target/arm: Fix bfdotadd_ebf vs nan selection

Implement FPProcessNaNs4 within bfdotadd_ebf, rather than
simply letting NaNs propagate through the function.

Cc: [email protected]
Fixes: 0e18501 ("target/arm: Implement FPCR.EBF=1 semantics for bfdotadd()")
Signed-off-by: Richard Henderson <[email protected]>
Reviewed-by: Peter Maydell <[email protected]>
Message-id: [email protected]
Signed-off-by: Peter Maydell <[email protected]>
(cherry picked from commit bf020eaa6741711902a425016e2c7585f222562d)
Signed-off-by: Michael Tokarev <[email protected]>

Author: rth7680

target/arm/tcg/vec_helper.c

@@ -2989,31 +2989,62 @@ float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2, float_status *fpst)
 float32 bfdotadd_ebf(float32 sum, uint32_t e1, uint32_t e2,
                      float_status *fpst, float_status *fpst_odd)
 {
-    /*
-     * Compare f16_dotadd() in sme_helper.c, but here we have
-     * bfloat16 inputs. In particular that means that we do not
-     * want the FPCR.FZ16 flush semantics, so we use the normal
-     * float_status for the input handling here.
-     */
-    float64 e1r = float32_to_float64(e1 << 16, fpst);
-    float64 e1c = float32_to_float64(e1 & 0xffff0000u, fpst);
-    float64 e2r = float32_to_float64(e2 << 16, fpst);
-    float64 e2c = float32_to_float64(e2 & 0xffff0000u, fpst);
-    float64 t64;
+    float32 s1r = e1 << 16;
+    float32 s1c = e1 & 0xffff0000u;
+    float32 s2r = e2 << 16;
+    float32 s2c = e2 & 0xffff0000u;
     float32 t32;
 
-    /*
-     * The ARM pseudocode function FPDot performs both multiplies
-     * and the add with a single rounding operation.  Emulate this
-     * by performing the first multiply in round-to-odd, then doing
-     * the second multiply as fused multiply-add, and rounding to
-     * float32 all in one step.
-     */
-    t64 = float64_mul(e1r, e2r, fpst_odd);
-    t64 = float64r32_muladd(e1c, e2c, t64, 0, fpst);
+    /* C.f. FPProcessNaNs4 */
+    if (float32_is_any_nan(s1r) || float32_is_any_nan(s1c) ||
+        float32_is_any_nan(s2r) || float32_is_any_nan(s2c)) {
+        if (float32_is_signaling_nan(s1r, fpst)) {
+            t32 = s1r;
+        } else if (float32_is_signaling_nan(s1c, fpst)) {
+            t32 = s1c;
+        } else if (float32_is_signaling_nan(s2r, fpst)) {
+            t32 = s2r;
+        } else if (float32_is_signaling_nan(s2c, fpst)) {
+            t32 = s2c;
+        } else if (float32_is_any_nan(s1r)) {
+            t32 = s1r;
+        } else if (float32_is_any_nan(s1c)) {
+            t32 = s1c;
+        } else if (float32_is_any_nan(s2r)) {
+            t32 = s2r;
+        } else {
+            t32 = s2c;
+        }
+        /*
+         * FPConvertNaN(FPProcessNaN(t32)) will be done as part
+         * of the final addition below.
+         */
+    } else {
+        /*
+         * Compare f16_dotadd() in sme_helper.c, but here we have
+         * bfloat16 inputs. In particular that means that we do not
+         * want the FPCR.FZ16 flush semantics, so we use the normal
+         * float_status for the input handling here.
+         */
+        float64 e1r = float32_to_float64(s1r, fpst);
+        float64 e1c = float32_to_float64(s1c, fpst);
+        float64 e2r = float32_to_float64(s2r, fpst);
+        float64 e2c = float32_to_float64(s2c, fpst);
+        float64 t64;
 
-    /* This conversion is exact, because we've already rounded. */
-    t32 = float64_to_float32(t64, fpst);
+        /*
+         * The ARM pseudocode function FPDot performs both multiplies
+         * and the add with a single rounding operation.  Emulate this
+         * by performing the first multiply in round-to-odd, then doing
+         * the second multiply as fused multiply-add, and rounding to
+         * float32 all in one step.
+         */
+        t64 = float64_mul(e1r, e2r, fpst_odd);
+        t64 = float64r32_muladd(e1c, e2c, t64, 0, fpst);
+
+        /* This conversion is exact, because we've already rounded. */
+        t32 = float64_to_float32(t64, fpst);
+    }
 
     /* The final accumulation step is not fused. */
     return float32_add(sum, t32, fpst);