i965/fs: Add an exec_size field to fs_inst

   This will, eventually, allow us to manage execution sizes of
   instructions in a much more natural way from the fs_visitor level.

i965/fs: Explicitly set instruction execute size a couple of places

i965/blorp: Explicitly set instruction execute sizes

   Since blorp is all 16-wide and nothing isn't, in general, very careful
   about register width, we'll just set it all explicitly.

Signed-off-by: Jason Ekstrand <[email protected]>
Reviewed-by: Matt Turner <[email protected]>

5 files changed, 126 insertions, 32 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_blorp_blit_eu.cpp b/src/mesa/drivers/dri/i965/brw_blorp_blit_eu.cpp
index 82ece732d18..3afe0e71519 100644
--- a/src/mesa/drivers/dri/i965/brw_blorp_blit_eu.cpp
+++ b/src/mesa/drivers/dri/i965/brw_blorp_blit_eu.cpp
@@ -96,7 +96,7 @@ brw_blorp_eu_emitter::emit_render_target_write(const struct brw_reg &src0,
                                                unsigned msg_length,
                                                bool use_header)
 {
-   fs_inst *inst = new (mem_ctx) fs_inst(FS_OPCODE_BLORP_FB_WRITE);
+   fs_inst *inst = new (mem_ctx) fs_inst(FS_OPCODE_BLORP_FB_WRITE, 16);
 
    inst->src[0] = src0;
    inst->base_mrf = msg_reg_nr;
diff --git a/src/mesa/drivers/dri/i965/brw_blorp_blit_eu.h b/src/mesa/drivers/dri/i965/brw_blorp_blit_eu.h
index 0459a7e1bd5..cd50da4883b 100644
--- a/src/mesa/drivers/dri/i965/brw_blorp_blit_eu.h
+++ b/src/mesa/drivers/dri/i965/brw_blorp_blit_eu.h
@@ -65,7 +65,7 @@ protected:
    {
       emit_cmp(op, x, y);
 
-      fs_inst *mv = new (mem_ctx) fs_inst(BRW_OPCODE_MOV, dst, src);
+      fs_inst *mv = new (mem_ctx) fs_inst(BRW_OPCODE_MOV, 16, dst, src);
       mv->predicate = BRW_PREDICATE_NORMAL;
       insts.push_tail(mv);
    }
@@ -82,17 +82,17 @@ protected:
                         const struct brw_reg &src3)
    {
       insts.push_tail(
-         new (mem_ctx) fs_inst(BRW_OPCODE_LRP, dst, src1, src2, src3));
+         new (mem_ctx) fs_inst(BRW_OPCODE_LRP, 16, dst, src1, src2, src3));
    }
 
    inline void emit_mov(const struct brw_reg& dst, const struct brw_reg& src)
    {
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_MOV, dst, src));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_MOV, 16, dst, src));
    }
 
    inline void emit_mov_8(const struct brw_reg& dst, const struct brw_reg& src)
    {
-      fs_inst *mv = new (mem_ctx) fs_inst(BRW_OPCODE_MOV, dst, src);
+      fs_inst *mv = new (mem_ctx) fs_inst(BRW_OPCODE_MOV, 8, dst, src);
       mv->force_uncompressed = true;
       insts.push_tail(mv);
    }
@@ -101,21 +101,21 @@ protected:
                         const struct brw_reg& src1,
                         const struct brw_reg& src2)
    {
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_AND, dst, src1, src2));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_AND, 16, dst, src1, src2));
    }
 
    inline void emit_add(const struct brw_reg& dst,
                         const struct brw_reg& src1,
                         const struct brw_reg& src2)
    {
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_ADD, dst, src1, src2));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_ADD, 16, dst, src1, src2));
    }
 
    inline void emit_add_8(const struct brw_reg& dst,
                           const struct brw_reg& src1,
                           const struct brw_reg& src2)
    {
-      fs_inst *add = new (mem_ctx) fs_inst(BRW_OPCODE_ADD, dst, src1, src2);
+      fs_inst *add = new (mem_ctx) fs_inst(BRW_OPCODE_ADD, 8, dst, src1, src2);
       add->force_uncompressed = true;
       insts.push_tail(add);
    }
@@ -124,40 +124,40 @@ protected:
                         const struct brw_reg& src1,
                         const struct brw_reg& src2)
    {
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_MUL, dst, src1, src2));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_MUL, 16, dst, src1, src2));
    }
 
    inline void emit_shr(const struct brw_reg& dst,
                         const struct brw_reg& src1,
                         const struct brw_reg& src2)
    {
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_SHR, dst, src1, src2));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_SHR, 16, dst, src1, src2));
    }
 
    inline void emit_shl(const struct brw_reg& dst,
                         const struct brw_reg& src1,
                         const struct brw_reg& src2)
    {
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_SHL, dst, src1, src2));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_SHL, 16, dst, src1, src2));
    }
 
    inline void emit_or(const struct brw_reg& dst,
                        const struct brw_reg& src1,
                        const struct brw_reg& src2)
    {
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_OR, dst, src1, src2));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_OR, 16, dst, src1, src2));
    }
 
    inline void emit_frc(const struct brw_reg& dst,
                         const struct brw_reg& src)
    {
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_FRC, dst, src));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_FRC, 16, dst, src));
    }
 
    inline void emit_rndd(const struct brw_reg& dst,
                          const struct brw_reg& src)
    {
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_RNDD, dst, src));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_RNDD, 16, dst, src));
    }
 
    inline void emit_cmp_if(enum brw_conditional_mod op,
@@ -165,17 +165,17 @@ protected:
                            const struct brw_reg &y)
    {
       emit_cmp(op, x, y);
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_IF));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_IF, 16));
    }
 
    inline void emit_else(void)
    {
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_ELSE));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_ELSE, 16));
    }
 
    inline void emit_endif(void)
    {
-      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_ENDIF));
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_ENDIF, 16));
    }
 
 private:
diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
index 9f30b379e79..ad61b33b754 100644
--- a/src/mesa/drivers/dri/i965/brw_fs.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
@@ -53,7 +53,8 @@ extern "C" {
 #include "glsl/glsl_types.h"
 
 void
-fs_inst::init(enum opcode opcode, const fs_reg &dst, fs_reg *src, int sources)
+fs_inst::init(enum opcode opcode, uint8_t exec_size, const fs_reg &dst,
+              fs_reg *src, int sources)
 {
    memset(this, 0, sizeof(*this));
 
@@ -61,6 +62,33 @@ fs_inst::init(enum opcode opcode, const fs_reg &dst, fs_reg *src, int sources)
    this->dst = dst;
    this->src = src;
    this->sources = sources;
+   this->exec_size = exec_size;
+
+   assert(dst.file != IMM && dst.file != UNIFORM);
+
+   /* If exec_size == 0, try to guess it from the registers.  Since all
+    * manner of things may use hardware registers, we first try to guess
+    * based on GRF registers.  If this fails, we will go ahead and take the
+    * width from the destination register.
+    */
+   if (this->exec_size == 0) {
+      if (dst.file == GRF) {
+         this->exec_size = dst.width;
+      } else {
+         for (int i = 0; i < sources; ++i) {
+            if (src[i].file != GRF)
+               continue;
+
+            if (this->exec_size <= 1)
+               this->exec_size = src[i].width;
+            assert(src[i].width == 1 || src[i].width == this->exec_size);
+         }
+      }
+
+      if (this->exec_size == 0 && dst.file != BAD_FILE)
+         this->exec_size = dst.width;
+   }
+   assert(this->exec_size != 0);
 
    this->conditional_mod = BRW_CONDITIONAL_NONE;
 
@@ -84,17 +112,46 @@ fs_inst::init(enum opcode opcode, const fs_reg &dst, fs_reg *src, int sources)
    this->writes_accumulator = false;
 }
 
+fs_inst::fs_inst()
+{
+   fs_reg *src = ralloc_array(this, fs_reg, 3);
+   init(BRW_OPCODE_NOP, 8, dst, src, 0);
+}
+
+fs_inst::fs_inst(enum opcode opcode, uint8_t exec_size)
+{
+   fs_reg *src = ralloc_array(this, fs_reg, 3);
+   init(opcode, exec_size, reg_undef, src, 0);
+}
+
 fs_inst::fs_inst(enum opcode opcode, const fs_reg &dst)
 {
    fs_reg *src = ralloc_array(this, fs_reg, 3);
-   init(opcode, dst, src, 0);
+   init(opcode, 0, dst, src, 0);
+}
+
+fs_inst::fs_inst(enum opcode opcode, uint8_t exec_size, const fs_reg &dst,
+                 const fs_reg &src0)
+{
+   fs_reg *src = ralloc_array(this, fs_reg, 3);
+   src[0] = src0;
+   init(opcode, exec_size, dst, src, 1);
 }
 
