pan/midgard: Optimize comparisions with similar operations
[mesa.git] / src / panfrost / midgard / midgard_opt_invert.c
index ffe43a1b176f05139b12f2bb7c78e8f895c0704c..4527102553b02d95a260027cfbbac540b9e0f3de 100644 (file)
@@ -40,26 +40,22 @@ midgard_lower_invert(compiler_context *ctx, midgard_block *block)
                 midgard_instruction not = {
                         .type = TAG_ALU_4,
                         .mask = ins->mask,
-                        .ssa_args = {
-                                .src0 = temp,
-                                .src1 = -1,
-                                .dest = ins->ssa_args.dest,
-                                .inline_constant = true
-                        },
+                        .src = { temp, ~0, ~0, ~0 },
+                        .swizzle = SWIZZLE_IDENTITY,
+                        .dest = ins->dest,
+                        .has_inline_constant = true,
                         .alu = {
                                 .op = midgard_alu_op_inor,
                                 /* TODO: i16 */
                                 .reg_mode = midgard_reg_mode_32,
                                 .dest_override = midgard_dest_override_none,
-                                .outmod = midgard_outmod_int_wrap,
-                                .src1 = vector_alu_srco_unsigned(blank_alu_src),
-                                .src2 = vector_alu_srco_unsigned(zero_alu_src)
+                                .outmod = midgard_outmod_int_wrap
                         },
                 };
 
-                ins->ssa_args.dest = temp;
+                ins->dest = temp;
                 ins->invert = false;
-                mir_insert_instruction_before(mir_next_op(ins), not);
+                mir_insert_instruction_before(ctx, mir_next_op(ins), not);
         }
 }
 
@@ -75,15 +71,15 @@ midgard_opt_not_propagate(compiler_context *ctx, midgard_block *block)
                 if (ins->alu.op != midgard_alu_op_imov) continue;
                 if (!ins->invert) continue;
                 if (mir_nontrivial_source2_mod_simple(ins)) continue;
-                if (ins->ssa_args.src1 & IS_REG) continue;
+                if (ins->src[1] & IS_REG) continue;
 
                 /* Is it beneficial to propagate? */
-                if (!mir_single_use(ctx, ins->ssa_args.src1)) continue;
+                if (!mir_single_use(ctx, ins->src[1])) continue;
 
                 /* We found an imov.not, propagate the invert back */
 
                 mir_foreach_instr_in_block_from_rev(block, v, mir_prev_op(ins)) {
-                        if (v->ssa_args.dest != ins->ssa_args.src1) continue;
+                        if (v->dest != ins->src[1]) continue;
                         if (v->type != TAG_ALU_4) break;
 
                         v->invert = !v->invert;
@@ -132,6 +128,32 @@ mir_invert_op(midgard_alu_op op)
         }
 }
 
+static midgard_alu_op
+mir_demorgan_op(midgard_alu_op op)
+{
+        switch (op) {
+        case midgard_alu_op_iand:
+                return midgard_alu_op_inor;
+        case midgard_alu_op_ior:
+                return midgard_alu_op_inand;
+        default:
+                unreachable("Op not De Morgan-able");
+        }
+}
+
+static midgard_alu_op
+mir_notright_op(midgard_alu_op op)
+{
+        switch (op) {
+        case midgard_alu_op_iand:
+                return midgard_alu_op_iandnot;
+        case midgard_alu_op_ior:
+                return midgard_alu_op_iornot;
+        default:
+                unreachable("Op not right able");
+        }
+}
+
 bool
 midgard_opt_fuse_dest_invert(compiler_context *ctx, midgard_block *block)
 {
@@ -150,3 +172,206 @@ midgard_opt_fuse_dest_invert(compiler_context *ctx, midgard_block *block)
 
