assembler/bdw: Add the support of align1 register-indirect addressing mode on Gen8

Otherwise it can't compile the following GPU shader that uses the
register-indirect addressing mode.
  >add.sat (16) r[a0.5,0]<1>:uw     r[a0.5,0]<16;16,1>:uw  0x0080:uw
  >add.sat (16) r[a0.5,32]<1>:uw    r[a0.5,32]<16;16,1>:uw 0x0080:uw

Signed-off-by: Zhao Yakui <yakui.zhao@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>

3 files changed, 101 insertions, 15 deletions

diff --git a/assembler/gen8_instruction.c b/assembler/gen8_instruction.c
index 7fb1eacd..31c15ca3 100644
--- a/assembler/gen8_instruction.c
+++ b/assembler/gen8_instruction.c
@@ -47,23 +47,38 @@ gen8_set_dst(struct gen8_instruction *inst, struct brw_reg reg)
    gen8_set_dst_reg_file(inst, reg.file);
    gen8_set_dst_reg_type(inst, reg.type);
 
-   assert(reg.address_mode == BRW_ADDRESS_DIRECT);
+   if (reg.address_mode == BRW_ADDRESS_DIRECT) {
+      gen8_set_dst_da_reg_nr(inst, reg.nr);
 
-   gen8_set_dst_da_reg_nr(inst, reg.nr);
+      if (gen8_access_mode(inst) == BRW_ALIGN_1) {
+         /* Set Dst.SubRegNum[4:0] */
+         gen8_set_dst_da1_subreg_nr(inst, reg.subnr);
 
-   if (gen8_access_mode(inst) == BRW_ALIGN_1) {
-      /* Set Dst.SubRegNum[4:0] */
-      gen8_set_dst_da1_subreg_nr(inst, reg.subnr);
+         /* Set Dst.HorzStride */
+         if (reg.hstride == BRW_HORIZONTAL_STRIDE_0)
+            reg.hstride = BRW_HORIZONTAL_STRIDE_1;
+         gen8_set_dst_da1_hstride(inst, reg.hstride);
+      } else {
+         /* Align16 SubRegNum only has a single bit (bit 4; bits 3:0 MBZ). */
+         assert(reg.subnr == 0 || reg.subnr == 16);
+         gen8_set_dst_da16_subreg_nr(inst, reg.subnr >> 4);
+         gen8_set_da16_writemask(inst, reg.dw1.bits.writemask);
+      }
+   } else {
+      /* Indirect mode */
+      assert (gen8_access_mode(inst) == BRW_ALIGN_1);
 
+      gen8_set_dst_addr_mode(inst, BRW_ADDRESS_REGISTER_INDIRECT_REGISTER);
       /* Set Dst.HorzStride */
       if (reg.hstride == BRW_HORIZONTAL_STRIDE_0)
          reg.hstride = BRW_HORIZONTAL_STRIDE_1;
       gen8_set_dst_da1_hstride(inst, reg.hstride);
-   } else {
-      /* Align16 SubRegNum only has a single bit (bit 4; bits 3:0 MBZ). */
-      assert(reg.subnr == 0 || reg.subnr == 16);
-      gen8_set_dst_da16_subreg_nr(inst, reg.subnr >> 4);
-      gen8_set_da16_writemask(inst, reg.dw1.bits.writemask);
+      gen8_set_dst_ida1_sub_nr(inst, reg.subnr);
+      gen8_set_dst_ida1_imm8(inst, (reg.dw1.bits.indirect_offset & IMM8_MASK));
+      if ((reg.dw1.bits.indirect_offset & IMM9_MASK) == IMM9_MASK)
+         gen8_set_dst_ida1_imm9(inst, 1);
+      else
+         gen8_set_dst_ida1_imm9(inst, 0);
    }
 
    /* Generators should set a default exec_size of either 8 (SIMD4x2 or SIMD8)
@@ -160,7 +175,6 @@ gen8_set_src0(struct gen8_instruction *inst, struct brw_reg reg)
    gen8_set_src0_abs(inst, reg.abs);
    gen8_set_src0_negate(inst, reg.negate);
 
-   assert(reg.address_mode == BRW_ADDRESS_DIRECT);
 
    if (reg.file == BRW_IMMEDIATE_VALUE) {
       inst->data[3] = reg.dw1.ud;
@@ -168,7 +182,7 @@ gen8_set_src0(struct gen8_instruction *inst, struct brw_reg reg)
       /* Required to set some fields in src1 as well: */
       gen8_set_src1_reg_file(inst, 0); /* arf */
       gen8_set_src1_reg_type(inst, reg.type);
-   } else {
+   } else if (reg.address_mode == BRW_ADDRESS_DIRECT) {
       gen8_set_src0_da_reg_nr(inst, reg.nr);
 
       if (gen8_access_mode(inst) == BRW_ALIGN_1) {
@@ -211,6 +225,25 @@ gen8_set_src0(struct gen8_instruction *inst, struct brw_reg reg)
          else
             gen8_set_src0_vert_stride(inst, reg.vstride);
       }
+   } else if (reg.address_mode == BRW_ADDRESS_REGISTER_INDIRECT_REGISTER) {
+      assert (gen8_access_mode(inst) == BRW_ALIGN_1);
+      if (reg.width == BRW_WIDTH_1 &&
+         gen8_exec_size(inst) == BRW_EXECUTE_1) {
+         gen8_set_src0_da1_hstride(inst, BRW_HORIZONTAL_STRIDE_0);
+         gen8_set_src0_vert_stride(inst, BRW_VERTICAL_STRIDE_0);
+      } else {
+         gen8_set_src0_da1_hstride(inst, reg.hstride);
+         gen8_set_src0_vert_stride(inst, reg.vstride);
+      }
+
+      gen8_set_src0_da1_width(inst, reg.width);
+      gen8_set_src0_ida1_sub_nr(inst, reg.subnr);
+      gen8_set_src0_addr_mode(inst, BRW_ADDRESS_REGISTER_INDIRECT_REGISTER);
+      gen8_set_src0_ida1_imm8(inst, (reg.dw1.bits.indirect_offset & IMM8_MASK));
+      if ((reg.dw1.bits.indirect_offset & IMM9_MASK) == IMM9_MASK)
+         gen8_set_src0_ida1_imm9(inst, 1);
+      else
+         gen8_set_src0_ida1_imm9(inst, 0);
    }
 }
 
