summaryrefslogtreecommitdiffstats
path: root/llvm/lib/Target/R600/SIInstructions.td
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/lib/Target/R600/SIInstructions.td')
-rw-r--r--llvm/lib/Target/R600/SIInstructions.td47
1 files changed, 17 insertions, 30 deletions
diff --git a/llvm/lib/Target/R600/SIInstructions.td b/llvm/lib/Target/R600/SIInstructions.td
index 0449ba8cd18..7137baabc39 100644
--- a/llvm/lib/Target/R600/SIInstructions.td
+++ b/llvm/lib/Target/R600/SIInstructions.td
@@ -1433,28 +1433,36 @@ defm V_RSQ_CLAMP_F64 : VOP1InstSI <vop1<0x32>, "v_rsq_clamp_f64",
// VINTRP Instructions
//===----------------------------------------------------------------------===//
+let Uses = [M0] in {
+
// FIXME: Specify SchedRW for VINTRP insturctions.
defm V_INTERP_P1_F32 : VINTRP_m <
0x00000000, "v_interp_p1_f32",
(outs VGPR_32:$dst),
- (ins VGPR_32:$i, i32imm:$attr_chan, i32imm:$attr, M0Reg:$m0),
- "v_interp_p1_f32 $dst, $i, $attr_chan, $attr, [$m0]",
- "$m0">;
+ (ins VGPR_32:$i, i32imm:$attr_chan, i32imm:$attr),
+ "v_interp_p1_f32 $dst, $i, $attr_chan, $attr, [m0]",
+ [(set f32:$dst, (AMDGPUinterp_p1 i32:$i, (i32 imm:$attr_chan),
+ (i32 imm:$attr)))]>;
defm V_INTERP_P2_F32 : VINTRP_m <
0x00000001, "v_interp_p2_f32",
(outs VGPR_32:$dst),
- (ins VGPR_32:$src0, VGPR_32:$j, i32imm:$attr_chan, i32imm:$attr, M0Reg:$m0),
- "v_interp_p2_f32 $dst, [$src0], $j, $attr_chan, $attr, [$m0]",
- "$src0,$m0",
+ (ins VGPR_32:$src0, VGPR_32:$j, i32imm:$attr_chan, i32imm:$attr),
+ "v_interp_p2_f32 $dst, [$src0], $j, $attr_chan, $attr, [m0]",
+ [(set f32:$dst, (AMDGPUinterp_p2 f32:$src0, i32:$j, (i32 imm:$attr_chan),
+ (i32 imm:$attr)))],
+ "$src0",
"$src0 = $dst">;
defm V_INTERP_MOV_F32 : VINTRP_m <
0x00000002, "v_interp_mov_f32",
(outs VGPR_32:$dst),
- (ins InterpSlot:$src0, i32imm:$attr_chan, i32imm:$attr, M0Reg:$m0),
- "v_interp_mov_f32 $dst, $src0, $attr_chan, $attr, [$m0]",
- "$m0">;
+ (ins InterpSlot:$src0, i32imm:$attr_chan, i32imm:$attr),
+ "v_interp_mov_f32 $dst, $src0, $attr_chan, $attr, [m0]",
+ [(set f32:$dst, (AMDGPUinterp_mov (i32 imm:$src0), (i32 imm:$attr_chan),
+ (i32 imm:$attr)))]>;
+
+} // End Uses = [M0]
//===----------------------------------------------------------------------===//
// VOP2 Instructions
@@ -2724,27 +2732,6 @@ def : Pat <
(S_MOV_B64 (f64 (bitcast_fpimm_to_i64 InlineFPImm<f64>:$imm)))
>;
-/********** ===================== **********/
-/********** Interpolation Paterns **********/
-/********** ===================== **********/
-
-// The value of $params is constant through out the entire kernel.
-// We need to use S_MOV_B32 $params, because CSE ignores copies, so
-// without it we end up with a lot of redundant moves.
-
-def : Pat <
- (int_SI_fs_constant imm:$attr_chan, imm:$attr, i32:$params),
- (V_INTERP_MOV_F32 INTERP.P0, imm:$attr_chan, imm:$attr, (S_MOV_B32 $params))
->;
-
-def : Pat <
- (int_SI_fs_interp imm:$attr_chan, imm:$attr, i32:$params, v2i32:$ij),
- (V_INTERP_P2_F32 (V_INTERP_P1_F32 (EXTRACT_SUBREG v2i32:$ij, sub0),
- imm:$attr_chan, imm:$attr, (S_MOV_B32 $params)),
- (EXTRACT_SUBREG $ij, sub1),
- imm:$attr_chan, imm:$attr, (S_MOV_B32 $params))
->;
-
/********** ================== **********/
/********** Intrinsic Patterns **********/
/********** ================== **********/
OpenPOWER on IntegriCloud