r300: merge together MOV and MAD instructions

Assuming they write different channels of the same destination
and they share at least one source or one of the sources is
RC_FILE_NONE.

shader-db with RV530:
total instructions in shared programs: 136033 -> 135673 (-0.26%)
instructions in affected programs: 22987 -> 22627 (-1.57%)
total temps in shared programs: 18977 -> 18965 (-0.06%)
temps in affected programs: 74 -> 62 (-16.22%)

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6188

Reviewed-by: Filip Gawin <filip@gawin.net>
Signed-off-by: Pavel Ondračka <pavel.ondracka@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18288>

diff --git a/src/gallium/drivers/r300/compiler/radeon_optimize.c b/src/gallium/drivers/r300/compiler/radeon_optimize.c
index e90005f..9a33827 100644 (file)
--- a/src/gallium/drivers/r300/compiler/radeon_optimize.c
+++ b/src/gallium/drivers/r300/compiler/radeon_optimize.c
@@ -1007,7 +1007,8 @@ static int have_shared_source(struct rc_instruction * inst1, struct rc_instructi
                for (unsigned j = 0; j < opcode2->NumSrcRegs; j++) {
                        if (inst1->U.I.SrcReg[i].File == inst2->U.I.SrcReg[j].File &&
                                inst1->U.I.SrcReg[i].Index == inst2->U.I.SrcReg[j].Index &&
-                               inst1->U.I.SrcReg[i].RelAddr == inst2->U.I.SrcReg[j].RelAddr)
+                               inst1->U.I.SrcReg[i].RelAddr == inst2->U.I.SrcReg[j].RelAddr &&
+                               inst1->U.I.SrcReg[i].Abs == inst2->U.I.SrcReg[j].Abs)
                                shared_src = i;
                }
        }
@@ -1105,6 +1106,97 @@ static int merge_mov_add_mul(
        return 1;
 }
 
