nir: Add a simple int64 lowering pass

The algorithms used by this pass, especially for division, are heavily
based on the work Ian Romanick did for the similar int64 lowering pass
in the GLSL compiler.

v2: Properly handle vectors

v3: Get rid of log2_denom stuff.  Since we're using bcsel, we do all the
    calculations anyway and this is just extra instructions.

v4:
 - Add back in the log2_denom stuff since it's needed for ensuring that
   the shifts don't overflow.
 - Rework the looping part of the pass to be easier to expand.

Reviewed-by: Matt Turner <[email protected]>

1 files changed, 279 insertions, 0 deletions

diff --git a/src/compiler/nir/nir_lower_int64.c b/src/compiler/nir/nir_lower_int64.c
new file mode 100644
index 00000000000..49140918100
--- /dev/null
+++ b/src/compiler/nir/nir_lower_int64.c
@@ -0,0 +1,279 @@
+/*
+ * Copyright © 2016 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "nir.h"
+#include "nir_builder.h"
+
+static nir_ssa_def *
+lower_umul64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
+{
+   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
+   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
+   nir_ssa_def *y_lo = nir_unpack_64_2x32_split_x(b, y);
+   nir_ssa_def *y_hi = nir_unpack_64_2x32_split_y(b, y);
+
+   nir_ssa_def *res_lo = nir_imul(b, x_lo, y_lo);
+   nir_ssa_def *res_hi = nir_iadd(b, nir_umul_high(b, x_lo, y_lo),
+                         nir_iadd(b, nir_imul(b, x_lo, y_hi),
+                                     nir_imul(b, x_hi, y_lo)));
+
+   return nir_pack_64_2x32_split(b, res_lo, res_hi);
+}
+
+static nir_ssa_def *
+lower_isign64(nir_builder *b, nir_ssa_def *x)
+{
+   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
+   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
+
+   nir_ssa_def *is_non_zero = nir_i2b(b, nir_ior(b, x_lo, x_hi));
+   nir_ssa_def *res_hi = nir_ishr(b, x_hi, nir_imm_int(b, 31));
+   nir_ssa_def *res_lo = nir_ior(b, res_hi, nir_b2i(b, is_non_zero));
+
+   return nir_pack_64_2x32_split(b, res_lo, res_hi);
+}
+
+static void
+lower_udiv64_mod64(nir_builder *b, nir_ssa_def *n, nir_ssa_def *d,
+                   nir_ssa_def **q, nir_ssa_def **r)
+{
+   /* TODO: We should specially handle the case where the denominator is a
+    * constant.  In that case, we should be able to reduce it to a multiply by
+    * a constant, some shifts, and an add.
+    */
+   nir_ssa_def *n_lo = nir_unpack_64_2x32_split_x(b, n);
+   nir_ssa_def *n_hi = nir_unpack_64_2x32_split_y(b, n);
+   nir_ssa_def *d_lo = nir_unpack_64_2x32_split_x(b, d);
+   nir_ssa_def *d_hi = nir_unpack_64_2x32_split_y(b, d);
+
+   nir_const_value v = { .u32 = { 0, 0, 0, 0 } };
+   nir_ssa_def *q_lo = nir_build_imm(b, n->num_components, 32, v);
+   nir_ssa_def *q_hi = nir_build_imm(b, n->num_components, 32, v);
+
+   nir_ssa_def *n_hi_before_if = n_hi;
+   nir_ssa_def *q_hi_before_if = q_hi;
+
+   /* If the upper 32 bits of denom are non-zero, it is impossible for shifts
+    * greater than 32 bits to occur.  If the upper 32 bits of the numerator
+    * are zero, it is impossible for (denom << [63, 32]) <= numer unless
+    * denom == 0.
+    */
+   nir_ssa_def *need_high_div =
+      nir_iand(b, nir_ieq(b, d_hi, nir_imm_int(b, 0)), nir_uge(b, n_hi, d_lo));
