Implement extended LWSP/SWSP instruction subdecoding for the purpose of
unaligned GP-relative memory access emulation.
With the introduction of the MIPS16e2 ASE[1] the previously must-be-zero
3-bit field at bits 7..5 of the extended encodings of the instructions
selected with the LWSP and SWSP major opcodes has become a `sel' field,
acting as an opcode extension for additional operations. In both cases
the `sel' value of 0 has retained the original operation, that is:
LW rx, offset(sp)
and:
SW rx, offset(sp)
for LWSP and SWSP respectively. In hardware predating the MIPS16e2 ASE
other values may or may not have been decoded, architecturally yielding
unpredictable results, and in our unaligned memory access emulation we
have treated the 3-bit field as a don't-care, that is effectively making
all the possible encodings of the field alias to the architecturally
defined encoding of 0.
For the non-zero values of the `sel' field the MIPS16e2 ASE has in
particular defined these GP-relative operations:
LW rx, offset(gp) # sel = 1
LH rx, offset(gp) # sel = 2
LHU rx, offset(gp) # sel = 4
and
SW rx, offset(gp) # sel = 1
SH rx, offset(gp) # sel = 2
for LWSP and SWSP respectively, which will trap with an Address Error
exception if the effective address calculated is not naturally-aligned
for the operation requested. These operations have been selected for
unaligned access emulation, for consistency with the corresponding
regular MIPS and microMIPS operations.
For other non-zero values of the `sel' field the MIPS16e2 ASE has
defined further operations, which however either never trap with an
Address Error exception, such as LWL or GP-relative SB, or are not
supposed to be emulated, such as LL or SC. These operations have been
selected to exclude from unaligned access emulation, should an Address
Error exception ever happen with them.
Subdecode the `sel' field in unaligned access emulation then for the
extended encodings of the instructions selected with the LWSP and SWSP
major opcodes, whenever support for the MIPS16e2 ASE has been detected
in hardware, and either emulate the operation requested or send SIGBUS
to the originating process, according to the selection described above.
For hardware implementing the MIPS16 ASE, however lacking MIPS16e2 ASE
support retain the original interpretation of the `sel' field.
The effects of this change are illustrated with the following user
program:
$ cat mips16e2-test.c
#include <inttypes.h>
#include <stdio.h>
int main(void)
{
int64_t scratch[16] = { 0 };
int32_t *tmp0, *tmp1, *tmp2;
int i;
scratch[0] = 0xc8c7c6c5c4c3c2c1;
scratch[1] = 0xd0cfcecdcccbcac9;
asm volatile(
"move %0, $sp\n\t"
"move %1, $gp\n\t"
"move $sp, %4\n\t"
"addiu %2, %4, 8\n\t"
"move $gp, %2\n\t"
"lw %2, 2($sp)\n\t"
"sw %2, 16(%4)\n\t"
"lw %2, 2($gp)\n\t"
"sw %2, 24(%4)\n\t"
"lw %2, 1($sp)\n\t"
"sw %2, 32(%4)\n\t"
"lh %2, 1($gp)\n\t"
"sw %2, 40(%4)\n\t"
"lw %2, 3($sp)\n\t"
"sw %2, 48(%4)\n\t"
"lhu %2, 3($gp)\n\t"
"sw %2, 56(%4)\n\t"
"lw %2, 0(%4)\n\t"
"sw %2, 66($sp)\n\t"
"lw %2, 8(%4)\n\t"
"sw %2, 82($gp)\n\t"
"lw %2, 0(%4)\n\t"
"sw %2, 97($sp)\n\t"
"lw %2, 8(%4)\n\t"
"sh %2, 113($gp)\n\t"
"move $gp, %1\n\t"
"move $sp, %0"
: "=&d" (tmp0), "=&d" (tmp1), "=&d" (tmp2), "=m" (scratch)
: "d" (scratch));
for (i = 0; i < sizeof(scratch) / sizeof(*scratch); i += 2)
printf("%016" PRIx64 "\t%016" PRIx64 "\n",
scratch[i], scratch[i + 1]);
return 0;
}
$
to be compiled with:
$ gcc -mips16 -mips32r2 -Wa,-mmips16e2 -o mips16e2-test mips16e2-test.c
$
With 74Kf hardware, which does not implement the MIPS16e2 ASE, this
program produces the following output:
$ ./mips16e2-test
c8c7c6c5c4c3c2c1 d0cfcecdcccbcac9
00000000c6c5c4c3 00000000c6c5c4c3
00000000c5c4c3c2 00000000c5c4c3c2
00000000c7c6c5c4 00000000c7c6c5c4
0000c4c3c2c10000 0000000000000000
0000cccbcac90000 0000000000000000
000000c4c3c2c100 0000000000000000
000000cccbcac900 0000000000000000
$
regardless of whether the change has been applied or not.