+/**
+ * This function will try to merge MOV and MAD instructions with the same
+ * destination, making use of the constant swizzles. This only works
+ * if there is a shared source or one of the sources is RC_FILE_NONE.
+ *
+ * For example:
+ *   MOV temp[0].x const[0].x
+ *   MAD temp[0].yz const[0].yz const[1].yz input[0].xw
+ *
+ * becomes
+ *   MAD temp[0].xyz const[0].xyz const[2].1yz input[0].0xw
+ */
+static bool merge_mov_mad(
+       struct radeon_compiler * c,
+       struct rc_instruction * inst1,
+       struct rc_instruction * inst2)
+{
+       struct rc_instruction * mov, * mad;
+       if (inst1->U.I.Opcode == RC_OPCODE_MOV) {
+               mov = inst1;
+               mad = inst2;
+       } else {
+               mov = inst2;
+               mad = inst1;
+       }
+
+       int shared_index = have_shared_source(mad, mov);
+       unsigned wmask = mov->U.I.DstReg.WriteMask | mad->U.I.DstReg.WriteMask;
+       struct rc_src_register src[3];
+       src[0] = mad->U.I.SrcReg[0];
+       src[1] = mad->U.I.SrcReg[1];
+       src[2] = mad->U.I.SrcReg[2];
+
+       /* Shared source is the one for multiplication. */
+       if (shared_index == 0 || shared_index == 1) {
+               src[shared_index].Negate = merge_negates(src[shared_index], mov->U.I.SrcReg[0]);
+               src[1 - shared_index].Negate = clean_negate(src[1 - shared_index]);
+               src[shared_index].Swizzle = merge_swizzles(src[shared_index].Swizzle,
+                               mov->U.I.SrcReg[0].Swizzle);
+               src[1 - shared_index].Swizzle = fill_swizzle(
+                               src[1 - shared_index].Swizzle, wmask, RC_SWIZZLE_ONE);
+               src[2].Swizzle =  fill_swizzle(src[2].Swizzle, wmask, RC_SWIZZLE_ZERO);
+
+       /* Shared source is the one for used for addition, or it is none. Additionally,
+        * if the mov SrcReg is none, we merge it with the addition (third) reg as well
+        * because than we have the highest change the swizzles will be legal.
+        */
+       } else if (shared_index == 2 || mov->U.I.SrcReg[0].File == RC_FILE_NONE ||
+                       src[2].File == RC_FILE_NONE) {
+               src[2].Negate = merge_negates(src[2], mov->U.I.SrcReg[0]);
+               src[2].Swizzle = merge_swizzles(src[2].Swizzle, mov->U.I.SrcReg[0].Swizzle);
+               src[0].Swizzle = fill_swizzle(src[0].Swizzle, wmask, RC_SWIZZLE_ZERO);
+               src[1].Swizzle = fill_swizzle(src[1].Swizzle, wmask, RC_SWIZZLE_ZERO);
+               if (src[2].File == RC_FILE_NONE) {
+                       src[2].File = mov->U.I.SrcReg[0].File;
+                       src[2].Index = mov->U.I.SrcReg[0].Index;
+                       src[2].RelAddr = mov->U.I.SrcReg[0].RelAddr;
+                       src[2].Abs = mov->U.I.SrcReg[0].Abs;
+               }
+
+       /* First or the second MAD source is RC_FILE_NONE, we merge the mov into it,
+        * fill the other one with ones and the reg for addition with zeros.
+        */
+       } else if (src[0].File == RC_FILE_NONE || src[1].File == RC_FILE_NONE) {
+               unsigned none_src = src[0].File == RC_FILE_NONE ? 0 : 1;
+               src[none_src] = mov->U.I.SrcReg[0];
+               src[none_src].Negate = merge_negates(src[none_src], mad->U.I.SrcReg[none_src]);
+               src[none_src].Swizzle = merge_swizzles(src[none_src].Swizzle,
+                               mad->U.I.SrcReg[none_src].Swizzle);
+               src[1 - none_src].Negate = clean_negate(src[1 - none_src]);
+               src[1 - none_src].Swizzle = fill_swizzle(src[1 - none_src].Swizzle,
+                               wmask, RC_SWIZZLE_ONE);
+               src[2].Swizzle =  fill_swizzle(src[2].Swizzle, wmask, RC_SWIZZLE_ZERO);
+       } else {
+               return false;
+       }
+
+       if (!c->SwizzleCaps->IsNative(RC_OPCODE_MAD, src[0]) ||
+               !c->SwizzleCaps->IsNative(RC_OPCODE_MAD, src[1]) ||
+               !c->SwizzleCaps->IsNative(RC_OPCODE_MAD, src[2]))
+               return false;
+
+       inst2->U.I.Opcode = RC_OPCODE_MAD;
+       inst2->U.I.SrcReg[0] = src[0];
+       inst2->U.I.SrcReg[1] = src[1];
+       inst2->U.I.SrcReg[2] = src[2];
+       inst2->U.I.DstReg.WriteMask = wmask;
+       rc_remove_instruction(inst1);
+       return true;
+}
+
 static bool inst_combination(
        struct rc_instruction * inst1,
        struct rc_instruction * inst2,
@@ -1187,6 +1279,11 @@ static void merge_channels(struct radeon_compiler * c, struct rc_instruction * i
                                if (merge_mov_add_mul(c, inst, cur))
                                        return;
                        }
+
+                       if (inst_combination(cur, inst, RC_OPCODE_MOV, RC_OPCODE_MAD)) {
+                               if (merge_mov_mad(c, inst, cur))
+                                       return;
+                       }
                }
        }
 }
@@ -1220,6 +1317,7 @@ void rc_optimize(struct radeon_compiler * c, void *user)
                        inst = inst->Next;
                        if (cur->U.I.Opcode == RC_OPCODE_MOV ||
                                cur->U.I.Opcode == RC_OPCODE_ADD ||
+                               cur->U.I.Opcode == RC_OPCODE_MAD ||
                                cur->U.I.Opcode == RC_OPCODE_MUL)
                                merge_channels(c, cur);
                }