         return progress;
 }
+
+/* Next up, we can fuse inverts into the sources of bitwise ops:
+ *
+ * ~a & b = b & ~a = iandnot(b, a)
+ * a & ~b = iandnot(a, b)
+ * ~a & ~b = ~(a | b) = inor(a, b)
+ *
+ * ~a | b = b | ~a = iornot(b, a)
+ *  a | ~b = iornot(a, b)
+ * ~a | ~b = ~(a & b) = inand(a, b)
+ *
+ * ~a ^ b = ~(a ^ b) = inxor(a, b)
+ * a ^ ~b = ~(a ^ b) + inxor(a, b)
+ * ~a ^ ~b = a ^ b
+ * ~(a ^ b) = inxor(a, b)
+ */
+
+static bool
+mir_strip_inverted(compiler_context *ctx, unsigned node)
+{
+        if (node >= SSA_FIXED_MINIMUM)
+                return false;
+
+       /* Strips and returns the invert off a node */
+       mir_foreach_instr_global(ctx, ins) {
+               if (ins->compact_branch) continue;
+               if (ins->dest != node) continue;
+
+               bool status = ins->invert;
+               ins->invert = false;
+               return status;
+       }
+
+       unreachable("Invalid node stripped");
+}
+
+static bool
+is_ssa_or_constant(unsigned node)
+{
+        return !(node & IS_REG) || (node == SSA_FIXED_REGISTER(26));
+}
+
+bool
+midgard_opt_fuse_src_invert(compiler_context *ctx, midgard_block *block)
+{
+        bool progress = false;
+
+        mir_foreach_instr_in_block_safe(block, ins) {
+                /* Search for inverted bitwise */
+                if (ins->type != TAG_ALU_4) continue;
+                if (!mir_is_bitwise(ins)) continue;
+                if (ins->invert) continue;
+
+                if (!is_ssa_or_constant(ins->src[0])) continue;
+                if (!is_ssa_or_constant(ins->src[1])) continue;
+                if (!mir_single_use(ctx, ins->src[0])) continue;
+                if (!ins->has_inline_constant && !mir_single_use(ctx, ins->src[1])) continue;
+
+                bool not_a = mir_strip_inverted(ctx, ins->src[0]);
+                bool not_b =
+                        ins->has_inline_constant ? false :
+                        mir_strip_inverted(ctx, ins->src[1]);
+
+                /* Edge case: if src0 == src1, it'll've been stripped */
+                if ((ins->src[0] == ins->src[1]) && !ins->has_inline_constant)
+                        not_b = not_a;
+
+                progress |= (not_a || not_b);
+
+                /* No point */
+                if (!(not_a || not_b)) continue;
+
+                bool both = not_a && not_b;
+                bool left = not_a && !not_b;
+                bool right = !not_a && not_b;
+
+                /* No-op, but we got to strip the inverts */
+                if (both && ins->alu.op == midgard_alu_op_ixor)
+                        continue;
+
+                if (both) {
+                        ins->alu.op = mir_demorgan_op(ins->alu.op);
+                } else if (right || (left && !ins->has_inline_constant)) {
+                        /* Commute arguments */
+                        if (left)
+                                mir_flip(ins);
+
+                        ins->alu.op = mir_notright_op(ins->alu.op);
+                } else if (left && ins->has_inline_constant) {
+                        /* Some special transformations:
+                         *
+                         * ~A & c = ~(~(~A) | (~c)) = ~(A | ~c) = inor(A, ~c)
+                         * ~A | c = ~(~(~A) & (~c)) = ~(A & ~c) = inand(A, ~c)
+                         */
+
+                        ins->alu.op = mir_demorgan_op(ins->alu.op);
+                        ins->inline_constant = ~ins->inline_constant;
+                }
+        }
+
+        return progress;
+}
+
+/* Optimizes a .not away when used as the source of a conditional select:
+ *
+ * csel(a, b, c)  = { b if a, c if !a }
+ * csel(!a, b, c) = { b if !a, c if !(!a) } = { c if a, b if !a } = csel(a, c, b)
+ * csel(!a, b, c) = csel(a, c, b)
+ */
+
+bool
+midgard_opt_csel_invert(compiler_context *ctx, midgard_block *block)
+{
+        bool progress = false;
+
+        mir_foreach_instr_in_block_safe(block, ins) {
+                if (ins->type != TAG_ALU_4) continue;
+                if (!OP_IS_CSEL(ins->alu.op)) continue;
+                if (!mir_single_use(ctx, ins->src[2])) continue;
+                if (!mir_strip_inverted(ctx, ins->src[2])) continue;
+
+                mir_flip(ins);
+                progress |= true;
+        }
+
+        return progress;
+}
+
+
+static bool
+mir_is_inverted(compiler_context *ctx, unsigned node)
+{
+        mir_foreach_instr_global(ctx, ins) {
+                if (ins->compact_branch) continue;
+                if (ins->dest != node) continue;
+
+                return ins->invert;
+        }
+
+        unreachable("Invalid node passed");
+}
+
+
+
+/* Optimizes comparisions which invert both arguments
+ *
+ *
+ * ieq(not(a), not(b)) = ieq(a, b)
+ * ine(not(a), not(b)) = ine(a, b)
+ *
+ * This does apply for ilt and ile if we flip the argument order:
+ * Proofs below provided by Alyssa Rosenzweig
+ *
+ * not(x) = −(x+1)
+ *
+ * ( not(A) <= not(B) ) <=> ( −(A+1) <= −(B+1) )
+ *                      <=> ( A+1 >= B+1)
+ *                      <=> ( B <= A )
+ *
+ * On unsigned comparisons (ult / ule) we can perform the same optimization
+ * with the additional restriction that the source registers must
+ * have the same size.
+ *
+ * TODO: We may not need them to be of the same size, if we can
+ *       prove that they are the same after sext/zext
+ *
+ * not(x) = 2n−x−1
+ *
+ * ( not(A) <= not(B) ) <=> ( 2n−A−1 <= 2n−B−1 )
+ *                      <=> ( −A <= −B )
+ *                      <=> ( B <= A )
+ */
+bool
+midgard_opt_drop_cmp_invert(compiler_context *ctx, midgard_block *block)
+{
+
+        bool progress = false;
+
+        mir_foreach_instr_in_block_safe(block, ins) {
+                if (ins->type != TAG_ALU_4) continue;
+                if (!OP_IS_INTEGER_CMP(ins->alu.op)) continue;
+
+                if ((ins->src[0] & IS_REG) || (ins->src[1] & IS_REG)) continue;
+                if (!mir_single_use(ctx, ins->src[0]) || !mir_single_use(ctx, ins->src[1])) continue;
+
+                bool a_inverted = mir_is_inverted(ctx, ins->src[0]);
+                bool b_inverted = mir_is_inverted(ctx, ins->src[1]);
+
+                if (!a_inverted || !b_inverted) continue;
+                if (OP_IS_UNSIGNED_CMP(ins->alu.op) && mir_srcsize(ins, 0) != mir_srcsize(ins, 1)) continue;
+
+
+                mir_strip_inverted(ctx, ins->src[0]);
+                mir_strip_inverted(ctx, ins->src[1]);
+
+                if (ins->alu.op != midgard_alu_op_ieq && ins->alu.op != midgard_alu_op_ine)
+                        mir_flip(ins);
+
+                progress |= true;
+        }
+
+        return progress;
+}