB; break;
case EFSCTUI: ret = XB; break;
case EFSCTUIZ: ret = XB; break;
case EFSDIV: ret = AB; break;
case EFSMUL: ret = AB; break;
case EFSNABS: ret = XA; break;
case EFSNEG: ret = XA; break;
case EFSSUB: ret = AB; break;
case EFSCFSI: ret = XB; break;
case EVFSABS: ret = XA; break;
case EVFSADD: ret = AB; break;
case EVFSCMPEQ: ret = XCR; break;
case EVFSCMPGT: ret = XCR; break;
case EVFSCMPLT: ret = XCR; break;
case EVFSCTSF: ret = XB; break;
case EVFSCTSI: ret = XB; break;
case EVFSCTSIZ: ret = XB; break;
case EVFSCTUF: ret = XB; break;
case EVFSCTUI: ret = XB; break;
case EVFSCTUIZ: ret = XB; break;
case EVFSDIV: ret = AB; break;
case EVFSMUL: ret = AB; break;
case EVFSNABS: ret = XA; break;
case EVFSNEG: ret = XA; break;
case EVFSSUB: ret = AB; break;
case EFDABS: ret = XA; break;
case EFDADD: ret = AB; break;
case EFDCFS: ret = XB; break;
case EFDCMPEQ: ret = XCR; break;
case EFDCMPGT: ret = XCR; break;
case EFDCMPLT: ret = XCR; break;
case EFDCTSF: ret = XB; break;
case EFDCTSI: ret = XB; break;
case EFDCTSIDZ: ret = XB; break;
case EFDCTSIZ: ret = XB; break;
case EFDCTUF: ret = XB; break;
case EFDCTUI: ret = XB; break;
case EFDCTUIDZ: ret = XB; break;
case EFDCTUIZ: ret = XB; break;
case EFDDIV: ret = AB; break;
case EFDMUL: ret = AB; break;
case EFDNABS: ret = XA; break;
case EFDNEG: ret = XA; break;
case EFDSUB: ret = AB; break;
}
return ret;
}
int do_spe_mathemu(struct pt_regs *regs)
{
FP_DECL_EX;
int IR, cmp;
unsigned long type, func, fc, fa, fb, src, speinsn;
union dw_union vc, va, vb;
if (get_user(speinsn, (unsigned int __user *) regs->nip))
return -EFAULT;
if ((speinsn >> 26) != EFAPU)
return -EINVAL; /* not an spe instruction */
type = insn_type(speinsn);
if (type == NOTYPE)
goto illegal;
func = speinsn & 0x7ff;
fc = (speinsn >> 21) & 0x1f;
fa = (speinsn >> 16) & 0x1f;
fb = (speinsn >> 11) & 0x1f;
src = (speinsn >> 5) & 0x7;
vc.wp[0] = current->thread.evr[fc];
vc.wp[1] = regs->gpr[fc];
va.wp[0] = current->thread.evr[fa];
va.wp[1] = regs->gpr[fa];
vb.wp[0] = current->thread.evr[fb];
vb.wp[1] = regs->gpr[fb];
__FPU_FPSCR = mfspr(SPRN_SPEFSCR);
pr_debug("speinsn:%08lx spefscr:%08lx\n", speinsn, __FPU_FPSCR);
pr_debug("vc: %08x %08x\n", vc.wp[0], vc.wp[1]);
pr_debug("va: %08x %08x\n", va.wp[0], va.wp[1]);
pr_debug("vb: %08x %08x\n", vb.wp[0], vb.wp[1]);
switch (src) {
case SPFP: {
FP_DECL_S(SA); FP_DECL_S(SB); FP_DECL_S(SR);
switch (type) {
case AB:
case XCR:
FP_UNPACK_SP(SA, va.wp + 1);
case XB:
FP_UNPACK_SP(SB, vb.wp + 1);
break;
case XA:
FP_UNPACK_SP(SA, va.wp + 1);
break;
}
pr_debug("SA: %ld %08lx %ld (%ld)\n", SA_s, SA_f, SA_e, SA_c);
pr_debug("SB: %ld %08lx %ld (%ld)\n", SB_s, SB_f, SB_e, SB_c);
switch (func) {
case EFSABS:
vc.wp[1] = va.wp[1] & ~SIGN_BIT_S;
goto update_regs;
case EFSNABS:
vc.wp[1] = va.wp[1] | SIGN_BIT_S;
goto update_regs;
case EFSNEG:
vc.wp[1] = va.wp[1] ^ SIGN_BIT_S;
goto update_regs;
case EFSADD:
FP_ADD_S(SR, SA, SB);
goto pack_s;
case EFSSUB:
FP_SUB_S(SR, SA, SB);
goto pack_s;
case EFSMUL:
FP_MUL_S(SR, SA, SB);
goto pack_s;
case EFSDIV:
FP_DIV_S(SR, SA, SB);
goto pack_s;
case EFSCMPEQ:
cmp = 0;
goto cmp_s;
case EFSCMPGT:
cmp = 1;
goto cmp_s;
case EFSCMPLT:
cmp = -1;
goto cmp_s;
case EFSCTSF:
case EFSCTUF:
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (SB_c == FP_CLS_NAN) {
vc.wp[1] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
SB_e += (func == EFSCTSF ? 31 : 32);
FP_TO_INT_ROUND_S(vc.wp[1], SB, 32,
(func == EFSCTSF));
}
goto update_regs;
case EFSCFD: {
FP_DECL_D(DB);
FP_CLEAR_EXCEPTIONS;
FP_UNPACK_DP(DB, vb.dp);
pr_debug("DB: %ld %08lx %08lx %ld (%ld)\n",
DB_s, DB_f1, DB_f0, DB_e, DB_c);
FP_CONV(S, D, 1, 2, SR, DB);
goto pack_s;
}
case EFSCTSI:
case EFSCTUI:
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (SB_c == FP_CLS_NAN) {
vc.wp[1] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
FP_TO_INT_ROUND_S(vc.wp[1], SB, 32,
((func & 0x3) != 0));
}
goto update_regs;
case EFSCTSIZ:
case EFSCTUIZ:
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (SB_c == FP_CLS_NAN) {
vc.wp[1] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
FP_TO_INT_S(vc.wp[1], SB, 32,
((func & 0x3) != 0));
}
goto update_regs;
default:
goto illegal;
}
break;
pack_s:
pr_debug("SR: %ld %08lx %ld (%ld)\n", SR_s, SR_f, SR_e, SR_c);
FP_PACK_SP(vc.wp + 1, SR);
goto update_regs;
cmp_s:
FP_CMP_S(IR, SA, SB, 3);
if (IR == 3 && (FP_ISSIGNAN_S(SA) || FP_ISSIGNAN_S(SB)))
FP_SET_EXCEPTION(FP_EX_INVALID);
if (IR == cmp) {
IR = 0x4;
} else {
IR = 0;
}
goto update_ccr;
}
case DPFP: {
FP_DECL_D(DA); FP_DECL_D(DB); FP_DECL_D(DR);
switch (type) {
case AB:
case XCR:
FP_UNPACK_DP(DA, va.dp);
case XB:
FP_UNPACK_DP(DB, vb.dp);
break;
case XA:
FP_UNPACK_DP(DA, va.dp);
break;
}
pr_debug("DA: %ld %08lx %08lx %ld (%ld)\n",
DA_s, DA_f1, DA_f0, DA_e, DA_c);
pr_debug("DB: %ld %08lx %08lx %ld (%ld)\n",
DB_s, DB_f1, DB_f0, DB_e, DB_c);
switch (func) {
case EFDABS:
vc.dp[0] = va.dp[0] & ~SIGN_BIT_D;
goto update_regs;
case EFDNABS:
vc.dp[0] = va.dp[0] | SIGN_BIT_D;
goto update_regs;
case EFDNEG:
vc.dp[0] = va.dp[0] ^ SIGN_BIT_D;
goto update_regs;
case EFDADD:
FP_ADD_D(DR, DA, DB);
goto pack_d;
case EFDSUB:
FP_SUB_D(DR, DA, DB);
goto pack_d;
case EFDMUL:
FP_MUL_D(DR, DA, DB);
goto pack_d;
case EFDDIV:
FP_DIV_D(DR, DA, DB);
goto pack_d;
case EFDCMPEQ:
cmp = 0;
goto cmp_d;
case EFDCMPGT:
cmp = 1;
goto cmp_d;
case EFDCMPLT:
cmp = -1;
goto cmp_d;
case EFDCTSF:
case EFDCTUF:
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (DB_c == FP_CLS_NAN) {
vc.wp[1] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
DB_e += (func == EFDCTSF ? 31 : 32);
FP_TO_INT_ROUND_D(vc.wp[1], DB, 32,
(func == EFDCTSF));
}
goto update_regs;
case EFDCFS: {
FP_DECL_S(SB);