+   nir_push_if(b, nir_bany(b, need_high_div));
+   {
+      /* If we only have one component, then the bany above goes away and
+       * this is always true within the if statement.
+       */
+      if (n->num_components == 1)
+         need_high_div = nir_imm_int(b, NIR_TRUE);
+
+      nir_ssa_def *log2_d_lo = nir_ufind_msb(b, d_lo);
+
+      for (int i = 31; i >= 0; i--) {
+         /* if ((d.x << i) <= n.y) {
+          *    n.y -= d.x << i;
+          *    quot.y |= 1U << i;
+          * }
+          */
+         nir_ssa_def *d_shift = nir_ishl(b, d_lo, nir_imm_int(b, i));
+         nir_ssa_def *new_n_hi = nir_isub(b, n_hi, d_shift);
+         nir_ssa_def *new_q_hi = nir_ior(b, q_hi, nir_imm_int(b, 1u << i));
+         nir_ssa_def *cond = nir_iand(b, need_high_div,
+                                         nir_uge(b, n_hi, d_shift));
+         if (i != 0) {
+            /* log2_d_lo is always <= 31, so we don't need to bother with it
+             * in the last iteration.
+             */
+            cond = nir_iand(b, cond,
+                               nir_ige(b, nir_imm_int(b, 31 - i), log2_d_lo));
+         }
+         n_hi = nir_bcsel(b, cond, new_n_hi, n_hi);
+         q_hi = nir_bcsel(b, cond, new_q_hi, q_hi);
+      }
+   }
+   nir_pop_if(b, NULL);
+   n_hi = nir_if_phi(b, n_hi, n_hi_before_if);
+   q_hi = nir_if_phi(b, q_hi, q_hi_before_if);
+
+   nir_ssa_def *log2_denom = nir_ufind_msb(b, d_hi);
+
+   n = nir_pack_64_2x32_split(b, n_lo, n_hi);
+   d = nir_pack_64_2x32_split(b, d_lo, d_hi);
+   for (int i = 31; i >= 0; i--) {
+      /* if ((d64 << i) <= n64) {
+       *    n64 -= d64 << i;
+       *    quot.x |= 1U << i;
+       * }
+       */
+      nir_ssa_def *d_shift = nir_ishl(b, d, nir_imm_int(b, i));
+      nir_ssa_def *new_n = nir_isub(b, n, d_shift);
+      nir_ssa_def *new_q_lo = nir_ior(b, q_lo, nir_imm_int(b, 1u << i));
+      nir_ssa_def *cond = nir_uge(b, n, d_shift);
+      if (i != 0) {
+         /* log2_denom is always <= 31, so we don't need to bother with it
+          * in the last iteration.
+          */
+         cond = nir_iand(b, cond,
+                            nir_ige(b, nir_imm_int(b, 31 - i), log2_denom));
+      }
+      n = nir_bcsel(b, cond, new_n, n);
+      q_lo = nir_bcsel(b, cond, new_q_lo, q_lo);
+   }
+
+   *q = nir_pack_64_2x32_split(b, q_lo, q_hi);
+   *r = n;
+}
+
+static nir_ssa_def *
+lower_udiv64(nir_builder *b, nir_ssa_def *n, nir_ssa_def *d)
+{
+   nir_ssa_def *q, *r;
+   lower_udiv64_mod64(b, n, d, &q, &r);
+   return q;
+}
+
+static nir_ssa_def *
+lower_idiv64(nir_builder *b, nir_ssa_def *n, nir_ssa_def *d)
+{
+   nir_ssa_def *n_hi = nir_unpack_64_2x32_split_y(b, n);
+   nir_ssa_def *d_hi = nir_unpack_64_2x32_split_y(b, d);
+
+   nir_ssa_def *negate = nir_ine(b, nir_ilt(b, n_hi, nir_imm_int(b, 0)),
+                                    nir_ilt(b, d_hi, nir_imm_int(b, 0)));
+   nir_ssa_def *q, *r;
+   lower_udiv64_mod64(b, nir_iabs(b, n), nir_iabs(b, d), &q, &r);
+   return nir_bcsel(b, negate, nir_ineg(b, q), q);
+}
+
+static nir_ssa_def *
+lower_umod64(nir_builder *b, nir_ssa_def *n, nir_ssa_def *d)
+{
+   nir_ssa_def *q, *r;
+   lower_udiv64_mod64(b, n, d, &q, &r);
+   return r;
+}
+
+static nir_ssa_def *
