aco: more carefully apply constant offsets into scratch accesses

Death stranding does scratch_arr[80-idx]. This doesn't seem to work if we
try to combine the subtraction into the access.

fossil-db (navi21):
Totals from 52 (0.04% of 135636) affected shaders:
Instrs: 78560 -> 79036 (+0.61%)
CodeSize: 427940 -> 431188 (+0.76%)
Latency: 1313809 -> 1318142 (+0.33%)
InvThroughput: 292833 -> 293842 (+0.34%)
VClause: 2361 -> 2555 (+8.22%); split: -0.51%, +8.73%
Copies: 8767 -> 8746 (-0.24%); split: -0.35%, +0.11%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: 0e783d687a3 ("aco: use scratch_* for scratch load/store on GFX9+")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/7735
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20117>
(cherry picked from commit 381de3c809fce5427308c696bbd313360194eff4)

2 files changed, 20 insertions, 5 deletions

diff --git a/.pick_status.json b/.pick_status.json
index 454a214bff0..bf0a1988d32 100644
--- a/.pick_status.json
+++ b/.pick_status.json
@@ -3307,7 +3307,7 @@
         "description": "aco: more carefully apply constant offsets into scratch accesses",
         "nominated": true,
         "nomination_type": 1,
-        "resolution": 0,
+        "resolution": 1,
         "main_sha": null,
         "because_sha": "0e783d687a3e13284eeae4081f16ee22033ff89b"
     },
diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp
index adb959a64e9..68144868c8c 100644
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -1255,12 +1255,14 @@ is_op_canonicalized(opt_ctx& ctx, Operand op)
 }
 
 bool
-is_scratch_offset_valid(opt_ctx& ctx, Instruction* instr, int32_t offset)
+is_scratch_offset_valid(opt_ctx& ctx, Instruction* instr, int64_t offset0, int64_t offset1)
 {
    bool negative_unaligned_scratch_offset_bug = ctx.program->gfx_level == GFX10;
    int32_t min = ctx.program->dev.scratch_global_offset_min;
    int32_t max = ctx.program->dev.scratch_global_offset_max;
 
+   int64_t offset = offset0 + offset1;
+
    bool has_vgpr_offset = instr && !instr->operands[0].isUndefined();
    if (negative_unaligned_scratch_offset_bug && has_vgpr_offset && offset < 0 && offset % 4)
       return false;
@@ -1467,15 +1469,28 @@ label_instruction(opt_ctx& ctx, aco_ptr<Instruction>& instr)
          while (info.is_temp())
             info = ctx.info[info.temp.id()];
 
-         if (i <= 1 && parse_base_offset(ctx, instr.get(), i, &base, &offset, false) &&
+         /* The hardware probably does: 'scratch_base + u2u64(saddr) + i2i64(offset)'. This means
+          * we can't combine the addition if the unsigned addition overflows and offset is
+          * positive. In theory, there is also issues if
+          * 'ilt(offset, 0) && ige(saddr, 0) && ilt(saddr + offset, 0)', but that just
+          * replaces an already out-of-bounds access with a larger one since 'saddr + offset'
+          * would be larger than INT32_MAX.
+          */
+         if (i <= 1 && parse_base_offset(ctx, instr.get(), i, &base, &offset, true) &&
              base.regClass() == instr->operands[i].regClass() &&
-             is_scratch_offset_valid(ctx, instr.get(), scratch.offset + (int32_t)offset)) {
+             is_scratch_offset_valid(ctx, instr.get(), scratch.offset, (int32_t)offset)) {
+            instr->operands[i].setTemp(base);
+            scratch.offset += (int32_t)offset;
+            continue;
+         } else if (i <= 1 && parse_base_offset(ctx, instr.get(), i, &base, &offset, false) &&
+                    base.regClass() == instr->operands[i].regClass() && (int32_t)offset < 0 &&
+                    is_scratch_offset_valid(ctx, instr.get(), scratch.offset, (int32_t)offset)) {
             instr->operands[i].setTemp(base);
             scratch.offset += (int32_t)offset;
             continue;
          } else if (i <= 1 && info.is_constant_or_literal(32) &&
                     ctx.program->gfx_level >= GFX10_3 &&
-                    is_scratch_offset_valid(ctx, NULL, scratch.offset + (int32_t)info.val)) {
+                    is_scratch_offset_valid(ctx, NULL, scratch.offset, (int32_t)info.val)) {
             /* GFX10.3+ can disable both SADDR and ADDR. */
             instr->operands[i] = Operand(instr->operands[i].regClass());
             scratch.offset += (int32_t)info.val;