freedreno/ir3: fix fixing-up register footprint

It isn't just vertex shaders that need to fixup reg footprint for inputs
populated before shader starts.

This problem showed up with compute shaders.  If you have (for example)
a localregid sysval, but only the .x component is used, the hw still
writes the .yz components, which could overflow into other threads
causing corruption.  Showed up in cl cts 'basic/test_basic intmath_int'.
But in theory the same problem could crop up elsewhere.

Signed-off-by: Rob Clark <robdclark@gmail.com>

diff --git a/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c b/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c
index 1c8d836a87f720e7fae2efd9a242a61f2551d517..8644bc1921890c5c652a7702d746a86006356aa4 100644 (file)
--- a/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c
@@ -3418,7 +3418,7 @@ ir3_compile_shader_nir(struct ir3_compiler *compiler,
                        so->varying_in++;
                        so->inputs[i].compmask = (1 << maxcomp) - 1;
                        inloc += maxcomp;
-               } else {
+               } else if (!so->inputs[i].sysval){
                        so->inputs[i].compmask = compmask;
                }
                so->inputs[i].regid = regid;

diff --git a/src/gallium/drivers/freedreno/ir3/ir3_shader.c b/src/gallium/drivers/freedreno/ir3/ir3_shader.c
index 3d6cae9f80ee31b50a0e1d3fee6ed3d4527a235d..555c654374e4fee7a82d364db3ded4b26760108e 100644 (file)
--- a/src/gallium/drivers/freedreno/ir3/ir3_shader.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3_shader.c
@@ -70,26 +70,35 @@ delete_variant(struct ir3_shader_variant *v)
 static void
 fixup_regfootprint(struct ir3_shader_variant *v)
 {
-       if (v->type == SHADER_VERTEX) {
-               unsigned i;
-               for (i = 0; i < v->inputs_count; i++) {
-                       /* skip frag inputs fetch via bary.f since their reg's are
-                        * not written by gpu before shader starts (and in fact the
-                        * regid's might not even be valid)
-                        */
-                       if (v->inputs[i].bary)
-                               continue;
+       unsigned i;
 
-                       if (v->inputs[i].compmask) {
-                               int32_t regid = (v->inputs[i].regid + 3) >> 2;
-                               v->info.max_reg = MAX2(v->info.max_reg, regid);
-                       }
-               }
-               for (i = 0; i < v->outputs_count; i++) {
-                       int32_t regid = (v->outputs[i].regid + 3) >> 2;
+       for (i = 0; i < v->inputs_count; i++) {
+               /* skip frag inputs fetch via bary.f since their reg's are
+                * not written by gpu before shader starts (and in fact the
+                * regid's might not even be valid)
+                */
+               if (v->inputs[i].bary)
+                       continue;
+
+               /* ignore high regs that are global to all threads in a warp
+                * (they exist by default) (a5xx+)
+                */
+               if (v->inputs[i].regid >= regid(48,0))
+                       continue;
+
+               if (v->inputs[i].compmask) {
+                       unsigned n = util_last_bit(v->inputs[i].compmask) - 1;
+                       int32_t regid = (v->inputs[i].regid + n) >> 2;
                        v->info.max_reg = MAX2(v->info.max_reg, regid);
                }
-       } else if (v->type == SHADER_FRAGMENT) {
+       }
+
+       for (i = 0; i < v->outputs_count; i++) {
+               int32_t regid = (v->outputs[i].regid + 3) >> 2;
+               v->info.max_reg = MAX2(v->info.max_reg, regid);
+       }
+
+       if (v->type == SHADER_FRAGMENT) {
                /* NOTE: not sure how to turn pos_regid off..  but this could
                 * be, for example, r1.x while max reg used by the shader is
                 * r0.*, in which case we need to fixup the reg footprint: