ir3/lower_io_offsets: Try propagate SSBO's SHR into a previous shift instruction
authorEduardo Lima Mitev <elima@igalia.com>
Thu, 28 Feb 2019 17:17:50 +0000 (18:17 +0100)
committerEduardo Lima Mitev <elima@igalia.com>
Wed, 13 Mar 2019 20:19:44 +0000 (21:19 +0100)
While we lack value range tracking, this patch tries to 'manually' propogate
the division by 4 to calculate SSBO element-offset, into a possible previous
shift operation (shift left or right); checking that it is safe to do so.

This should help in cases like ie. when accessing a field in an array of
structs, where the offset is likely defined as base plus a multiplication
by a struct or array element size.

See dEQP test 'dEQP-GLES31.functional.ssbo.atomic.xor.highp_uint'
for an example of a shader that benefits from this.

Reviewed-by: Rob Clark <robdclark@gmail.com>
src/freedreno/ir3/ir3_nir_lower_io_offsets.c

index 264f106454518628bacf883a79c8726a3f820be9..bc50fa09f3bf2993b132ffa7df525602c410ab45 100644 (file)
@@ -83,6 +83,81 @@ get_ir3_intrinsic_for_ssbo_intrinsic(unsigned intrinsic,
        return -1;
 }
 
+static nir_ssa_def *
+check_and_propagate_bit_shift32(nir_builder *b, nir_ssa_def *offset,
+                                                               nir_alu_instr *alu_instr, int32_t direction,
+                                                               int32_t shift)
+{
+       debug_assert(alu_instr->src[1].src.is_ssa);
+       nir_ssa_def *shift_ssa = alu_instr->src[1].src.ssa;
+
+       /* Only propagate if the shift is a const value so we can check value range
+        * statically.
+        */
+       nir_const_value *const_val = nir_src_as_const_value(alu_instr->src[1].src);
+       if (!const_val)
+               return NULL;
+
+       int32_t current_shift = const_val->i32[0] * direction;
+       int32_t new_shift = current_shift + shift;
+
+       /* If the merge would reverse the direction, bail out.
+        * e.g, 'x << 2' then 'x >> 4' is not 'x >> 2'.
+        */
+       if (current_shift * new_shift < 0)
+               return NULL;
+
+       /* If the propagation would overflow an int32_t, bail out too to be on the
+        * safe side.
+        */
+       if (new_shift < -31 || new_shift > 31)
+               return NULL;
+
+       b->cursor = nir_before_instr(&alu_instr->instr);
+
+       /* Add or substract shift depending on the final direction (SHR vs. SHL). */
+       if (shift * direction < 0)
+               shift_ssa = nir_isub(b, shift_ssa, nir_imm_int(b, abs(shift)));
+       else
+               shift_ssa = nir_iadd(b, shift_ssa, nir_imm_int(b, abs(shift)));
+
+       return shift_ssa;
+}
+
+static nir_ssa_def *
+try_propagate_bit_shift(nir_builder *b, nir_ssa_def *offset, int32_t shift)
+{
+       nir_instr *offset_instr = offset->parent_instr;
+       if (offset_instr->type != nir_instr_type_alu)
+               return NULL;
+
+       nir_alu_instr *alu = nir_instr_as_alu(offset_instr);
+       nir_ssa_def *shift_ssa;
+       nir_ssa_def *new_offset = NULL;
+
+       switch (alu->op) {
+       case nir_op_ishl:
+               shift_ssa = check_and_propagate_bit_shift32(b, offset, alu, 1, shift);
+               if (shift_ssa)
+                       new_offset = nir_ishl(b, alu->src[0].src.ssa, shift_ssa);
+               break;
+       case nir_op_ishr:
+               shift_ssa = check_and_propagate_bit_shift32(b, offset, alu, -1, shift);
+               if (shift_ssa)
+                       new_offset = nir_ishr(b, alu->src[0].src.ssa, shift_ssa);
+               break;
+       case nir_op_ushr:
+               shift_ssa = check_and_propagate_bit_shift32(b, offset, alu, -1, shift);
+               if (shift_ssa)
+                       new_offset = nir_ushr(b, alu->src[0].src.ssa, shift_ssa);
+               break;
+       default:
+               return NULL;
+       }
+
+       return new_offset;
+}
+
 static bool
 lower_offset_for_ssbo(nir_intrinsic_instr *intrinsic, nir_builder *b,
                                          unsigned ir3_ssbo_opcode, uint8_t offset_src_idx)
@@ -104,6 +179,16 @@ lower_offset_for_ssbo(nir_intrinsic_instr *intrinsic, nir_builder *b,
        debug_assert(intrinsic->src[offset_src_idx].is_ssa);
        nir_ssa_def *offset = intrinsic->src[offset_src_idx].ssa;
 
+       /* Since we don't have value range checking, we first try to propagate
+        * the division by 4 ('offset >> 2') into another bit-shift instruction that
+        * possibly defines the offset. If that's the case, we emit a similar
+        * instructions adjusting (merging) the shift value.
+        *
+        * Here we use the convention that shifting right is negative while shifting
+        * left is positive. So 'x / 4' ~ 'x >> 2' or 'x << -2'.
+        */
+       nir_ssa_def *new_offset = try_propagate_bit_shift(b, offset, -2);
+
        /* The new source that will hold the dword-offset is always the last
         * one for every intrinsic.
         */
@@ -127,11 +212,16 @@ lower_offset_for_ssbo(nir_intrinsic_instr *intrinsic, nir_builder *b,
        new_intrinsic->num_components = intrinsic->num_components;
 
        b->cursor = nir_before_instr(&intrinsic->instr);
-       nir_ssa_def *offset_div_4 = nir_ushr(b, offset, nir_imm_int(b, 2));
-       debug_assert(offset_div_4);
+
+       /* If we managed to propagate the division by 4, just use the new offset
+        * register and don't emit the SHR.
+        */
+       if (new_offset)
+               offset = new_offset;
+       else
+               offset = nir_ushr(b, offset, nir_imm_int(b, 2));
 
        /* Insert the new intrinsic right before the old one. */
-       b->cursor = nir_before_instr(&intrinsic->instr);
        nir_builder_instr_insert(b, &new_intrinsic->instr);
 
        /* Replace the last source of the new intrinsic by the result of
@@ -139,7 +229,7 @@ lower_offset_for_ssbo(nir_intrinsic_instr *intrinsic, nir_builder *b,
         */
        nir_instr_rewrite_src(&new_intrinsic->instr,
                                                  target_src,
-                                                 nir_src_for_ssa(offset_div_4));
+                                                 nir_src_for_ssa(offset));
 
        if (has_dest) {
                /* Replace the uses of the original destination by that