agx: Add image_load opcode

This is equivalent to texture_load but cannot be reordered, since it might be
writeable.

It also sets bit 43. This needs more investigation, but it fixes
KHR-GLES31.core.shader_image_load_store.basic-glsl-misc-fs. Some sort of cache
control bit.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24258>

src/asahi/compiler/agx_lower_uniform_sources.c

@@ -22,6 +22,7 @@ should_lower(enum agx_opcode op, agx_index uniform, unsigned src_index)
    bool high = uniform.value >= 256;
 
    switch (op) {
+   case AGX_OPCODE_IMAGE_LOAD:
    case AGX_OPCODE_TEXTURE_LOAD:
    case AGX_OPCODE_TEXTURE_SAMPLE:
       return src_index != 1 && src_index != 2;

src/asahi/compiler/agx_opcodes.py

@@ -254,9 +254,10 @@ op("texture_sample",
       encoding_32 = (0x31, 0x7F, 8, 10), # XXX WRONG SIZE
       srcs = 6, imms = [DIM, LOD_MODE, MASK, SCOREBOARD, OFFSET, SHADOW,
 								GATHER])
-op("texture_load",
-      encoding_32 = (0x71, 0x7F, 8, 10), # XXX WRONG SIZE
-      srcs = 6, imms = [DIM, LOD_MODE, MASK, SCOREBOARD, OFFSET])
+for memory, can_reorder in [("texture", True), ("image", False)]:
+    op(f"{memory}_load", encoding_32 = (0x71, 0x7F, 8, 10), # XXX WRONG SIZE
+       srcs = 6, imms = [DIM, LOD_MODE, MASK, SCOREBOARD, OFFSET],
+       can_reorder = can_reorder)
 
 # sources are base, index
 op("device_load",

src/asahi/compiler/agx_optimizer.c

@@ -265,7 +265,7 @@ agx_optimizer_forward(agx_context *ctx)
       /* Inline immediates if we can. TODO: systematic */
       if (I->op != AGX_OPCODE_ST_VARY && I->op != AGX_OPCODE_COLLECT &&
           I->op != AGX_OPCODE_TEXTURE_SAMPLE &&
-          I->op != AGX_OPCODE_TEXTURE_LOAD &&
+          I->op != AGX_OPCODE_IMAGE_LOAD && I->op != AGX_OPCODE_TEXTURE_LOAD &&
           I->op != AGX_OPCODE_UNIFORM_STORE &&
           I->op != AGX_OPCODE_BLOCK_IMAGE_STORE)
          agx_optimizer_inline_imm(defs, I, info.nr_srcs, info.is_float);

src/asahi/compiler/agx_pack.c

@@ -774,6 +774,7 @@ agx_pack_instr(struct util_dynarray *emission, struct util_dynarray *fixups,
    }
 
    case AGX_OPCODE_TEXTURE_LOAD:
+   case AGX_OPCODE_IMAGE_LOAD:
    case AGX_OPCODE_TEXTURE_SAMPLE: {
       assert(I->mask != 0);
       assert(I->format <= 0x10);
@@ -795,6 +796,18 @@ agx_pack_instr(struct util_dynarray *emission, struct util_dynarray *fixups,
       unsigned q3 = 12;  // XXX
       unsigned kill = 0; // helper invocation kill bit
 
+      /* Set bit 43 for image loads. This seems to makes sure that image loads
+       * get the value written by the latest image store, not some other image
+       * store that was already in flight, fixing
+       *
+       *    KHR-GLES31.core.shader_image_load_store.basic-glsl-misc-fs
+       *
+       * Apple seems to set this bit unconditionally for read/write image loads
+       * and never for readonly image loads. Some sort of cache control.
+       */
+      if (I->op == AGX_OPCODE_IMAGE_LOAD)
+         q3 |= 1;
+
       uint32_t extend = ((U & BITFIELD_MASK(5)) << 0) | (kill << 5) |
                         ((I->dim >> 3) << 7) | ((R >> 6) << 8) |
                         ((C >> 6) << 10) | ((D >> 6) << 12) | ((T >> 6) << 14) |
@@ -804,7 +817,7 @@ agx_pack_instr(struct util_dynarray *emission, struct util_dynarray *fixups,
       bool L = (extend != 0);
 
       uint64_t raw =
-         0x31 | ((I->op == AGX_OPCODE_TEXTURE_LOAD) ? (1 << 6) : 0) |
+         0x31 | ((I->op != AGX_OPCODE_TEXTURE_SAMPLE) ? (1 << 6) : 0) |
          (Rt ? (1 << 8) : 0) | ((R & BITFIELD_MASK(6)) << 9) |
          (L ? (1 << 15) : 0) | ((C & BITFIELD_MASK(6)) << 16) |
          (Ct ? (1 << 22) : 0) | (q1 << 23) | ((D & BITFIELD_MASK(6)) << 24) |

src/asahi/compiler/agx_register_allocate.c

@@ -44,6 +44,7 @@ agx_write_registers(const agx_instr *I, unsigned d)
       assert(1 <= I->channels && I->channels <= 4);
       return I->channels * size;
 
+   case AGX_OPCODE_IMAGE_LOAD:
    case AGX_OPCODE_TEXTURE_LOAD:
    case AGX_OPCODE_TEXTURE_SAMPLE:
       /* Even when masked out, these clobber 4 registers */
@@ -238,6 +239,7 @@ agx_read_registers(const agx_instr *I, unsigned s)
       else
          return size;
 
+   case AGX_OPCODE_IMAGE_LOAD:
    case AGX_OPCODE_TEXTURE_LOAD:
    case AGX_OPCODE_TEXTURE_SAMPLE:
       if (s == 0) {