panfrost/midgard: imov workaround

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>

diff --git a/src/gallium/drivers/panfrost/midgard/midgard_compile.c b/src/gallium/drivers/panfrost/midgard/midgard_compile.c
index 9098727aa15d09a2485808697aa9d4b1deaeaed1..2b97bbc86f636cd8bc4458bb61ff71456cd80280 100644 (file)
--- a/src/gallium/drivers/panfrost/midgard/midgard_compile.c
+++ b/src/gallium/drivers/panfrost/midgard/midgard_compile.c
@@ -1134,12 +1134,7 @@ emit_alu(compiler_context *ctx, nir_alu_instr *instr)
                 ALU_CASE(isub, isub);
                 ALU_CASE(imul, imul);
                 ALU_CASE(iabs, iabs);
-
-                /* XXX: Use fmov, not imov for now, since NIR does not
-                 * differentiate well (it'll happily emits imov for floats,
-                 * which the hardware rather dislikes and breaks e.g
-                 * -bjellyfish */
-                ALU_CASE(imov, fmov);
+                ALU_CASE(imov, imov);
 
                 ALU_CASE(feq32, feq);
                 ALU_CASE(fne32, fne);
@@ -3236,6 +3231,31 @@ midgard_opt_copy_prop_tex(compiler_context *ctx, midgard_block *block)
         return progress;
 }
 
+/* We don't really understand the imov/fmov split, so always use fmov (but let
+ * it be imov in the IR so we don't do unsafe floating point "optimizations"
+ * and break things */
+
+static void
+midgard_imov_workaround(compiler_context *ctx, midgard_block *block)
+{
+        mir_foreach_instr_in_block_safe(block, ins) {
+                if (ins->type != TAG_ALU_4) continue;
+                if (ins->alu.op != midgard_alu_op_imov) continue;
+
+                ins->alu.op = midgard_alu_op_fmov;
+                ins->alu.outmod = midgard_outmod_none;
+
+                /* Remove flags that don't make sense */
+
+                midgard_vector_alu_src s =
+                        vector_alu_from_unsigned(ins->alu.src2);
+
+                s.mod = 0;
+
+                ins->alu.src2 = vector_alu_srco_unsigned(s);
+        }
+}
+
 /* The following passes reorder MIR instructions to enable better scheduling */
 
 static void
@@ -3487,6 +3507,7 @@ emit_block(compiler_context *ctx, nir_block *block)
 
         midgard_emit_store(ctx, this_block);
         midgard_pair_load_store(ctx, this_block);
+        midgard_imov_workaround(ctx, this_block);
 
         /* Append fragment shader epilogue (value writeout) */
         if (ctx->stage == MESA_SHADER_FRAGMENT) {