+lower_imod64(nir_builder *b, nir_ssa_def *n, nir_ssa_def *d)
+{
+   nir_ssa_def *n_hi = nir_unpack_64_2x32_split_y(b, n);
+   nir_ssa_def *d_hi = nir_unpack_64_2x32_split_y(b, d);
+
+   nir_ssa_def *negate = nir_ine(b, nir_ilt(b, n_hi, nir_imm_int(b, 0)),
+                                    nir_ilt(b, d_hi, nir_imm_int(b, 0)));
+   nir_ssa_def *q, *r;
+   lower_udiv64_mod64(b, nir_iabs(b, n), nir_iabs(b, d), &q, &r);
+   return nir_bcsel(b, negate, nir_ineg(b, r), r);
+}
+
+static bool
+lower_int64_impl(nir_function_impl *impl, nir_lower_int64_options options)
+{
+   nir_builder b;
+   nir_builder_init(&b, impl);
+
+   bool progress = false;
+   nir_foreach_block(block, impl) {
+      nir_foreach_instr_safe(instr, block) {
+         if (instr->type != nir_instr_type_alu)
+            continue;
+
+         nir_alu_instr *alu = nir_instr_as_alu(instr);
+         assert(alu->dest.dest.is_ssa);
+         if (alu->dest.dest.ssa.bit_size != 64)
+            continue;
+
+         b.cursor = nir_before_instr(instr);
+
+         nir_ssa_def *new_def = NULL;
+         switch (alu->op) {
+         case nir_op_imul:
+            if (options & nir_lower_imul64) {
+               new_def = lower_umul64(&b, nir_ssa_for_alu_src(&b, alu, 0),
+                                          nir_ssa_for_alu_src(&b, alu, 1));
+            }
+            break;
+
+         case nir_op_isign:
+            if (options & nir_lower_isign64)
+               new_def = lower_isign64(&b, nir_ssa_for_alu_src(&b, alu, 0));
+            break;
+
+         case nir_op_udiv:
+            if (options & nir_lower_divmod64) {
+               new_def = lower_udiv64(&b, nir_ssa_for_alu_src(&b, alu, 0),
+                                          nir_ssa_for_alu_src(&b, alu, 1));
+            }
+            break;
+
+         case nir_op_idiv:
+            if (options & nir_lower_divmod64) {
+               new_def = lower_idiv64(&b, nir_ssa_for_alu_src(&b, alu, 0),
+                                          nir_ssa_for_alu_src(&b, alu, 1));
+            }
+            break;
+
+         case nir_op_umod:
+            if (options & nir_lower_divmod64) {
+               new_def = lower_umod64(&b, nir_ssa_for_alu_src(&b, alu, 0),
+                                          nir_ssa_for_alu_src(&b, alu, 1));
+            }
+            break;
+
+         case nir_op_imod:
+            if (options & nir_lower_divmod64) {
+               new_def = lower_imod64(&b, nir_ssa_for_alu_src(&b, alu, 0),
+                                          nir_ssa_for_alu_src(&b, alu, 1));
+            }
+            break;
+
+         default:
+            /* Not lowered */
+            continue;
+         }
+
+         if (new_def) {
+            nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa,
+                                     nir_src_for_ssa(new_def));
+            nir_instr_remove(&alu->instr);
+            progress = true;
+         }
+      }
+   }
+
+   return progress;
+}
+
+bool
+nir_lower_int64(nir_shader *shader, nir_lower_int64_options options)
+{
+   bool progress = false;
+
+   nir_foreach_function(function, shader) {
+      if (function->impl)
+         progress |= lower_int64_impl(function->impl, options);
+   }
+
+   return progress;
+}