Mesa (master): vc4: Add register allocation support for MUL output rotation.

[Eric Anholt]

Module: Mesa
Branch: master
Commit: e763e19808a84ae0218117c89864ff50cb6b0d16
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=e763e19808a84ae0218117c89864ff50cb6b0d16

Author: Eric Anholt <eric at anholt.net>
Date:   Thu Aug 25 14:32:47 2016 -0700

vc4: Add register allocation support for MUL output rotation.

We need the source to be in r0-r3, so make a new register class for it.
It will be up to the surrounding passes to make sure that the r0-r3
allocation of its source won't conflict with anything other class
requirements on that temp.

---

 src/gallium/drivers/vc4/vc4_context.h           |  1 +
 src/gallium/drivers/vc4/vc4_register_allocate.c | 13 +++++++++++++
 2 files changed, 14 insertions(+)

diff --git a/src/gallium/drivers/vc4/vc4_context.h b/src/gallium/drivers/vc4/vc4_context.h
index c3474a0..63a1dfb 100644
--- a/src/gallium/drivers/vc4/vc4_context.h
+++ b/src/gallium/drivers/vc4/vc4_context.h
@@ -283,6 +283,7 @@ struct vc4_context {
         struct ra_regs *regs;
         unsigned int reg_class_any;
         unsigned int reg_class_a_or_b_or_acc;
+        unsigned int reg_class_r0_r3;
         unsigned int reg_class_r4_or_a;
         unsigned int reg_class_a;
 
diff --git a/src/gallium/drivers/vc4/vc4_register_allocate.c b/src/gallium/drivers/vc4/vc4_register_allocate.c
index 203b459..fc44764 100644
--- a/src/gallium/drivers/vc4/vc4_register_allocate.c
+++ b/src/gallium/drivers/vc4/vc4_register_allocate.c
@@ -119,6 +119,7 @@ vc4_alloc_reg_set(struct vc4_context *vc4)
         vc4->reg_class_a_or_b_or_acc = ra_alloc_reg_class(vc4->regs);
         vc4->reg_class_r4_or_a = ra_alloc_reg_class(vc4->regs);
         vc4->reg_class_a = ra_alloc_reg_class(vc4->regs);
+        vc4->reg_class_r0_r3 = ra_alloc_reg_class(vc4->regs);
         for (uint32_t i = 0; i < ARRAY_SIZE(vc4_regs); i++) {
                 /* Reserve ra31/rb31 for spilling fixup_raddr_conflict() in
                  * vc4_qpu_emit.c
@@ -135,6 +136,9 @@ vc4_alloc_reg_set(struct vc4_context *vc4)
                         continue;
                 }
 
+                if (vc4_regs[i].mux <= QPU_MUX_R3)
+                        ra_class_add_reg(vc4->regs, vc4->reg_class_r0_r3, i);
+
                 ra_class_add_reg(vc4->regs, vc4->reg_class_any, i);
                 ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b_or_acc, i);
         }
@@ -164,6 +168,7 @@ node_to_temp_priority(const void *in_a, const void *in_b)
 #define CLASS_BIT_A			(1 << 0)
 #define CLASS_BIT_B_OR_ACC		(1 << 1)
 #define CLASS_BIT_R4			(1 << 2)
+#define CLASS_BIT_R0_R3			(1 << 4)
 
 /**
  * Returns a mapping from QFILE_TEMP indices to struct qpu_regs.
@@ -240,6 +245,11 @@ vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
                                         AB_INDEX + QPU_R_FRAG_PAYLOAD_ZW * 2);
                         break;
 
+                case QOP_ROT_MUL:
+                        assert(inst->src[0].file == QFILE_TEMP);
+                        class_bits[inst->src[0].index] &= ~CLASS_BIT_R0_R3;
+                        break;
+
                 default:
                         break;
                 }
@@ -287,6 +297,9 @@ vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
                 case CLASS_BIT_A:
                         ra_set_node_class(g, node, vc4->reg_class_a);
                         break;
+                case CLASS_BIT_R0_R3:
+                        ra_set_node_class(g, node, vc4->reg_class_r0_r3);
+                        break;
                 default:
                         fprintf(stderr, "temp %d: bad class bits: 0x%x\n",
                                 i, class_bits[i]);