gm107/ir: Emit third src 'bound' and optional predicate output of SHFL

v2: Emit the original hard-coded 0x1c03 when OP_SHFL is used in gm107's
    lowering (Samuel Pitoiset)

Signed-off-by: Boyan Ding <[email protected]>
Reviewed-by: Ilia Mirkin <[email protected]>

2 files changed, 29 insertions, 9 deletions

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp
index c3c0dcd9fc1..944563c93cf 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp
@@ -967,11 +967,26 @@ CodeEmitterGM107::emitSHFL()
       break;
    }
 
-   /*XXX: what is this arg? hardcode immediate for now */
-   emitField(0x22, 13, 0x1c03);
-   type |= 2;
+   switch (insn->src(2).getFile()) {
+   case FILE_GPR:
+      emitGPR(0x27, insn->src(2));
+      break;
+   case FILE_IMMEDIATE:
+      emitIMMD(0x22, 13, insn->src(2));
+      type |= 2;
+      break;
+   default:
+      assert(!"invalid src2 file");
+      break;
+   }
+
+   if (!insn->defExists(1))
+      emitPRED(0x30);
+   else {
+      assert(insn->def(1).getFile() == FILE_PREDICATE);
+      emitPRED(0x30, insn->def(1));
+   }
 
-   emitPRED (0x30);
    emitField(0x1e, 2, insn->subOp);
    emitField(0x1c, 2, type);
    emitGPR  (0x08, insn->src(0));
diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp
index 371ebae40c1..6b9edd48645 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp
@@ -41,6 +41,8 @@ namespace nv50_ir {
    ((QOP_##q << 6) | (QOP_##r << 4) |           \
     (QOP_##s << 2) | (QOP_##t << 0))
 
+#define SHFL_BOUND_QUAD 0x1c03
+
 void
 GM107LegalizeSSA::handlePFETCH(Instruction *i)
 {
@@ -120,7 +122,8 @@ GM107LoweringPass::handleManualTXD(TexInstruction *i)
       // mov coordinates from lane l to all lanes
       bld.mkOp(OP_QUADON, TYPE_NONE, NULL);
       for (c = 0; c < dim; ++c) {
-         bld.mkOp2(OP_SHFL, TYPE_F32, crd[c], i->getSrc(c + array), bld.mkImm(l));
+         bld.mkOp3(OP_SHFL, TYPE_F32, crd[c], i->getSrc(c + array),
+                   bld.mkImm(l), bld.mkImm(SHFL_BOUND_QUAD));
          add = bld.mkOp2(OP_QUADOP, TYPE_F32, crd[c], crd[c], zero);
          add->subOp = 0x00;
          add->lanes = 1; /* abused for .ndv */
@@ -128,7 +131,8 @@ GM107LoweringPass::handleManualTXD(TexInstruction *i)
 
       // add dPdx from lane l to lanes dx
       for (c = 0; c < dim; ++c) {
-         bld.mkOp2(OP_SHFL, TYPE_F32, tmp, i->dPdx[c].get(), bld.mkImm(l));
+         bld.mkOp3(OP_SHFL, TYPE_F32, tmp, i->dPdx[c].get(), bld.mkImm(l),
+                   bld.mkImm(SHFL_BOUND_QUAD));
          add = bld.mkOp2(OP_QUADOP, TYPE_F32, crd[c], tmp, crd[c]);
          add->subOp = qOps[l][0];
          add->lanes = 1; /* abused for .ndv */
@@ -136,7 +140,8 @@ GM107LoweringPass::handleManualTXD(TexInstruction *i)
 
       // add dPdy from lane l to lanes dy
       for (c = 0; c < dim; ++c) {
-         bld.mkOp2(OP_SHFL, TYPE_F32, tmp, i->dPdy[c].get(), bld.mkImm(l));
+         bld.mkOp3(OP_SHFL, TYPE_F32, tmp, i->dPdy[c].get(), bld.mkImm(l),
+                   bld.mkImm(SHFL_BOUND_QUAD));
          add = bld.mkOp2(OP_QUADOP, TYPE_F32, crd[c], tmp, crd[c]);
          add->subOp = qOps[l][1];
          add->lanes = 1; /* abused for .ndv */
@@ -203,8 +208,8 @@ GM107LoweringPass::handleDFDX(Instruction *insn)
       break;
    }
 
-   shfl = bld.mkOp2(OP_SHFL, TYPE_F32, bld.getScratch(),
-                    insn->getSrc(0), bld.mkImm(xid));
+   shfl = bld.mkOp3(OP_SHFL, TYPE_F32, bld.getScratch(), insn->getSrc(0),
+                    bld.mkImm(xid), bld.mkImm(SHFL_BOUND_QUAD));
    shfl->subOp = NV50_IR_SUBOP_SHFL_BFLY;
    insn->op = OP_QUADOP;
    insn->subOp = qop;