Hexagon (target/hexagon) Fix predicated assignment to .tmp and .cur

Here are example instructions with a predicated .tmp/.cur assignment
    if (p1) v12.tmp = vmem(r7 + #0)
    if (p0) v12.cur = vmem(r9 + #0)
The .tmp/.cur indicates that references to v12 in the same packet
take the result of the load.  However, when the predicate is false,
the value at the start of the packet should be used.  After the packet
commits, the .tmp value is dropped, but the .cur value is maintained.

To fix this bug, we preload the original value from the HVX register
into the temporary used for the result.

Test cases added to tests/tcg/hexagon/hvx_misc.c

Acked-by: Richard Henderson <richard.henderson@linaro.org>
Co-authored-by: Matheus Tavares Bernardino <quic_mathbern@quicinc.com>
Signed-off-by: Matheus Tavares Bernardino <quic_mathbern@quicinc.com>
Signed-off-by: Taylor Simpson <tsimpson@quicinc.com>
Message-Id: <20221108162906.3166-3-tsimpson@quicinc.com>

1 files changed, 72 insertions, 0 deletions

diff --git a/tests/tcg/hexagon/hvx_misc.c b/tests/tcg/hexagon/hvx_misc.c
index 6e2c9ab3cd..53d5c9b44f 100644
--- a/tests/tcg/hexagon/hvx_misc.c
+++ b/tests/tcg/hexagon/hvx_misc.c
@@ -541,6 +541,75 @@ static void test_vshuff(void)
     check_output_b(__LINE__, 1);
 }
 
+static void test_load_tmp_predicated(void)
+{
+    void *p0 = buffer0;
+    void *p1 = buffer1;
+    void *pout = output;
+    bool pred = true;
+
+    for (int i = 0; i < BUFSIZE; i++) {
+        /*
+         * Load into v12 as .tmp with a predicate
+         * When the predicate is true, we get the vector from buffer1[i]
+         * When the predicate is false, we get a vector of all 1's
+         * Regardless of the predicate, the next packet should have
+         * a vector of all 1's
+         */
+        asm("v3 = vmem(%0 + #0)\n\t"
+            "r1 = #1\n\t"
+            "v12 = vsplat(r1)\n\t"
+            "p1 = !cmp.eq(%3, #0)\n\t"
+            "{\n\t"
+            "    if (p1) v12.tmp = vmem(%1 + #0)\n\t"
+            "    v4.w = vadd(v12.w, v3.w)\n\t"
+            "}\n\t"
+            "v4.w = vadd(v4.w, v12.w)\n\t"
+            "vmem(%2 + #0) = v4\n\t"
+            : : "r"(p0), "r"(p1), "r"(pout), "r"(pred)
+            : "r1", "p1", "v12", "v3", "v4", "v6", "memory");
+        p0 += sizeof(MMVector);
+        p1 += sizeof(MMVector);
+        pout += sizeof(MMVector);
+
+        for (int j = 0; j < MAX_VEC_SIZE_BYTES / 4; j++) {
+            expect[i].w[j] =
+                pred ? buffer0[i].w[j] + buffer1[i].w[j] + 1
+                     : buffer0[i].w[j] + 2;
+        }
+        pred = !pred;
+    }
+
+    check_output_w(__LINE__, BUFSIZE);
+}
+
+static void test_load_cur_predicated(void)
+{
+    bool pred = true;
+    for (int i = 0; i < BUFSIZE; i++) {
+        asm volatile("p0 = !cmp.eq(%3, #0)\n\t"
+                     "v3 = vmem(%0+#0)\n\t"
+                     /*
+                      * Preload v4 to make sure that the assignment from the
+                      * packet below is not being ignored when pred is false.
+                      */
+                     "r0 = #0x01237654\n\t"
+                     "v4 = vsplat(r0)\n\t"
+                     "{\n\t"
+                     "    if (p0) v3.cur = vmem(%1+#0)\n\t"
+                     "    v4 = v3\n\t"
+                     "}\n\t"
+                     "vmem(%2+#0) = v4\n\t"
+                     :
+                     : "r"(&buffer0[i]), "r"(&buffer1[i]),
+                       "r"(&output[i]), "r"(pred)
+                     : "r0", "p0", "v3", "v4", "memory");
+        expect[i] = pred ? buffer1[i] : buffer0[i];
+        pred = !pred;
+    }
+    check_output_w(__LINE__, BUFSIZE);
+}
+
 int main()
 {
     init_buffers();
@@ -578,6 +647,9 @@ int main()
 
     test_vshuff();
 
+    test_load_tmp_predicated();
+    test_load_cur_predicated();
+
     puts(err ? "FAIL" : "PASS");
     return err ? 1 : 0;
 }