diff --git a/src/asahi/compiler/agx_compiler.h b/src/asahi/compiler/agx_compiler.h
index 94f75a6e985..e4f76813d8e 100644
--- a/src/asahi/compiler/agx_compiler.h
+++ b/src/asahi/compiler/agx_compiler.h
@@ -43,6 +43,9 @@ enum agx_dbg {
 
 extern int agx_debug;
 
+/* r0-r127 inclusive, as pairs of 16-bits, gives 256 registers */
+#define AGX_NUM_REGS (256)
+
 enum agx_index_type {
    AGX_INDEX_NULL = 0,
    AGX_INDEX_NORMAL = 1,
@@ -325,6 +328,9 @@ typedef struct agx_block {
    BITSET_WORD *live_in;
    BITSET_WORD *live_out;
 
+   /* Register allocation */
+   BITSET_DECLARE(regs_out, AGX_NUM_REGS);
+
    /* Offset of the block in the emitted binary */
    off_t offset;
 
diff --git a/src/asahi/compiler/agx_register_allocate.c b/src/asahi/compiler/agx_register_allocate.c
index 0977d2be2d5..0f1acdc7c57 100644
--- a/src/asahi/compiler/agx_register_allocate.c
+++ b/src/asahi/compiler/agx_register_allocate.c
@@ -37,8 +37,6 @@ agx_read_registers(agx_instr *I, unsigned s)
    unsigned size = I->src[s].size == AGX_SIZE_32 ? 2 : 1;
 
    switch (I->op) {
-   case AGX_OPCODE_DEVICE_LOAD:
-      return 8;
    default:
       return size;
    }
@@ -58,43 +56,134 @@ agx_write_registers(agx_instr *I, unsigned d)
       return 8;
    case AGX_OPCODE_LD_VARY_FLAT:
       return 6;
+   case AGX_OPCODE_P_COMBINE:
+   {
+      unsigned components = 0;
+
+      for (unsigned i = 0; i < 4; ++i) {
+         if (!agx_is_null(I->src[i]))
+            components = i + 1;
+      }
+
+      return components * size;
+   }
    default:
       return size;
    }
 }
 
+static unsigned
+agx_assign_regs(BITSET_WORD *used_regs, unsigned count, unsigned align)
+{
+   for (unsigned reg = 0; reg < AGX_NUM_REGS; reg += align) {
+      bool conflict = false;
+
+      for (unsigned j = 0; j < count; ++j)
+         conflict |= BITSET_TEST(used_regs, reg + j);
+
+      if (!conflict) {
+         for (unsigned j = 0; j < count; ++j)
+            BITSET_SET(used_regs, reg + j);
+
+         return reg;
+      }
+   }
+
+   unreachable("Could not find a free register");
+}
+
+/** Assign registers to SSA values in a block. */
+
+static void
+agx_ra_assign_local(agx_block *block, uint8_t *ssa_to_reg)
+{
+   BITSET_DECLARE(used_regs, AGX_NUM_REGS) = { 0 };
+
+   agx_foreach_predecessor(block, pred) {
+      for (unsigned i = 0; i < BITSET_WORDS(AGX_NUM_REGS); ++i)
+         used_regs[i] |= pred->regs_out[i];
+   }
+
+   BITSET_SET(used_regs, 0); // control flow writes r0l
+   BITSET_SET(used_regs, 5*2); // TODO: precolouring, don't overwrite vertex ID
+   BITSET_SET(used_regs, (5*2 + 1));
+
+   agx_foreach_instr_in_block(block, I) {
+      /* First, free killed sources */
+      agx_foreach_src(I, s) {
+         if (I->src[s].type == AGX_INDEX_NORMAL && I->src[s].kill) {
+            unsigned reg = ssa_to_reg[I->src[s].value];
+            unsigned count = agx_read_registers(I, s);
+
+            for (unsigned i = 0; i < count; ++i)
+               BITSET_CLEAR(used_regs, reg + i);
+         }
+      }
+
+      /* Next, assign destinations. Always legal in SSA form. */
+      agx_foreach_dest(I, d) {
+         if (I->dest[d].type == AGX_INDEX_NORMAL) {
+            unsigned count = agx_write_registers(I, d);
+            unsigned align = (I->dest[d].size == AGX_SIZE_16) ? 1 : 2;
+            unsigned reg = agx_assign_regs(used_regs, count, align);
+
+            ssa_to_reg[I->dest[d].value] = reg;
+         }
+      }
+   }
+
+   STATIC_ASSERT(sizeof(block->regs_out) == sizeof(used_regs));
+   memcpy(block->regs_out, used_regs, sizeof(used_regs));
+}
+
 void
 agx_ra(agx_context *ctx)
 {
    unsigned *alloc = calloc(ctx->alloc, sizeof(unsigned));
-   unsigned usage = 6*2;
+
+   agx_compute_liveness(ctx);
+   uint8_t *ssa_to_reg = calloc(ctx->alloc, sizeof(uint8_t));
+   agx_foreach_block(ctx, block)
+      agx_ra_assign_local(block, ssa_to_reg);
+
+   /* TODO: Coalesce combines */
 
