[PATCH v3 48/48] tcg/optimize: Propagate sign info for shifting

[Richard Henderson <richard.henderson@linaro.org>]

For constant shifts, we can simply shift the s_mask.

For variable shifts, we know that sar does not reduce
the s_mask, which helps for sequences like

    ext32s_i64  t, in
    sar_i64     t, t, v
    ext32s_i64  out, t

allowing the final extend to be eliminated.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/optimize.c | 50 +++++++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 47 insertions(+), 3 deletions(-)

diff --git a/tcg/optimize.c b/tcg/optimize.c
index 3dd9219d69..5e239e00ee 100644
--- a/tcg/optimize.c
+++ b/tcg/optimize.c
@@ -85,6 +85,18 @@ static uint64_t smask_from_zmask(uint64_t zmask)
     return ~(~0ull >> rep);
 }
 
+/*
+ * Recreate a properly left-aligned smask after manipulation.
+ * Some bit-shuffling, particularly shifts and rotates, may
+ * retain sign bits on the left, but may scatter disconnected
+ * sign bits on the right.  Retain only what remains to the left.
+ */
+static uint64_t smask_from_smask(int64_t smask)
+{
+    /* Only the 1 bits are significant for smask */
+    return smask_from_zmask(~smask);
+}
+
 static inline TempOptInfo *ts_info(TCGTemp *ts)
 {
     return ts->state_ptr;
@@ -1801,18 +1813,50 @@ static bool fold_sextract(OptContext *ctx, TCGOp *op)
 
 static bool fold_shift(OptContext *ctx, TCGOp *op)
 {
+    uint64_t s_mask, z_mask, sign;
+
     if (fold_const2(ctx, op) ||
         fold_ix_to_i(ctx, op, 0) ||
         fold_xi_to_x(ctx, op, 0)) {
         return true;
     }
 
+    s_mask = arg_info(op->args[1])->s_mask;
+    z_mask = arg_info(op->args[1])->z_mask;
+
     if (arg_is_const(op->args[2])) {
-        ctx->z_mask = do_constant_folding(op->opc, ctx->type,
-                                          arg_info(op->args[1])->z_mask,
-                                          arg_info(op->args[2])->val);
+        int sh = arg_info(op->args[2])->val;
+
+        ctx->z_mask = do_constant_folding(op->opc, ctx->type, z_mask, sh);
+
+        s_mask = do_constant_folding(op->opc, ctx->type, s_mask, sh);
+        ctx->s_mask = smask_from_smask(s_mask);
+
         return fold_masks(ctx, op);
     }
+
+    switch (op->opc) {
+    CASE_OP_32_64(sar):
+        /*
+         * Arithmetic right shift will not reduce the number of
+         * input sign repetitions.
+         */
+        ctx->s_mask = s_mask;
+        break;
+    CASE_OP_32_64(shr):
+        /*
+         * If the sign bit is known zero, then logical right shift
+         * will not reduced the number of input sign repetitions.
+         */
+        sign = (s_mask & -s_mask) >> 1;
+        if (!(z_mask & sign)) {
+            ctx->s_mask = s_mask;
+        }
+        break;
+    default:
+        break;
+    }
+
     return false;
 }
 
-- 
2.25.1