tcg/optimize: Fix sign_mask for logical right-shift

The 'sign' computation is attempting to locate the sign bit that has
been repeated, so that we can test if that bit is known zero.  That
computation can be zero if there are no known sign repetitions.

Cc: qemu-stable@nongnu.org
Fixes: 93a967fbb5 ("tcg/optimize: Propagate sign info for shifting")
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2248
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
(cherry picked from commit 2911e9b95f)
Signed-off-by: Michael Tokarev <mjt@tls.msk.ru>

tcg/optimize.c

@@ -2123,7 +2123,7 @@ static bool fold_shift(OptContext *ctx, TCGOp *op)
          * will not reduced the number of input sign repetitions.
          */
         sign = (s_mask & -s_mask) >> 1;
-        if (!(z_mask & sign)) {
+        if (sign && !(z_mask & sign)) {
             ctx->s_mask = s_mask;
         }
         break;

tests/tcg/aarch64/Makefile.target

@@ -10,6 +10,7 @@ VPATH 		+= $(AARCH64_SRC)
 
 # Base architecture tests
 AARCH64_TESTS=fcvt pcalign-a64 lse2-fault
+AARCH64_TESTS += test-2248
 
 fcvt: LDFLAGS+=-lm
 

tests/tcg/aarch64/test-2248.c

@@ -0,0 +1,28 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+/* See https://gitlab.com/qemu-project/qemu/-/issues/2248 */
+
+#include <assert.h>
+
+__attribute__((noinline))
+long test(long x, long y, long sh)
+{
+    long r;
+    asm("cmp   %1, %2\n\t"
+        "cset  x12, lt\n\t"
+        "and   w11, w12, #0xff\n\t"
+        "cmp   w11, #0\n\t"
+        "csetm x14, ne\n\t"
+        "lsr   x13, x14, %3\n\t"
+        "sxtb  %0, w13"
+        : "=r"(r)
+        : "r"(x), "r"(y), "r"(sh)
+        : "x11", "x12", "x13", "x14");
+    return r;
+}
+
+int main()
+{
+    long r = test(0, 1, 2);
+    assert(r == -1);
+    return 0;
+}