FP_CLEAR_EXCEPTIONS;
FP_UNPACK_SP(SB, vb.wp + 1);
pr_debug("SB: %ld %08lx %ld (%ld)\n",
SB_s, SB_f, SB_e, SB_c);
FP_CONV(D, S, 2, 1, DR, SB);
goto pack_d;
}
case EFDCTUIDZ:
case EFDCTSIDZ:
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (DB_c == FP_CLS_NAN) {
vc.dp[0] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
FP_TO_INT_D(vc.dp[0], DB, 64,
((func & 0x1) == 0));
}
goto update_regs;
case EFDCTUI:
case EFDCTSI:
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (DB_c == FP_CLS_NAN) {
vc.wp[1] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
FP_TO_INT_ROUND_D(vc.wp[1], DB, 32,
((func & 0x3) != 0));
}
goto update_regs;
case EFDCTUIZ:
case EFDCTSIZ:
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (DB_c == FP_CLS_NAN) {
vc.wp[1] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
FP_TO_INT_D(vc.wp[1], DB, 32,
((func & 0x3) != 0));
}
goto update_regs;
default:
goto illegal;
}
break;
pack_d:
pr_debug("DR: %ld %08lx %08lx %ld (%ld)\n",
DR_s, DR_f1, DR_f0, DR_e, DR_c);
FP_PACK_DP(vc.dp, DR);
goto update_regs;
cmp_d:
FP_CMP_D(IR, DA, DB, 3);
if (IR == 3 && (FP_ISSIGNAN_D(DA) || FP_ISSIGNAN_D(DB)))
FP_SET_EXCEPTION(FP_EX_INVALID);
if (IR == cmp) {
IR = 0x4;
} else {
IR = 0;
}
goto update_ccr;
}
case VCT: {
FP_DECL_S(SA0); FP_DECL_S(SB0); FP_DECL_S(SR0);
FP_DECL_S(SA1); FP_DECL_S(SB1); FP_DECL_S(SR1);
int IR0, IR1;
switch (type) {
case AB:
case XCR:
FP_UNPACK_SP(SA0, va.wp);
FP_UNPACK_SP(SA1, va.wp + 1);
case XB:
FP_UNPACK_SP(SB0, vb.wp);
FP_UNPACK_SP(SB1, vb.wp + 1);
break;
case XA:
FP_UNPACK_SP(SA0, va.wp);
FP_UNPACK_SP(SA1, va.wp + 1);
break;
}
pr_debug("SA0: %ld %08lx %ld (%ld)\n",
SA0_s, SA0_f, SA0_e, SA0_c);
pr_debug("SA1: %ld %08lx %ld (%ld)\n",
SA1_s, SA1_f, SA1_e, SA1_c);
pr_debug("SB0: %ld %08lx %ld (%ld)\n",
SB0_s, SB0_f, SB0_e, SB0_c);
pr_debug("SB1: %ld %08lx %ld (%ld)\n",
SB1_s, SB1_f, SB1_e, SB1_c);
switch (func) {
case EVFSABS:
vc.wp[0] = va.wp[0] & ~SIGN_BIT_S;
vc.wp[1] = va.wp[1] & ~SIGN_BIT_S;
goto update_regs;
case EVFSNABS:
vc.wp[0] = va.wp[0] | SIGN_BIT_S;
vc.wp[1] = va.wp[1] | SIGN_BIT_S;
goto update_regs;
case EVFSNEG:
vc.wp[0] = va.wp[0] ^ SIGN_BIT_S;
vc.wp[1] = va.wp[1] ^ SIGN_BIT_S;
goto update_regs;
case EVFSADD:
FP_ADD_S(SR0, SA0, SB0);
FP_ADD_S(SR1, SA1, SB1);
goto pack_vs;
case EVFSSUB:
FP_SUB_S(SR0, SA0, SB0);
FP_SUB_S(SR1, SA1, SB1);
goto pack_vs;
case EVFSMUL:
FP_MUL_S(SR0, SA0, SB0);
FP_MUL_S(SR1, SA1, SB1);
goto pack_vs;
case EVFSDIV:
FP_DIV_S(SR0, SA0, SB0);
FP_DIV_S(SR1, SA1, SB1);
goto pack_vs;
case EVFSCMPEQ:
cmp = 0;
goto cmp_vs;
case EVFSCMPGT:
cmp = 1;
goto cmp_vs;
case EVFSCMPLT:
cmp = -1;
goto cmp_vs;
case EVFSCTUF:
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
case EVFSCTSF:
if (SB0_c == FP_CLS_NAN) {
vc.wp[0] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
SB0_e += (func == EVFSCTSF ? 31 : 32);
