nv50/ir: flip shl(add, imm) into add(shl, imm)

This works when the add also has an immediate. This often happens in
address calculations. These addresses can then be inlined as well.

On code targeted to SM35:

total instructions in shared programs : 6223346 -> 6206257 (-0.27%)
total gprs used in shared programs    : 911075 -> 911045 (-0.00%)
total local used in shared programs   : 39072 -> 39072 (0.00%)

                local        gpr       inst      bytes
    helped           0         119        3664        3664
      hurt           0          74          15          15

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
index b79e465b4faa53e7d47dfa56637e3f05a4ade5b4..9e49c19b8e1a5e28d5dbd044cab2ac1d7aa7d5a9 100644 (file)
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
@@ -1132,13 +1132,43 @@ ConstantFolding::opnd(Instruction *i, ImmediateValue &imm0, int s)
          break;
       // try to concatenate shifts
       Instruction *si = i->getSrc(0)->getInsn();
-      if (!si || si->op != OP_SHL)
+      if (!si)
          break;
       ImmediateValue imm1;
-      if (si->src(1).getImmediate(imm1)) {
+      switch (si->op) {
+      case OP_SHL:
+         if (si->src(1).getImmediate(imm1)) {
+            bld.setPosition(i, false);
+            i->setSrc(0, si->getSrc(0));
+            i->setSrc(1, bld.loadImm(NULL, imm0.reg.data.u32 + imm1.reg.data.u32));
+         }
+         break;
+      case OP_SUB:
+      case OP_ADD:
+         int adds;
+         if (isFloatType(si->dType))
+            return;
+         if (si->op != OP_SUB && si->src(0).getImmediate(imm1))
+            adds = 0;
+         else if (si->src(1).getImmediate(imm1))
+            adds = 1;
+         else
+            return;
+         // SHL(ADD(x, y), z) = ADD(SHL(x, z), SHL(y, z))
+
+         // This is more operations, but if one of x, y is an immediate, then
+         // we can get a situation where (a) we can use ISCADD, or (b)
+         // propagate the add bit into an indirect load.
          bld.setPosition(i, false);
-         i->setSrc(0, si->getSrc(0));
-         i->setSrc(1, bld.loadImm(NULL, imm0.reg.data.u32 + imm1.reg.data.u32));
+         i->op = si->op;
+         i->setSrc(adds, bld.loadImm(NULL, imm1.reg.data.u32 << imm0.reg.data.u32));
+         i->setSrc(!adds, bld.mkOp2v(OP_SHL, i->dType,
+                                     bld.getSSA(i->def(0).getSize(), i->def(0).getFile()),
+                                     si->getSrc(!adds),
+                                     bld.mkImm(imm0.reg.data.u32)));
+         break;
+      default:
+         return;
       }
    }
       break;