@@ -236,11 +269,9 @@ gen8_set_src1(struct gen8_instruction *inst, struct brw_reg reg)
    /* Only src1 can be an immediate in two-argument instructions. */
    assert(gen8_src0_reg_file(inst) != BRW_IMMEDIATE_VALUE);
 
-   assert(reg.address_mode == BRW_ADDRESS_DIRECT);
-
    if (reg.file == BRW_IMMEDIATE_VALUE) {
       inst->data[3] = reg.dw1.ud;
-   } else {
+   } else if (reg.address_mode == BRW_ADDRESS_DIRECT) {
       gen8_set_src1_da_reg_nr(inst, reg.nr);
 
       if (gen8_access_mode(inst) == BRW_ALIGN_1) {
@@ -282,6 +313,25 @@ gen8_set_src1(struct gen8_instruction *inst, struct brw_reg reg)
          else
             gen8_set_src1_vert_stride(inst, reg.vstride);
       }
+   } else if (reg.address_mode == BRW_ADDRESS_REGISTER_INDIRECT_REGISTER) {
+      assert (gen8_access_mode(inst) == BRW_ALIGN_1);
+      if (reg.width == BRW_WIDTH_1 &&
+         gen8_exec_size(inst) == BRW_EXECUTE_1) {
+         gen8_set_src1_da1_hstride(inst, BRW_HORIZONTAL_STRIDE_0);
+         gen8_set_src1_vert_stride(inst, BRW_VERTICAL_STRIDE_0);
+      } else {
+         gen8_set_src1_da1_hstride(inst, reg.hstride);
+         gen8_set_src1_vert_stride(inst, reg.vstride);
+      }
+
+      gen8_set_src1_da1_width(inst, reg.width);
+      gen8_set_src1_ida1_sub_nr(inst, reg.subnr);
+      gen8_set_src1_addr_mode(inst, BRW_ADDRESS_REGISTER_INDIRECT_REGISTER);
+      gen8_set_src1_ida1_imm8(inst, (reg.dw1.bits.indirect_offset & IMM8_MASK));
+      if ((reg.dw1.bits.indirect_offset & IMM9_MASK) == IMM9_MASK)
+         gen8_set_src1_ida1_imm9(inst, 1);
+      else
+         gen8_set_src1_ida1_imm9(inst, 0);
    }
 }
 
diff --git a/assembler/gen8_instruction.h b/assembler/gen8_instruction.h
index 87fa3a92..5e72e457 100644
--- a/assembler/gen8_instruction.h
+++ b/assembler/gen8_instruction.h
@@ -224,8 +224,32 @@ F(cre_message_type,        110, 109)
 F(cre_binding_table_index, 103,  96)
 /** @} */
 
+/* Addr Mode */
+
+F(dst_addr_mode,	  63, 63)
+F(src0_addr_mode,	  79, 79)
+F(src1_addr_mode,	  111, 111)
+
+/* Indirect access mode for Align1. */
+F(dst_ida1_sub_nr,        60,  57)
+F(src0_ida1_sub_nr,       76,  73)
+F(src1_ida1_sub_nr,      108, 105)
+
+/* Imm[8:0] of Immediate addr offset under Indirect mode */
+F(dst_ida1_imm8,         56,  48)
+F(src0_ida1_imm8,        72,  64)
+F(src1_ida1_imm8,        104,  96)
+
+/* Imm Bit9 of Immediate addr offset under Indirect mode */
+F(dst_ida1_imm9,         47,  47)
+F(src0_ida1_imm9,        95,  95)
+F(src1_ida1_imm9,        121, 121)
+
 #undef F
 
+#define IMM8_MASK	0x1FF
+#define IMM9_MASK	0x200
+
 /**
 * Flow control instruction bits:
 *  @{
diff --git a/assembler/gram.y b/assembler/gram.y
index 5671a0f4..03f1be64 100644
--- a/assembler/gram.y
+++ b/assembler/gram.y
@@ -294,6 +294,12 @@ static bool validate_dst_reg(struct brw_program_instruction *insn, struct brw_re
 	return false;
     }
 
+    if (reg->address_mode == BRW_ADDRESS_REGISTER_INDIRECT_REGISTER &&
+	access_mode(insn) == BRW_ALIGN_16) {
+	fprintf(stderr, "error: indirect Dst addr mode in align16 instruction\n");
+	return false;
+    }
+
     return true;
 }
 
@@ -317,6 +323,12 @@ static bool validate_src_reg(struct brw_program_instruction *insn,
 	return false;
     }
 
+    if (reg.address_mode == BRW_ADDRESS_REGISTER_INDIRECT_REGISTER &&
+	access_mode(insn) == BRW_ALIGN_16) {
+	fprintf(stderr, "error: indirect Source addr mode in align16 instruction\n");
+	return false;
+    }
+
     assert(reg.hstride >= 0 && reg.hstride < ARRAY_SIZE(hstride_for_reg));
     hstride = hstride_for_reg[reg.hstride];