FP_TO_INT_ROUND_S(vc.wp[0], SB0, 32,
(func == EVFSCTSF));
}
if (SB1_c == FP_CLS_NAN) {
vc.wp[1] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
SB1_e += (func == EVFSCTSF ? 31 : 32);
FP_TO_INT_ROUND_S(vc.wp[1], SB1, 32,
(func == EVFSCTSF));
}
goto update_regs;
case EVFSCTUI:
case EVFSCTSI:
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (SB0_c == FP_CLS_NAN) {
vc.wp[0] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
FP_TO_INT_ROUND_S(vc.wp[0], SB0, 32,
((func & 0x3) != 0));
}
if (SB1_c == FP_CLS_NAN) {
vc.wp[1] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
FP_TO_INT_ROUND_S(vc.wp[1], SB1, 32,
((func & 0x3) != 0));
}
goto update_regs;
case EVFSCTUIZ:
case EVFSCTSIZ:
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (SB0_c == FP_CLS_NAN) {
vc.wp[0] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
FP_TO_INT_S(vc.wp[0], SB0, 32,
((func & 0x3) != 0));
}
if (SB1_c == FP_CLS_NAN) {
vc.wp[1] = 0;
FP_SET_EXCEPTION(FP_EX_INVALID);
} else {
FP_TO_INT_S(vc.wp[1], SB1, 32,
((func & 0x3) != 0));
}
goto update_regs;
default:
goto illegal;
}
break;
pack_vs:
pr_debug("SR0: %ld %08lx %ld (%ld)\n",
SR0_s, SR0_f, SR0_e, SR0_c);
pr_debug("SR1: %ld %08lx %ld (%ld)\n",
SR1_s, SR1_f, SR1_e, SR1_c);
FP_PACK_SP(vc.wp, SR0);
FP_PACK_SP(vc.wp + 1, SR1);
goto update_regs;
cmp_vs:
{
int ch, cl;
FP_CMP_S(IR0, SA0, SB0, 3);
FP_CMP_S(IR1, SA1, SB1, 3);
if (IR0 == 3 && (FP_ISSIGNAN_S(SA0) || FP_ISSIGNAN_S(SB0)))
FP_SET_EXCEPTION(FP_EX_INVALID);
if (IR1 == 3 && (FP_ISSIGNAN_S(SA1) || FP_ISSIGNAN_S(SB1)))
FP_SET_EXCEPTION(FP_EX_INVALID);
ch = (IR0 == cmp) ? 1 : 0;
cl = (IR1 == cmp) ? 1 : 0;
IR = (ch << 3) | (cl << 2) | ((ch | cl) << 1) |
((ch & cl) << 0);
goto update_ccr;
}
}
default:
return -EINVAL;
}
update_ccr:
regs->ccr &= ~(15 << ((7 - ((speinsn >> 23) & 0x7)) << 2));
regs->ccr |= (IR << ((7 - ((speinsn >> 23) & 0x7)) << 2));
update_regs:
powerpc: fix exception clearing in e500 SPE float emulation 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>
2013-12-11 07:07:45 +08:00
/*
* If the "invalid" exception sticky bit was set by the
* processor for non-finite input, but was not set before the
* instruction being emulated, clear it. Likewise for the
* "underflow" bit, which may have been set by the processor
* for exact underflow, not just inexact underflow when the
* flag should be set for IEEE 754 semantics. Other sticky
* exceptions will only be set by the processor when they are
* correct according to IEEE 754 semantics, and we must not
* clear sticky bits that were already set before the emulated
* instruction as they represent the user-visible sticky
* exception status. "inexact" traps to kernel are not
* required for IEEE semantics and are not enabled by default,
* so the "inexact" sticky bit may have been set by a previous
* instruction without the kernel being aware of it.
