summaryrefslogtreecommitdiffstats
path: root/results/classifier/accel-gemma3:12b/tcg/1612
diff options
context:
space:
mode:
Diffstat (limited to 'results/classifier/accel-gemma3:12b/tcg/1612')
-rw-r--r--results/classifier/accel-gemma3:12b/tcg/161252
1 files changed, 52 insertions, 0 deletions
diff --git a/results/classifier/accel-gemma3:12b/tcg/1612 b/results/classifier/accel-gemma3:12b/tcg/1612
new file mode 100644
index 00000000..896b4067
--- /dev/null
+++ b/results/classifier/accel-gemma3:12b/tcg/1612
@@ -0,0 +1,52 @@
+
+SVE first-faulting gather loads return incorrect data
+Description of problem:
+The results of `ldff1(b|h|w|d)` seem to be incorrect when `<Zt> == <Zm>`. The first element is duplicated throughout the vector while the FFR indicates that all elements were successfully loaded. This happens since https://gitlab.com/qemu-project/qemu/-/commit/50de9b78cec06e6d16e92a114a505779359ca532, and still happens on the latest master.
+Steps to reproduce:
+1. This assembly sequence loads data with an `ldff1d` instruction (and also loads the ffr). Note that with `ldff1d`, `<Zt> == <Zm>`.
+
+asmtest.s
+```
+ .type asmtest, @function
+ .balign 16
+ .global asmtest
+asmtest:
+ setffr
+ ptrue p0.d
+ index z1.d, #0, #1
+ ldff1d z1.d, p0/z, [x0, z1.d, LSL #3]
+ rdffr p1.b
+ st1d {z1.d}, p0, [x1]
+ str p1, [x2]
+ ret
+```
+
+This harness for convenience intialises some data and checks the element at index 1, which should be 1.
+
+test.c
+```
+#include <arm_sve.h>
+#include <stdio.h>
+
+void asmtest(int64_t const * data, svint64_t * loaded, svbool_t * ffr);
+
+int main() {
+ const int64_t data[] = {42, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
+ 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21,
+ 22, 23, 24, 25, 26, 27, 28, 29, 30, 31};
+ svint64_t loaded;
+ svbool_t ffr;
+
+ asmtest(data, &loaded, &ffr);
+
+ // Check value of element at index 1
+ svuint64_t lanes = svindex_u64(0, 1);
+ svbool_t lane = svcmpeq_n_u64(svptrue_b64(), lanes, 1);
+ printf("%ld\n", svaddv_s64(lane, loaded));
+}
+```
+
+2. ```clang-15 -fuse-ld=lld -march=armv8-a+sve2 -target aarch64-unknown-linux-gnu -static *.c *.s -o svldffgathertest```
+3. ```qemu-aarch64 svldffgathertest``` - the value printed should be 1, but it can be seen that all values in the loaded vector are 42.
+Additional information:
+The above code was successfully tested on real SVE hardware. Normal gathers work fine in QEMU, as does a non-gather first-fault load.