def TEX_SHADOW : PatLeaf<
(imm),
[{uint32_t TType = (uint32_t)N->getZExtValue();
- return (TType >= 6 && TType <= 8) || (TType >= 11 && TType <= 13);
+ return (TType >= 6 && TType <= 8) || TType == 13;
}]
>;
def load_param_exti8 : LoadParamFrag<az_extloadi8>;
def load_param_exti16 : LoadParamFrag<az_extloadi16>;
-def isR600 : Predicate<"Subtarget.getGeneration() <= AMDGPUSubtarget::R700">;
+def isR600 : Predicate<"Subtarget->getGeneration() <= AMDGPUSubtarget::R700">;
-def isR600toCayman : Predicate<
- "Subtarget.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS">;
+def isR600toCayman
+ : Predicate<
+ "Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS">;
//===----------------------------------------------------------------------===//
// R600 SDNodes
multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
(ExportInst
- (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
+ (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), $reg, sub0),
0, 61, 0, 7, 7, 7, cf_inst, 0)
>;
def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
(ExportInst
- (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
+ (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), $reg, sub0),
0, 61, 7, 0, 7, 7, cf_inst, 0)
>;
// Stream1
def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
(i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
- (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
+ (ExportInst $src, 0, imm:$arraybase,
4095, imm:$mask, buf1inst, 0)>;
// Stream2
def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
(i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
- (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
+ (ExportInst $src, 0, imm:$arraybase,
4095, imm:$mask, buf2inst, 0)>;
// Stream3
def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
(i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
- (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
+ (ExportInst $src, 0, imm:$arraybase,
4095, imm:$mask, buf3inst, 0)>;
}
// Non-IEEE MUL: 0 * anything = 0
def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
-def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
-def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
+// TODO: Do these actually match the regular fmin/fmax behavior?
+def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax_legacy>;
+def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin_legacy>;
// For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
// so some of the instruction names don't match the asm string.
def SNE : R600_2OP <
0xB, "SETNE",
- [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_UNE))]
+ [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_UNE_NE))]
>;
def SETE_DX10 : R600_2OP <
[(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_OGE))]
>;
+// FIXME: This should probably be COND_ONE
def SETNE_DX10 : R600_2OP <
0xF, "SETNE_DX10",
- [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_UNE))]
+ [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_UNE_NE))]
>;
def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
-def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
+def TRUNC : R600_1OP_Helper <0x11, "TRUNC", ftrunc>;
def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
-// Add also ftrunc intrinsic pattern
-def : Pat<(ftrunc f32:$src0), (TRUNC $src0)>;
-
def MOV : R600_1OP <0x19, "MOV", []>;
let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
[(set f32:$dst, (fadd (fmul f32:$src0, f32:$src1), f32:$src2))]
>;
+class FMA_Common <bits<5> inst> : R600_3OP <
+ inst, "FMA",
+ [(set f32:$dst, (fma f32:$src0, f32:$src1, f32:$src2))], VecALU
+>;
+
class CNDE_Common <bits<5> inst> : R600_3OP <
inst, "CNDE",
[(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_OEQ))]
}
class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
- inst, "RECIP_IEEE", [(set f32:$dst, (fdiv FP_ONE, f32:$src0))]
+ inst, "RECIP_IEEE", [(set f32:$dst, (AMDGPUrcp f32:$src0))]
> {
let Itinerary = TransALU;
}
let Itinerary = TransALU;
}
+// Clamped to maximum.
class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
- inst, "RECIPSQRT_CLAMPED", AMDGPUrsq
+ inst, "RECIPSQRT_CLAMPED", AMDGPUrsq_clamped
> {
let Itinerary = TransALU;
}
-class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
- inst, "RECIPSQRT_IEEE", []
+class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
+ inst, "RECIPSQRT_IEEE", AMDGPUrsq_legacy
> {
let Itinerary = TransALU;
}
+// TODO: There is also RECIPSQRT_FF which clamps to zero.
+
class SIN_Common <bits<11> inst> : R600_1OP <
inst, "SIN", [(set f32:$dst, (SIN_HW f32:$src0))]>{
let Trig = 1;
// Helper patterns for complex intrinsics
//===----------------------------------------------------------------------===//
+// FIXME: Should be predicated on unsafe fp math.
multiclass DIV_Common <InstR600 recip_ieee> {
def : Pat<
(int_AMDGPU_div f32:$src0, f32:$src1),
(fdiv f32:$src0, f32:$src1),
(MUL_IEEE $src0, (recip_ieee $src1))
>;
+
+def : RcpPat<recip_ieee, f32>;
}
class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee>
(exp_ieee (mul_lit (log_clamped (MAX $src_y, (f32 ZERO))), $src_w, $src_x))
>;
-// FROUND pattern
-class FROUNDPat<Instruction CNDGE> : Pat <
- (AMDGPUround f32:$x),
- (CNDGE (ADD (FNEG_R600 (f32 HALF)), (FRACT $x)), (CEIL $x), (FLOOR $x))
->;
-
-
//===----------------------------------------------------------------------===//
// R600 / R700 Instructions
//===----------------------------------------------------------------------===//
def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_r600 $src))>;
- def : FROUNDPat <CNDGE_r600>;
+ def : RsqPat<RECIPSQRT_IEEE_r600, f32>;
def R600_ExportSwz : ExportSwzInst {
let Word1{20-17} = 0; // BURST_COUNT
def CF_PUSH_ELSE_R600 : CF_CLAUSE_R600<12, (ins i32imm:$ADDR),
"PUSH_ELSE @$ADDR"> {
let CNT = 0;
+ let POP_COUNT = 0; // FIXME?
}
def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
"ELSE @$ADDR POP:$POP_COUNT"> {
let Pattern =
[(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
let AsmString = "CONST_COPY";
- let neverHasSideEffects = 1;
+ let hasSideEffects = 0;
let isAsCheapAsAMove = 1;
let Itinerary = NullALU;
}
let mayLoad = 0;
let mayStore = 0;
let hasSideEffects = 0;
+ let isCodeGenOnly = 1;
}
multiclass BranchConditional<SDNode Op, RegisterClass rci, RegisterClass rcf> {