Skip to contentSkip to conversation

target/hexagon: accept valid packets rejected by check

PATCH In review Brian Cain 2 participants 6 messages Jul 24 – Jul 27

lore
Share & tools
lore

Brian Cain

author
#0 lore raw

PATCH 1/2 target/hexagon: accept valid packets rejected by check

has_valid_slot_assignment() rejected any packet with two instructions assigned to the same slot. That is too strict. When a memory instruction is encoded before a slot-flexible instruction in a packet, the descending slot assignment places the memory op in slot 1 and the other in slot 0, then the "mem insns to slot 0" fixup moves the memory op to slot 0 as well, leaving both in slot 0. Such a packet is valid and executes correctly, but the uniqueness test flagged it as HEX_CAUSE_INVALID_PACKET, raising SIGILL in linux-user and a precise exception in system mode.

For example this packet, with the load encoded first, was wrongly rejected:

    { r6 = memw(r3+#-4)
      r7 = #0x4ae6 }

Replace the uniqueness test with a slot-exhaustion check: walk the instructions in encoding order handing out slots in strictly decreasing order and fail only if an instruction has no valid slot at or below the running slot. This accepts packets that legally share a slot while still rejecting genuinely unassignable packets, such as a memory instruction grouped with a duplex, or a load followed by an instruction that requires a high slot.

---
 target/hexagon/decode.c | 30 ++++++++++++++++++++++--------
 1 file changed, 22 insertions(+), 8 deletions(-)
target/hexagon/decode.c +22 -8
index 6eddcca26ed..e6bb7733be8 100644
@@ -549,21 +549,35 @@ static bool decode_parsebits_is_loopend(uint32_t encoding32)
     return bits == 0x2;
 }
 
+/*
+ * Check that the packet's instructions can be grouped into slots: walk them
+ * in encoding order handing out slots in strictly decreasing order, and fail
+ * if an instruction has no valid slot at or below the running slot.  Two
+ * instructions may legally share a slot, so this does not require unique
+ * slots, only that every instruction fits.
+ */
 static bool has_valid_slot_assignment(Packet *pkt)
 {
-    int used_slots = 0;
-    for (int i = 0; i < pkt->num_insns; i++) {
-        int slot_mask;
-        Insn *insn = &pkt->insn[i];
-        if (decode_opcode_ends_loop(insn->opcode)) {
+    int i;
+    int slot = 3;
+
+    for (i = 0; i < pkt->num_insns; i++) {
+        SlotMask valid_slots;
+        if (decode_opcode_ends_loop(pkt->insn[i].opcode)) {
             /* We overload slot 0 for endloop. */
             continue;
         }
-        slot_mask = 1 << insn->slot;
-        if (used_slots & slot_mask) {
+        if (slot < 0) {
             return false;
         }
-        used_slots |= slot_mask;
+        valid_slots = get_valid_slots(pkt, i);
+        while (!(valid_slots & (1 << slot))) {
+            if (slot <= 0) {
+                return false;
+            }
+            slot--;
+        }
+        slot--;
     }
     return true;
 }
Signature · 1 line
2.34.1
author
#1 lore raw

PATCH 2/2 tests/tcg/hexagon: add slot-assignment tests

valid-slots: packets that legally share a slot and were wrongly rejected before the fix (load and transfer, load encoded first; dczeroa packed last with three transfers).

invalid-slots: unassignable packets that must still be rejected:
store + duplex, load + indirect jump, three logical ops competing for slots 2 and 3, and five ops for four slots.

---
 tests/tcg/hexagon/Makefile.target |  1 +
 tests/tcg/hexagon/invalid-slots.c | 72 +++++++++++++++++++++++++++++++
 tests/tcg/hexagon/valid-slots.c   | 62 ++++++++++++++++++++++++++
 3 files changed, 135 insertions(+)
 create mode 100644 tests/tcg/hexagon/valid-slots.c
tests/tcg/hexagon/Makefile.target +1 -0
index a2a0ffc69b2..80e56f7d055 100644
@@ -51,6 +51,7 @@ HEX_TESTS += scatter_gather
 HEX_TESTS += hvx_misc
 HEX_TESTS += hvx_histogram
 HEX_TESTS += invalid-slots
+HEX_TESTS += valid-slots
 HEX_TESTS += invalid-encoding
 HEX_TESTS += multiple-writes
 HEX_TESTS += unaligned_pc
tests/tcg/hexagon/invalid-slots.c +72 -0

84 lines not on this page · raw message

tests/tcg/hexagon/valid-slots.c new file +62 -0

62 lines not on this page · raw message

Signature · 1 line
2.34.1
Reviewed-by on v1 2/2
#2 lore raw

Re: [PATCH 2/2] tests/tcg/hexagon: add slot-assignment tests

Quoted 9 lines · On Fri, 24 Jul 2026 10:18:45 -0700 Brian Cain <[email protected]>…

valid-slots: packets that legally share a slot and were wrongly rejected before the fix (load and transfer, load encoded first; dczeroa packed last with three transfers).

invalid-slots: unassignable packets that must still be rejected:
store + duplex, load + indirect jump, three logical ops competing for slots 2 and 3, and five ops for four slots.