*/
__FPU_FPSCR
&= ~(FP_EX_INVALID | FP_EX_UNDERFLOW) | current->thread.spefscr_last;
__FPU_FPSCR |= (FP_CUR_EXCEPTIONS & FP_EX_MASK);
mtspr(SPRN_SPEFSCR, __FPU_FPSCR);
powerpc: fix exception clearing in e500 SPE float emulation 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>
2013-12-11 07:07:45 +08:00
current->thread.spefscr_last = __FPU_FPSCR;
current->thread.evr[fc] = vc.wp[0];
regs->gpr[fc] = vc.wp[1];
pr_debug("ccr = %08lx\n", regs->ccr);
pr_debug("cur exceptions = %08x spefscr = %08lx\n",
FP_CUR_EXCEPTIONS, __FPU_FPSCR);
pr_debug("vc: %08x %08x\n", vc.wp[0], vc.wp[1]);
pr_debug("va: %08x %08x\n", va.wp[0], va.wp[1]);
pr_debug("vb: %08x %08x\n", vb.wp[0], vb.wp[1]);
if (current->thread.fpexc_mode & PR_FP_EXC_SW_ENABLE) {
if ((FP_CUR_EXCEPTIONS & FP_EX_DIVZERO)
&& (current->thread.fpexc_mode & PR_FP_EXC_DIV))
return 1;
if ((FP_CUR_EXCEPTIONS & FP_EX_OVERFLOW)
&& (current->thread.fpexc_mode & PR_FP_EXC_OVF))
return 1;
if ((FP_CUR_EXCEPTIONS & FP_EX_UNDERFLOW)
&& (current->thread.fpexc_mode & PR_FP_EXC_UND))
return 1;
if ((FP_CUR_EXCEPTIONS & FP_EX_INEXACT)
&& (current->thread.fpexc_mode & PR_FP_EXC_RES))
return 1;
if ((FP_CUR_EXCEPTIONS & FP_EX_INVALID)
&& (current->thread.fpexc_mode & PR_FP_EXC_INV))
return 1;
}
return 0;
illegal:
if (have_e500_cpu_a005_erratum) {
/* according to e500 cpu a005 erratum, reissue efp inst */
regs->nip -= 4;
pr_debug("re-issue efp inst: %08lx\n", speinsn);
return 0;
}
printk(KERN_ERR "\nOoops! IEEE-754 compliance handler encountered un-supported instruction.\ninst code: %08lx\n", speinsn);
return -ENOSYS;
}
int speround_handler(struct pt_regs *regs)
{
union dw_union fgpr;
int s_lo, s_hi;
int lo_inexact, hi_inexact;
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
int fp_result;
unsigned long speinsn, type, fb, fc, fptype, func;
if (get_user(speinsn, (unsigned int __user *) regs->nip))
return -EFAULT;
if ((speinsn >> 26) != 4)
return -EINVAL; /* not an spe instruction */
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
func = speinsn & 0x7ff;
type = insn_type(func);
if (type == XCR) return -ENOSYS;
__FPU_FPSCR = mfspr(SPRN_SPEFSCR);
pr_debug("speinsn:%08lx spefscr:%08lx\n", speinsn, __FPU_FPSCR);
fptype = (speinsn >> 5) & 0x7;
/* No need to round if the result is exact */
lo_inexact = __FPU_FPSCR & (SPEFSCR_FG | SPEFSCR_FX);
hi_inexact = __FPU_FPSCR & (SPEFSCR_FGH | SPEFSCR_FXH);
if (!(lo_inexact || (hi_inexact && fptype == VCT)))
return 0;
fc = (speinsn >> 21) & 0x1f;
s_lo = regs->gpr[fc] & SIGN_BIT_S;
s_hi = current->thread.evr[fc] & SIGN_BIT_S;
fgpr.wp[0] = current->thread.evr[fc];
fgpr.wp[1] = regs->gpr[fc];
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
fb = (speinsn >> 11) & 0x1f;
switch (func) {
case EFSCTUIZ:
case EFSCTSIZ:
case EVFSCTUIZ:
case EVFSCTSIZ:
case EFDCTUIDZ:
case EFDCTSIDZ:
case EFDCTUIZ:
case EFDCTSIZ:
/*
* These instructions always round to zero,
* independent of the rounding mode.