    agx_foreach_instr_global_safe(ctx, ins) {
       /* Lower away RA pseudo-instructions */
       if (ins->op == AGX_OPCODE_P_COMBINE) {
          /* TODO: Optimize out the moves! */
-         unsigned components = 0;
-
-         for (unsigned i = 0; i < 4; ++i) {
-            if (!agx_is_null(ins->src[i]))
-               components = i + 1;
-         }
-
-         unsigned size = ins->dest[0].size == AGX_SIZE_32 ? 2 : 1;
-         if (size == 2 && usage & 1) usage++;
-         unsigned base = usage;
          assert(ins->dest[0].type == AGX_INDEX_NORMAL);
-         alloc[ins->dest[0].value] = base;
-         usage += (components * size);
+         enum agx_size common_size = ins->dest[0].size;
+         unsigned base = ssa_to_reg[ins->dest[0].value];
+         unsigned size = common_size == AGX_SIZE_32 ? 2 : 1;
 
          /* Move the sources */
          agx_builder b = agx_init_builder(ctx, agx_after_instr(ins));
 
+         /* TODO: Eliminate the intermediate copy by handling parallel copies */
          for (unsigned i = 0; i < 4; ++i) {
             if (agx_is_null(ins->src[i])) continue;
-            assert(ins->src[0].type == AGX_INDEX_NORMAL);
-            agx_mov_to(&b, agx_register(base + (i * size), ins->dest[0].size),
-                  agx_register(alloc[ins->src[i].value], ins->src[0].size));
+            unsigned base = ins->src[i].value;
+            if (ins->src[i].type == AGX_INDEX_NORMAL)
+               base = ssa_to_reg[base];
+            else
+               assert(ins->src[i].type == AGX_INDEX_REGISTER);
+
+            assert(ins->src[i].size == common_size);
+
+            agx_mov_to(&b, agx_register(124*2 + (i * size), common_size),
+                  agx_register(base, common_size));
+         }
+
+         for (unsigned i = 0; i < 4; ++i) {
+            if (agx_is_null(ins->src[i])) continue;
+            agx_index src = ins->src[i];
+
+            if (src.type == AGX_INDEX_NORMAL)
+               src = agx_register(alloc[src.value], src.size);
+
+            agx_mov_to(&b, agx_register(base + (i * size), common_size),
+                  agx_register(124*2 + (i * size), common_size));
          }
 
          /* We've lowered away, delete the old */
@@ -102,34 +191,42 @@ agx_ra(agx_context *ctx)
          continue;
       } else if (ins->op == AGX_OPCODE_P_EXTRACT) {
          assert(ins->dest[0].type == AGX_INDEX_NORMAL);
-         assert(ins->src[0].type == AGX_INDEX_NORMAL);
          assert(ins->dest[0].size == ins->src[0].size);
+         unsigned base = ins->src[0].value;
+
+         if (ins->src[0].type != AGX_INDEX_REGISTER) {
+            assert(ins->src[0].type == AGX_INDEX_NORMAL);
+            base = alloc[base];
+         }
 
          unsigned size = ins->dest[0].size == AGX_SIZE_32 ? 2 : 1;
-         alloc[ins->dest[0].value] = alloc[ins->src[0].value] + (size * ins->imm);
+         unsigned left = ssa_to_reg[ins->dest[0].value];
+         unsigned right = ssa_to_reg[ins->src[0].value] + (size * ins->imm);
+
+         if (left != right) {
+            agx_builder b = agx_init_builder(ctx, agx_after_instr(ins));
+            agx_mov_to(&b, agx_register(left, ins->dest[0].size),
+                  agx_register(right, ins->src[0].size));
+         }
+
          agx_remove_instruction(ins);
          continue;
       }
 
       agx_foreach_src(ins, s) {
          if (ins->src[s].type == AGX_INDEX_NORMAL) {
-            unsigned v = alloc[ins->src[s].value];
+            unsigned v = ssa_to_reg[ins->src[s].value];
             ins->src[s] = agx_replace_index(ins->src[s], agx_register(v, ins->src[s].size));
          }
       }
 
       agx_foreach_dest(ins, d) {
          if (ins->dest[d].type == AGX_INDEX_NORMAL) {
-            unsigned size = ins->dest[d].size == AGX_SIZE_32 ? 2 : 1;
-            if (size == 2 && usage & 1) usage++;
-            unsigned v = usage;
-            usage += agx_write_registers(ins, d);
-            alloc[ins->dest[d].value] = v;
+            unsigned v = ssa_to_reg[ins->dest[d].value];
             ins->dest[d] = agx_replace_index(ins->dest[d], agx_register(v, ins->dest[d].size));
          }
       }
    }
 
-   assert(usage < 256 && "dummy RA");
    free(alloc);
 }