Quoted 7 lines
---
 tests/tcg/hexagon/Makefile.target |  1 +
 tests/tcg/hexagon/invalid-slots.c | 72 +++++++++++++++++++++++++++++++
 tests/tcg/hexagon/valid-slots.c   | 62 ++++++++++++++++++++++++++
 3 files changed, 135 insertions(+)
 create mode 100644 tests/tcg/hexagon/valid-slots.c
Reviewed-by on v1 2/2
#3 lore raw

Re: [PATCH 2/2] tests/tcg/hexagon: add slot-assignment tests

In reply to Brian Cain
Quoted 9 lines · On Fri, 24 Jul 2026 10:18:45 -0700 Brian Cain <[email protected]>…

valid-slots: packets that legally share a slot and were wrongly rejected before the fix (load and transfer, load encoded first; dczeroa packed last with three transfers).

invalid-slots: unassignable packets that must still be rejected:
store + duplex, load + indirect jump, three logical ops competing for slots 2 and 3, and five ops for four slots.

Quoted 7 lines
---
 tests/tcg/hexagon/Makefile.target |  1 +
 tests/tcg/hexagon/invalid-slots.c | 72 +++++++++++++++++++++++++++++++
 tests/tcg/hexagon/valid-slots.c   | 62 ++++++++++++++++++++++++++
 3 files changed, 135 insertions(+)
 create mode 100644 tests/tcg/hexagon/valid-slots.c
Reviewed-by on v1 2/2
#4 lore raw

Re: [PATCH 2/2] tests/tcg/hexagon: add slot-assignment tests

In reply to Brian Cain
Quoted 9 lines · On Fri, 24 Jul 2026 10:18:45 -0700 Brian Cain <[email protected]>…

valid-slots: packets that legally share a slot and were wrongly rejected before the fix (load and transfer, load encoded first; dczeroa packed last with three transfers).

invalid-slots: unassignable packets that must still be rejected:
store + duplex, load + indirect jump, three logical ops competing for slots 2 and 3, and five ops for four slots.

Quoted 7 lines
---
 tests/tcg/hexagon/Makefile.target |  1 +
 tests/tcg/hexagon/invalid-slots.c | 72 +++++++++++++++++++++++++++++++
 tests/tcg/hexagon/valid-slots.c   | 62 ++++++++++++++++++++++++++
 3 files changed, 135 insertions(+)
 create mode 100644 tests/tcg/hexagon/valid-slots.c
Reviewed-by on v1 1/2
#5 lore raw
In reply to Brian Cain
Quoted 25 lines · On Fri, 24 Jul 2026 10:18:44 -0700 Brian Cain <[email protected]>…

has_valid_slot_assignment() rejected any packet with two instructions assigned to the same slot. That is too strict. When a memory instruction is encoded before a slot-flexible instruction in a packet, the descending slot assignment places the memory op in slot 1 and the other in slot 0, then the "mem insns to slot 0" fixup moves the memory op to slot 0 as well, leaving both in slot 0. Such a packet is valid and executes correctly, but the uniqueness test flagged it as HEX_CAUSE_INVALID_PACKET, raising SIGILL in linux-user and a precise exception in system mode.

For example this packet, with the load encoded first, was wrongly rejected:

    { r6 = memw(r3+#-4)
      r7 = #0x4ae6 }

Replace the uniqueness test with a slot-exhaustion check: walk the instructions in encoding order handing out slots in strictly decreasing order and fail only if an instruction has no valid slot at or below the running slot. This accepts packets that legally share a slot while still rejecting genuinely unassignable packets, such as a memory instruction grouped with a duplex, or a load followed by an instruction that requires a high slot.

Quoted 4 lines
---
 target/hexagon/decode.c | 30 ++++++++++++++++++++++--------
 1 file changed, 22 insertions(+), 8 deletions(-)

(BTW, apologies for the repeated replies in the second patch. I had a bug
in my setting here)