With the change not applied and interAptive MR2 hardware[2], which does
implement the MIPS16e2 ASE, it produces the following output:
$ ./mips16e2-test
c8c7c6c5c4c3c2c1 d0cfcecdcccbcac9
00000000c6c5c4c3 00000000cecdcccb
00000000c5c4c3c2 00000000cdcccbca
00000000c7c6c5c4 00000000cfcecdcc
0000c4c3c2c10000 0000000000000000
0000000000000000 0000cccbcac90000
000000c4c3c2c100 0000000000000000
0000000000000000 000000cccbcac900
$
which shows that for GP-relative operations the correct trapping address
calculated from $gp has been obtained from the CP0 BadVAddr register and
so has data from the source operand, however masking and extension has
not been applied for halfword operations.
With the change applied and interAptive MR2 hardware the program
produces the following output:
$ ./mips16e2-test
c8c7c6c5c4c3c2c1 d0cfcecdcccbcac9
00000000c6c5c4c3 00000000cecdcccb
00000000c5c4c3c2 00000000ffffcbca
00000000c7c6c5c4 000000000000cdcc
0000c4c3c2c10000 0000000000000000
0000000000000000 0000cccbcac90000
000000c4c3c2c100 0000000000000000
0000000000000000 0000000000cac900
$
as expected.
References:
[1] "MIPS32 Architecture for Programmers: MIPS16e2 Application-Specific
Extension Technical Reference Manual", Imagination Technologies
Ltd., Document Number: MD01172, Revision 01.00, April 26, 2016
[2] "MIPS32 interAptiv Multiprocessing System Software User's Manual",
Imagination Technologies Ltd., Document Number: MD00904, Revision
02.01, June 15, 2016, Chapter 24 "MIPS16e Application-Specific
Extension to the MIPS32 Instruction Set", pp. 871-883
Signed-off-by: Maciej W. Rozycki <macro@xxxxxxxxxx>
---
NB a recent binutils version, as from commit 25499ac7ee92 ("MIPS16e2:
Add MIPS16e2 ASE support"), is required to build the test program.
Maciej
linux-mips16e2-ase-emul.diff
Index: linux-sfr-test/arch/mips/kernel/unaligned.c
===================================================================
--- linux-sfr-test.orig/arch/mips/kernel/unaligned.c 2017-05-22 22:42:16.000000000 +0100
+++ linux-sfr-test/arch/mips/kernel/unaligned.c 2017-05-22 22:54:28.686096000 +0100
@@ -1984,6 +1984,8 @@ static void emulate_load_store_MIPS16e(s
u16 __user *pc16;
unsigned long origpc;
union mips16e_instruction mips16inst, oldinst;
+ unsigned int opcode;
+ int extended = 0;
origpc = regs->cp0_epc;
orig31 = regs->regs[31];
@@ -1996,6 +1998,7 @@ static void emulate_load_store_MIPS16e(s
/* skip EXTEND instruction */
if (mips16inst.ri.opcode == MIPS16e_extend_op) {
+ extended = 1;
pc16++;
__get_user(mips16inst.full, pc16);
} else if (delay_slot(regs)) {
@@ -2008,7 +2011,8 @@ static void emulate_load_store_MIPS16e(s
goto sigbus;
}
- switch (mips16inst.ri.opcode) {
+ opcode = mips16inst.ri.opcode;
+ switch (opcode) {
case MIPS16e_i64_op: /* I64 or RI64 instruction */
switch (mips16inst.i64.func) { /* I64/RI64 func field check */
case MIPS16e_ldpc_func:
@@ -2028,9 +2032,40 @@ static void emulate_load_store_MIPS16e(s
goto sigbus;
case MIPS16e_swsp_op:
+ reg = reg16to32[mips16inst.ri.rx];
+ if (extended && cpu_has_mips16e2)
+ switch (mips16inst.ri.imm >> 5) {
+ case 0: /* SWSP */
+ case 1: /* SWGP */
+ break;
+ case 2: /* SHGP */
+ opcode = MIPS16e_sh_op;
+ break;
+ default:
+ goto sigbus;
+ }
+ break;
+
case MIPS16e_lwpc_op:
+ reg = reg16to32[mips16inst.ri.rx];
+ break;
+
case MIPS16e_lwsp_op:
reg = reg16to32[mips16inst.ri.rx];
+ if (extended && cpu_has_mips16e2)
+ switch (mips16inst.ri.imm >> 5) {
+ case 0: /* LWSP */
+ case 1: /* LWGP */
+ break;
+ case 2: /* LHGP */
+ opcode = MIPS16e_lh_op;
+ break;
+ case 4: /* LHUGP */
+ opcode = MIPS16e_lhu_op;
+ break;
+ default:
+ goto sigbus;
+ }
break;
case MIPS16e_i8_op:
@@ -2044,7 +2079,7 @@ static void emulate_load_store_MIPS16e(s
break;
}
- switch (mips16inst.ri.opcode) {
+ switch (opcode) {
case MIPS16e_lb_op:
case MIPS16e_lbu_op:
