gk110/ir: Use the new rcp/rsq in library

v2: (Karol Herbst <kherbst@redhat.com>
 * fix Value setup for the builtins

Signed-off-by: Boyan Ding <boyan.j.ding@gmail.com>
[imirkin: track the fp64 flag when switching ops to calls]
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: 19.0 <mesa-stable@lists.freedesktop.org>
(cherry picked from commit 656ad06051)

src/gallium/drivers/nouveau/codegen/nv50_ir.cpp

@@ -1119,6 +1119,7 @@ Program::Program(Type type, Target *arch)
    binSize = 0;
 
    maxGPR = -1;
+   fp64 = false;
 
    main = new Function(this, "MAIN", ~0);
    calls.insert(&main->call);

src/gallium/drivers/nouveau/codegen/nv50_ir.h

@@ -1311,6 +1311,7 @@ public:
    uint32_t tlsSize; // size required for FILE_MEMORY_LOCAL
 
    int maxGPR;
+   bool fp64;
 
    MemoryPool mem_Instruction;
    MemoryPool mem_CmpInstruction;

src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp

@@ -83,6 +83,38 @@ NVC0LegalizeSSA::handleDIV(Instruction *i)
    delete_Instruction(prog, i);
 }
 
+void
+NVC0LegalizeSSA::handleRCPRSQLib(Instruction *i, Value *src[])
+{
+   FlowInstruction *call;
+   Value *def[2];
+   int builtin;
+
+   def[0] = bld.mkMovToReg(0, src[0])->getDef(0);
+   def[1] = bld.mkMovToReg(1, src[1])->getDef(0);
+
+   if (i->op == OP_RCP)
+      builtin = NVC0_BUILTIN_RCP_F64;
+   else
+      builtin = NVC0_BUILTIN_RSQ_F64;
+
+   call = bld.mkFlow(OP_CALL, NULL, CC_ALWAYS, NULL);
+   def[0] = bld.getSSA();
+   def[1] = bld.getSSA();
+   bld.mkMovFromReg(def[0], 0);
+   bld.mkMovFromReg(def[1], 1);
+   bld.mkClobber(FILE_GPR, 0x3fc, 2);
+   bld.mkClobber(FILE_PREDICATE, i->op == OP_RSQ ? 0x3 : 0x1, 0);
+   bld.mkOp2(OP_MERGE, TYPE_U64, i->getDef(0), def[0], def[1]);
+
+   call->fixed = 1;
+   call->absolute = call->builtin = 1;
+   call->target.builtin = builtin;
+   delete_Instruction(prog, i);
+
+   prog->fp64 = true;
+}
+
 void
 NVC0LegalizeSSA::handleRCPRSQ(Instruction *i)
 {
@@ -96,6 +128,12 @@ NVC0LegalizeSSA::handleRCPRSQ(Instruction *i)
    Value *src[2], *dst[2], *def = i->getDef(0);
    bld.mkSplit(src, 4, i->getSrc(0));
 
+   int chip = prog->getTarget()->getChipset();
+   if (chip >= NVISA_GK20A_CHIPSET && chip < NVISA_GM107_CHIPSET) {
+      handleRCPRSQLib(i, src);
+      return;
+   }
+
    // 2. We don't care about the low 32 bits of the destination. Stick a 0 in.
    dst[0] = bld.loadImm(NULL, 0);
    dst[1] = bld.getSSA();

src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h

@@ -62,6 +62,7 @@ private:
 
    // we want to insert calls to the builtin library only after optimization
    void handleDIV(Instruction *); // integer division, modulus
+   void handleRCPRSQLib(Instruction *, Value *[]);
    void handleRCPRSQ(Instruction *); // double precision float recip/rsqrt
    void handleFTZ(Instruction *);
    void handleSET(CmpInstruction *);

src/gallium/drivers/nouveau/codegen/nv50_ir_target.cpp

@@ -399,6 +399,7 @@ Program::emitBinary(struct nv50_ir_prog_info *info)
          }
       }
    }
+   info->io.fp64 |= fp64;
    info->bin.relocData = emit->getRelocInfo();
    info->bin.fixupData = emit->getFixupInfo();