*/
return 0;
case EFSCTUI:
case EFSCTUF:
case EVFSCTUI:
case EVFSCTUF:
case EFDCTUI:
case EFDCTUF:
fp_result = 0;
s_lo = 0;
s_hi = 0;
break;
case EFSCTSI:
case EFSCTSF:
fp_result = 0;
/* Recover the sign of a zero result if possible. */
if (fgpr.wp[1] == 0)
s_lo = regs->gpr[fb] & SIGN_BIT_S;
break;
case EVFSCTSI:
case EVFSCTSF:
fp_result = 0;
/* Recover the sign of a zero result if possible. */
if (fgpr.wp[1] == 0)
s_lo = regs->gpr[fb] & SIGN_BIT_S;
if (fgpr.wp[0] == 0)
s_hi = current->thread.evr[fb] & SIGN_BIT_S;
break;
case EFDCTSI:
case EFDCTSF:
fp_result = 0;
s_hi = s_lo;
/* Recover the sign of a zero result if possible. */
if (fgpr.wp[1] == 0)
s_hi = current->thread.evr[fb] & SIGN_BIT_S;
break;
default:
fp_result = 1;
break;
}
pr_debug("round fgpr: %08x %08x\n", fgpr.wp[0], fgpr.wp[1]);
switch (fptype) {
/* Since SPE instructions on E500 core can handle round to nearest
* and round toward zero with IEEE-754 complied, we just need
* to handle round toward +Inf and round toward -Inf by software.
*/
case SPFP:
if ((FP_ROUNDMODE) == FP_RND_PINF) {
if (!s_lo) fgpr.wp[1]++; /* Z > 0, choose Z1 */
} else { /* round to -Inf */
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (s_lo) {
if (fp_result)
fgpr.wp[1]++; /* Z < 0, choose Z2 */
else
fgpr.wp[1]--; /* Z < 0, choose Z2 */
}
}
break;
case DPFP:
if (FP_ROUNDMODE == FP_RND_PINF) {
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (!s_hi) {
if (fp_result)
fgpr.dp[0]++; /* Z > 0, choose Z1 */
else
fgpr.wp[1]++; /* Z > 0, choose Z1 */
}
} else { /* round to -Inf */
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (s_hi) {
if (fp_result)
fgpr.dp[0]++; /* Z < 0, choose Z2 */
else
fgpr.wp[1]--; /* Z < 0, choose Z2 */
}
}
break;
case VCT:
if (FP_ROUNDMODE == FP_RND_PINF) {
if (lo_inexact && !s_lo)
fgpr.wp[1]++; /* Z_low > 0, choose Z1 */
if (hi_inexact && !s_hi)
fgpr.wp[0]++; /* Z_high word > 0, choose Z1 */
} else { /* round to -Inf */
powerpc: fix e500 SPE float to integer and fixed-point conversions 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 afc0a07d4a283599ac3a6a31d7454e9baaeccca0. 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>
2013-11-05 00:54:46 +08:00
if (lo_inexact && s_lo) {
if (fp_result)
fgpr.wp[1]++; /* Z_low < 0, choose Z2 */
else
fgpr.wp[1]--; /* Z_low < 0, choose Z2 */
}
if (hi_inexact && s_hi) {
if (fp_result)
fgpr.wp[0]++; /* Z_high < 0, choose Z2 */
else
fgpr.wp[0]--; /* Z_high < 0, choose Z2 */
}
}
break;
default:
return -EINVAL;
}
current->thread.evr[fc] = fgpr.wp[0];
regs->gpr[fc] = fgpr.wp[1];
pr_debug(" to fgpr: %08x %08x\n", fgpr.wp[0], fgpr.wp[1]);
if (current->thread.fpexc_mode & PR_FP_EXC_SW_ENABLE)
return (current->thread.fpexc_mode & PR_FP_EXC_RES) ? 1 : 0;
return 0;
}
int __init spe_mathemu_init(void)
{
u32 pvr, maj, min;
pvr = mfspr(SPRN_PVR);
if ((PVR_VER(pvr) == PVR_VER_E500V1) ||
(PVR_VER(pvr) == PVR_VER_E500V2)) {
maj = PVR_MAJ(pvr);
min = PVR_MIN(pvr);
/*
* E500 revision below 1.1, 2.3, 3.1, 4.1, 5.1
* need cpu a005 errata workaround
*/
switch (maj) {
case 1:
if (min < 1)
have_e500_cpu_a005_erratum = 1;
break;
case 2:
if (min < 3)
have_e500_cpu_a005_erratum = 1;
break;
case 3:
case 4:
case 5:
if (min < 1)
have_e500_cpu_a005_erratum = 1;
break;
default:
break;
}
}
return 0;
}
module_init(spe_mathemu_init);