nvc0/ir: disable support for LIMMs on MAD/FMA

I hit an assert in the emiter while toying around with optimizations, because
ConstantFolding immediated a big int into a mad.

There is special handling for FMA/MAD in insnCanLoad, which is broken. With
this patch the special path should be not hit anymore. Anyway, the constraints
for the LIMMS can't be guarenteed in SSA form and I have patches pending to
use it via a post-SSA optimization pass.

As a result, immediates get immediated for int mad/fmas as well.

changes in shader-db:
total instructions in shared programs : 3943335 -> 3941587 (-0.04%)
total gprs used in shared programs    : 481563 -> 481511 (-0.01%)
total local used in shared programs   : 27469 -> 27469 (0.00%)
total bytes used in shared programs   : 36139384 -> 36123344 (-0.04%)

Signed-off-by: Karol Herbst <karolherbst@gmail.com>
[imirkin: remove extra bit from insnCanLoad as well]
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_target_nvc0.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_target_nvc0.cpp
index 1d2e4c3c090f7b2ac08a1e45b7afd036a36f8a90..0df528f60cb276da6c5ea9fb27e66bdd45061c37 100644 (file)
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_target_nvc0.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_target_nvc0.cpp
@@ -104,8 +104,8 @@ static const struct opProperties _initProps[] =
    { OP_MUL,    0x3, 0x0, 0x0, 0x8, 0x2, 0x2 | 0x8 },
    { OP_MAX,    0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
    { OP_MIN,    0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
-   { OP_MAD,    0x7, 0x0, 0x0, 0x8, 0x6, 0x2 | 0x8 }, // special c[] constraint
-   { OP_FMA,    0x7, 0x0, 0x0, 0x8, 0x6, 0x2 | 0x8 }, // keep the same as OP_MAD
+   { OP_MAD,    0x7, 0x0, 0x0, 0x8, 0x6, 0x2 }, // special c[] constraint
+   { OP_FMA,    0x7, 0x0, 0x0, 0x8, 0x6, 0x2 }, // keep the same as OP_MAD
    { OP_SHLADD, 0x5, 0x0, 0x0, 0x0, 0x4, 0x6 },
    { OP_MADSP,  0x0, 0x0, 0x0, 0x0, 0x6, 0x2 },
    { OP_ABS,    0x0, 0x0, 0x0, 0x0, 0x1, 0x0 },
@@ -381,12 +381,6 @@ TargetNVC0::insnCanLoad(const Instruction *i, int s,
             return false;
          }
       } else
-      if (i->op == OP_MAD || i->op == OP_FMA) {
-         // requires src == dst, cannot decide before RA
-         // (except if we implement more constraints)
-         if (ld->getSrc(0)->asImm()->reg.data.u32 & 0xfff)
-            return false;
-      } else
       if (i->op == OP_ADD && i->sType == TYPE_F32) {
          // add f32 LIMM cannot saturate
          if (i->saturate && (reg.data.u32 & 0xfff))