 fs_inst::fs_inst(enum opcode opcode, const fs_reg &dst, const fs_reg &src0)
 {
    fs_reg *src = ralloc_array(this, fs_reg, 3);
    src[0] = src0;
-   init(opcode, dst, src, 1);
+   init(opcode, 0, dst, src, 1);
+}
+
+fs_inst::fs_inst(enum opcode opcode, uint8_t exec_size, const fs_reg &dst,
+                 const fs_reg &src0, const fs_reg &src1)
+{
+   fs_reg *src = ralloc_array(this, fs_reg, 3);
+   src[0] = src0;
+   src[1] = src1;
+   init(opcode, exec_size, dst, src, 2);
 }
 
 fs_inst::fs_inst(enum opcode opcode, const fs_reg &dst, const fs_reg &src0,
@@ -103,7 +160,17 @@ fs_inst::fs_inst(enum opcode opcode, const fs_reg &dst, const fs_reg &src0,
    fs_reg *src = ralloc_array(this, fs_reg, 3);
    src[0] = src0;
    src[1] = src1;
-   init(opcode, dst, src, 2);
+   init(opcode, 0, dst, src, 2);
+}
+
+fs_inst::fs_inst(enum opcode opcode, uint8_t exec_size, const fs_reg &dst,
+                 const fs_reg &src0, const fs_reg &src1, const fs_reg &src2)
+{
+   fs_reg *src = ralloc_array(this, fs_reg, 3);
+   src[0] = src0;
+   src[1] = src1;
+   src[2] = src2;
+   init(opcode, exec_size, dst, src, 3);
 }
 
 fs_inst::fs_inst(enum opcode opcode, const fs_reg &dst, const fs_reg &src0,
@@ -113,12 +180,18 @@ fs_inst::fs_inst(enum opcode opcode, const fs_reg &dst, const fs_reg &src0,
    src[0] = src0;
    src[1] = src1;
    src[2] = src2;
-   init(opcode, dst, src, 3);
+   init(opcode, 0, dst, src, 3);
 }
 
 fs_inst::fs_inst(enum opcode opcode, const fs_reg &dst, fs_reg src[], int sources)
 {
-   init(opcode, dst, src, sources);
+   init(opcode, 0, dst, src, sources);
+}
+
+fs_inst::fs_inst(enum opcode opcode, uint8_t exec_width, const fs_reg &dst,
+                 fs_reg src[], int sources)
+{
+   init(opcode, exec_width, dst, src, sources);
 }
 
 fs_inst::fs_inst(const fs_inst &that)
@@ -206,7 +279,7 @@ ALU2(MAC)
 fs_inst *
 fs_visitor::IF(enum brw_predicate predicate)
 {
-   fs_inst *inst = new(mem_ctx) fs_inst(BRW_OPCODE_IF);
+   fs_inst *inst = new(mem_ctx) fs_inst(BRW_OPCODE_IF, dispatch_width);
    inst->predicate = predicate;
    return inst;
 }
@@ -217,7 +290,7 @@ fs_visitor::IF(const fs_reg &src0, const fs_reg &src1,
                enum brw_conditional_mod condition)
 {
    assert(brw->gen == 6);
-   fs_inst *inst = new(mem_ctx) fs_inst(BRW_OPCODE_IF,
+   fs_inst *inst = new(mem_ctx) fs_inst(BRW_OPCODE_IF, dispatch_width,
                                         reg_null_d, src0, src1);
    inst->conditional_mod = condition;
    return inst;
@@ -358,6 +431,7 @@ fs_visitor::DEP_RESOLVE_MOV(int grf)
    /* The caller always wants uncompressed to emit the minimal extra
     * dependencies, and to avoid having to deal with aligning its regs to 2.
     */
+   inst->exec_size = 8;
    inst->force_uncompressed = true;
 
    return inst;
@@ -380,6 +454,7 @@ fs_inst::equals(fs_inst *inst) const
            eot == inst->eot &&
            header_present == inst->header_present &&
            shadow_compare == inst->shadow_compare &&
+           exec_size == inst->exec_size &&
            offset == inst->offset);
 }
 
@@ -605,6 +680,7 @@ fs_visitor::get_timestamp()
     */
    mov->force_writemask_all = true;
    mov->force_uncompressed = true;
+   mov->exec_size = 8;
 
