i965/eu: Refactor jump distance scaling to use a helper function.

Different generations of hardware measure jump distances in different
units.  Previously, every function that needed to set a jump target open
coded this scaling, or made a hardcoded assumption (i.e. just used 2).

Most functions start with the number of instructions to jump, and scale
up to the hardware-specific value.  So, I made the function match that.

Others start with a byte offset, and divide by a constant (8) to obtain
the jump distance.  This is actually 16 / 2 (the jump scale for Gen5-7).

v2: Make the helper a static inline defined in brw_eu.h, instead of
    an actual function in brw_eu_emit.c (as suggested by Matt).

Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Chris Forbes <[email protected]>
Reviewed-by: Matt Turner <[email protected]>

3 files changed, 32 insertions, 17 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_eu.h b/src/mesa/drivers/dri/i965/brw_eu.h
index 7efc02816c9..3e03ab82cc0 100644
--- a/src/mesa/drivers/dri/i965/brw_eu.h
+++ b/src/mesa/drivers/dri/i965/brw_eu.h
@@ -316,6 +316,26 @@ void brw_shader_time_add(struct brw_compile *p,
                          struct brw_reg payload,
                          uint32_t surf_index);
 
+/**
+ * Return the generation-specific jump distance scaling factor.
+ *
+ * Given the number of instructions to jump, we need to scale by
+ * some number to obtain the actual jump distance to program in an
+ * instruction.
+ */
+static inline unsigned
+brw_jump_scale(const struct brw_context *brw)
+{
+   /* Ironlake and later measure jump targets in 64-bit data chunks (in order
+    * (to support compaction), so each 128-bit instruction requires 2 chunks.
+    */
+   if (brw->gen >= 5)
+      return 2;
+
+   /* Gen4 simply uses the number of 128-bit instructions. */
+   return 1;
+}
+
 /* If/else/endif.  Works by manipulating the execution flags on each
  * channel.
  */
diff --git a/src/mesa/drivers/dri/i965/brw_eu_emit.c b/src/mesa/drivers/dri/i965/brw_eu_emit.c
index 3ab44333ec8..36ebed2dec8 100644
--- a/src/mesa/drivers/dri/i965/brw_eu_emit.c
+++ b/src/mesa/drivers/dri/i965/brw_eu_emit.c
@@ -1333,12 +1333,7 @@ patch_IF_ELSE(struct brw_compile *p,
    assert(endif_inst != NULL);
    assert(else_inst == NULL || brw_inst_opcode(brw, else_inst) == BRW_OPCODE_ELSE);
 
-   unsigned br = 1;
-   /* Jump count is for 64bit data chunk each, so one 128bit instruction
-    * requires 2 chunks.
-    */
-   if (brw->gen >= 5)
-      br = 2;
+   unsigned br = brw_jump_scale(brw);
 
    assert(brw_inst_opcode(brw, endif_inst) == BRW_OPCODE_ENDIF);
    brw_inst_set_exec_size(brw, endif_inst, brw_inst_exec_size(brw, if_inst));
@@ -1635,7 +1630,7 @@ brw_patch_break_cont(struct brw_compile *p, brw_inst *while_inst)
    struct brw_context *brw = p->brw;
    brw_inst *do_inst = get_inner_do_insn(p);
    brw_inst *inst;
-   int br = (brw->gen == 5) ? 2 : 1;
+   unsigned br = brw_jump_scale(brw);
 
    assert(brw->gen < 6);
 
@@ -1659,10 +1654,7 @@ brw_WHILE(struct brw_compile *p)
 {
    struct brw_context *brw = p->brw;
    brw_inst *insn, *do_insn;
-   unsigned br = 1;
-
-   if (brw->gen >= 5)
-      br = 2;
+   unsigned br = brw_jump_scale(brw);
 
    if (brw->gen >= 7) {
       insn = next_insn(p, BRW_OPCODE_WHILE);
@@ -2346,7 +2338,7 @@ brw_find_loop_end(struct brw_compile *p, int start_offset)
 {
    struct brw_context *brw = p->brw;
    int offset;
-   int scale = 8;
+   int scale = 16 / brw_jump_scale(brw);
    void *store = p->store;
 
    assert(brw->gen >= 6);
@@ -2378,7 +2370,8 @@ brw_set_uip_jip(struct brw_compile *p)
 {
    struct brw_context *brw = p->brw;
    int offset;
-   int scale = 8;
+   int br = brw_jump_scale(brw);
+   int scale = 16 / br;
    void *store = p->store;
 
    if (brw->gen < 6)
@@ -2418,7 +2411,7 @@ brw_set_uip_jip(struct brw_compile *p)
 
       case BRW_OPCODE_ENDIF:
          if (block_end_offset == 0)
-            brw_inst_set_jip(brw, insn, 2);
+            brw_inst_set_jip(brw, insn, 1 * br);
          else
             brw_inst_set_jip(brw, insn, (block_end_offset - offset) / scale);
 	 break;
diff --git a/src/mesa/drivers/dri/i965/brw_fs_generator.cpp b/src/mesa/drivers/dri/i965/brw_fs_generator.cpp
index cec2e826e17..d3509a0f1de 100644
--- a/src/mesa/drivers/dri/i965/brw_fs_generator.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_generator.cpp
@@ -67,6 +67,8 @@ fs_generator::patch_discard_jumps_to_fb_writes()
    if (brw->gen < 6 || this->discard_halt_patches.is_empty())
       return false;
 
+   int scale = brw_jump_scale(brw);
+
    /* There is a somewhat strange undocumented requirement of using
     * HALT, according to the simulator.  If some channel has HALTed to
     * a particular UIP, then by the end of the program, every channel
@@ -79,8 +81,8 @@ fs_generator::patch_discard_jumps_to_fb_writes()
     * tests.
     */
    brw_inst *last_halt = gen6_HALT(p);
-   brw_inst_set_uip(brw, last_halt, 2);
-   brw_inst_set_jip(brw, last_halt, 2);
+   brw_inst_set_uip(brw, last_halt, 1 * scale);
+   brw_inst_set_jip(brw, last_halt, 1 * scale);
 
    int ip = p->nr_insn;
 
@@ -89,7 +91,7 @@ fs_generator::patch_discard_jumps_to_fb_writes()
 
       assert(brw_inst_opcode(brw, patch) == BRW_OPCODE_HALT);
       /* HALT takes a half-instruction distance from the pre-incremented IP. */
-      brw_inst_set_uip(brw, patch, (ip - patch_ip->ip) * 2);
+      brw_inst_set_uip(brw, patch, (ip - patch_ip->ip) * scale);
    }
 
    this->discard_halt_patches.make_empty();