pan/bi: Don't optimize if without else

We need to emit a jump over the else body when constructing the IR, even if the
else body appears empty in NIR, because instructions could be added when
going out-of-SSA. This optimization should instead happen post-RA, but doing so
is nontrivial and the cycle count regression is minimal, so punting on this for
now.

total instructions in shared programs: 2728364 -> 2731495 (0.11%)
instructions in affected programs: 580462 -> 583593 (0.54%)
helped: 0
HURT: 1272
HURT stats (abs)   min: 1.0 max: 17.0 x̄: 2.46 x̃: 2
HURT stats (rel)   min: 0.09% max: 22.22% x̄: 0.88% x̃: 0.59%
95% mean confidence interval for instructions value: 2.37 2.56
95% mean confidence interval for instructions %-change: 0.80% 0.97%
Instructions are HURT.

total cycles in shared programs: 140628.52 -> 140631.50 (<.01%)
cycles in affected programs: 474.45 -> 477.44 (0.63%)
helped: 0
HURT: 116
HURT stats (abs)   min: 0.015625 max: 0.0625 x̄: 0.03 x̃: 0
HURT stats (rel)   min: 0.21% max: 9.09% x̄: 1.16% x̃: 0.39%
95% mean confidence interval for cycles value: 0.02 0.03
95% mean confidence interval for cycles %-change: 0.88% 1.43%
Cycles are HURT.

total cvt in shared programs: 14816.20 -> 14865.12 (0.33%)
cvt in affected programs: 4954.31 -> 5003.23 (0.99%)
helped: 0
HURT: 1272
HURT stats (abs)   min: 0.015625 max: 0.265625 x̄: 0.04 x̃: 0
HURT stats (rel)   min: 0.16% max: 28.57% x̄: 1.88% x̃: 1.17%
95% mean confidence interval for cvt value: 0.04 0.04
95% mean confidence interval for cvt %-change: 1.74% 2.02%
Cvt are HURT.

total quadwords in shared programs: 1476536 -> 1478512 (0.13%)
quadwords in affected programs: 54496 -> 56472 (3.63%)
helped: 0
HURT: 247
HURT stats (abs)   min: 8.0 max: 8.0 x̄: 8.00 x̃: 8
HURT stats (rel)   min: 0.93% max: 20.00% x̄: 5.78% x̃: 4.17%
95% mean confidence interval for quadwords value: 8.00 8.00
95% mean confidence interval for quadwords %-change: 5.21% 6.34%
Quadwords are HURT.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17794>

diff --git a/src/panfrost/bifrost/bifrost_compile.c b/src/panfrost/bifrost/bifrost_compile.c
index 56d51f8..52e50e5 100644 (file)
--- a/src/panfrost/bifrost/bifrost_compile.c
+++ b/src/panfrost/bifrost/bifrost_compile.c
@@ -3988,7 +3988,6 @@ emit_block(bi_context *ctx, nir_block *block)
 
         nir_foreach_instr(instr, block) {
                 bi_emit_instr(&_b, instr);
-                ++ctx->instruction_count;
         }
 
         return ctx->current_block;
@@ -4009,9 +4008,8 @@ emit_if(bi_context *ctx, nir_if *nif)
         bi_block *then_block = emit_cf_list(ctx, &nif->then_list);
         bi_block *end_then_block = ctx->current_block;
 
-        /* Emit second block, and check if it's empty */
+        /* Emit second block */
 
-        int count_in = ctx->instruction_count;
         bi_block *else_block = emit_cf_list(ctx, &nif->else_list);
         bi_block *end_else_block = ctx->current_block;
         ctx->after_block = create_empty_block(ctx);
@@ -4021,20 +4019,15 @@ emit_if(bi_context *ctx, nir_if *nif)
         assert(then_block);
         assert(else_block);
 
-        if (ctx->instruction_count == count_in) {
-                then_branch->branch_target = ctx->after_block;
-                bi_block_add_successor(end_then_block, ctx->after_block); /* fallthrough */
-        } else {
-                then_branch->branch_target = else_block;
+        then_branch->branch_target = else_block;
 
-                /* Emit a jump from the end of the then block to the end of the else */
-                _b.cursor = bi_after_block(end_then_block);
-                bi_instr *then_exit = bi_jump(&_b, bi_zero());
-                then_exit->branch_target = ctx->after_block;
+        /* Emit a jump from the end of the then block to the end of the else */
+        _b.cursor = bi_after_block(end_then_block);
+        bi_instr *then_exit = bi_jump(&_b, bi_zero());
+        then_exit->branch_target = ctx->after_block;
 
-                bi_block_add_successor(end_then_block, then_exit->branch_target);
-                bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
-        }
+        bi_block_add_successor(end_then_block, then_exit->branch_target);
+        bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
 
         bi_block_add_successor(before_block, then_branch->branch_target); /* then_branch */
         bi_block_add_successor(before_block, then_block); /* fallthrough */

diff --git a/src/panfrost/bifrost/compiler.h b/src/panfrost/bifrost/compiler.h
index bb8f582..2d70b2f 100644 (file)
--- a/src/panfrost/bifrost/compiler.h
+++ b/src/panfrost/bifrost/compiler.h
@@ -836,7 +836,6 @@ typedef struct {
        struct hash_table_u64 *allocated_vec;
 
        /* Stats for shader-db */
-       unsigned instruction_count;
        unsigned loop_count;
        unsigned spills;
        unsigned fills;