    /* The caller wants the low 32 bits of the timestamp.  Since it's running
     * at the GPU clock rate of ~1.2ghz, it will roll over every ~3 seconds,
@@ -760,7 +836,7 @@ fs_visitor::no16(const char *format, ...)
 fs_inst *
 fs_visitor::emit(enum opcode opcode)
 {
-   return emit(new(mem_ctx) fs_inst(opcode));
+   return emit(new(mem_ctx) fs_inst(opcode, dispatch_width));
 }
 
 fs_inst *
@@ -2129,7 +2205,7 @@ fs_visitor::demote_pull_constants()
          } else {
             fs_reg offset = fs_reg((unsigned)(pull_index * 4) & ~15);
             fs_inst *pull =
-               new(mem_ctx) fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD,
+               new(mem_ctx) fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD, 8,
                                     dst, surf_index, offset);
             inst->insert_before(block, pull);
             inst->src[i].set_smear(pull_index & 3);
@@ -2840,7 +2916,7 @@ fs_visitor::lower_uniform_pull_constant_loads()
           * by live variable analysis, or register allocation will explode.
           */
          fs_inst *setup = new(mem_ctx) fs_inst(FS_OPCODE_SET_SIMD4X2_OFFSET,
-                                               payload, const_offset_reg);
+                                               8, payload, const_offset_reg);
          setup->force_writemask_all = true;
 
          setup->ir = inst->ir;
diff --git a/src/mesa/drivers/dri/i965/brw_fs.h b/src/mesa/drivers/dri/i965/brw_fs.h
index 05fb71d724c..0e398026977 100644
--- a/src/mesa/drivers/dri/i965/brw_fs.h
+++ b/src/mesa/drivers/dri/i965/brw_fs.h
@@ -193,18 +193,29 @@ public:
 class fs_inst : public backend_instruction {
    fs_inst &operator=(const fs_inst &);
 
+   void init(enum opcode opcode, uint8_t exec_width, const fs_reg &dst,
+             fs_reg *src, int sources);
+
 public:
    DECLARE_RALLOC_CXX_OPERATORS(fs_inst)
 
-   void init(enum opcode opcode, const fs_reg &dst, fs_reg *src, int sources);
-
-   fs_inst(enum opcode opcode = BRW_OPCODE_NOP, const fs_reg &dst = reg_undef);
+   fs_inst();
+   fs_inst(enum opcode opcode, uint8_t exec_size);
+   fs_inst(enum opcode opcode, const fs_reg &dst);
+   fs_inst(enum opcode opcode, uint8_t exec_size, const fs_reg &dst,
+           const fs_reg &src0);
    fs_inst(enum opcode opcode, const fs_reg &dst, const fs_reg &src0);
+   fs_inst(enum opcode opcode, uint8_t exec_size, const fs_reg &dst,
+           const fs_reg &src0, const fs_reg &src1);
    fs_inst(enum opcode opcode, const fs_reg &dst, const fs_reg &src0,
            const fs_reg &src1);
+   fs_inst(enum opcode opcode, uint8_t exec_size, const fs_reg &dst,
+           const fs_reg &src0, const fs_reg &src1, const fs_reg &src2);
    fs_inst(enum opcode opcode, const fs_reg &dst, const fs_reg &src0,
            const fs_reg &src1, const fs_reg &src2);
    fs_inst(enum opcode opcode, const fs_reg &dst, fs_reg src[], int sources);
+   fs_inst(enum opcode opcode, uint8_t exec_size, const fs_reg &dst,
+           fs_reg src[], int sources);
    fs_inst(const fs_inst &that);
 
    void resize_sources(uint8_t num_sources);
@@ -224,6 +235,13 @@ public:
 
    uint8_t sources; /**< Number of fs_reg sources. */
 
+   /**
+    * Execution size of the instruction.  This is used by the generator to
+    * generate the correct binary for the given fs_inst.  Current valid
+    * values are 1, 8, 16.
+    */
+   uint8_t exec_size;
+
    /* Chooses which flag subregister (f0.0 or f0.1) is used for conditional
     * mod and predication.
     */
diff --git a/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp b/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
index da6f00145be..640987ed333 100644
--- a/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
@@ -944,7 +944,7 @@ fs_visitor::visit(ir_expression *ir)
          packed_consts.type = result.type;
 
          fs_reg const_offset_reg = fs_reg(const_offset->value.u[0] & ~15);
-         emit(new(mem_ctx) fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD,
+         emit(new(mem_ctx) fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD, 8,
                                    packed_consts, surf_index, const_offset_reg));
 
          for (int i = 0; i < ir->type->vector_elements; i++) {