grad_func, flib_def,
                                          visited_functions));
    GradientDef grad;
    grad.set_function_name(function_name);
    grad.set_gradient_func(grad_func.getName().str());
    TF_RETURN_IF_ERROR(flib_def->AddGradientDef(grad));
  }

  auto stateful_string = mlir::TF::TensorFlowDialect::GetStatefulAttrName();

	"MOVDLE",
	"MOVDLT",
	"MOVDNE",
	"LOCR",
	"LOCGR",
	"FLOGR",
	"POPCNT",
	"AND",
	"ANDW",
	"OR",
	"ORW",
	"XOR",
	"XORW",
	"SLW",
	"SLD",
	"SRW",
	"SRAW",
	"SRD",
	"SRAD",
	"RLL",
	"RLLG",
	"RNSBG",
	"RXSBG",
	"ROSBG",
	"RNSBGT",
	"RXSBGT",
	"ROSBGT",
	"RISBG",
	"RISBGN",
	"RISBGZ",
	"RISBGNZ",
	"RISBHG",
	"RISBLG",
	"RISBHGZ",
	"RISBLGZ",

// with coefficients determined with a Remez exchange algorithm.

TEXT	·asinAsm(SB), NOSPLIT, $0-16
	FMOVD	x+0(FP), F0
	MOVD	$·asinrodataL15<>+0(SB), R9
	LGDR	F0, R7
	FMOVD	F0, F8
	SRAD	$32, R7
	WORD	$0xC0193FE6 //iilf  %r1,1072079005
	BYTE	$0xA0
	BYTE	$0x9D
	WORD	$0xB91700C7 //llgtr %r12,%r7
	MOVW	R12, R8
	MOVW	R1, R6
	CMPBGT	R8, R6, L2
	WORD	$0xC0193BFF //iilf  %r1,1006632959
	BYTE	$0xFF

	MOVD	$·log1pc6<>+0(SB), R1
	WFSDB	V0, V3, V4
	VLEG	$0, 0(R1), V18
	MOVD	$·log1pc5<>+0(SB), R1
	VLEG	$0, 0(R1), V16
	MOVD	R2, R5
	LGDR	F4, R3
	WORD	$0xC0190006	//iilf	%r1,425983
	BYTE	$0x7F
	BYTE	$0xFF
	SRAD	$32, R3, R3
	SUBW	R3, R1
	SRW	$16, R1, R1
	BYTE	$0x18	//lr	%r4,%r1
	BYTE	$0x41
	RISBGN	$0, $15, $48, R4, R2
	RISBGN	$16, $31, $32, R4, R5
	MOVW	R0, R6
	MOVW	R3, R7
	CMPBGT	R6, R7, L8

	"RLDICL",
	"RLDICLCC",
	"RLDIC",
	"RLDICCC",
	"CLRLSLDI",
	"ROTL",
	"ROTLW",
	"SLBIA",
	"SLBIE",
	"SLBMFEE",
	"SLBMFEV",
	"SLBMTE",
	"SLD",
	"SLDCC",
	"SRD",
	"SRAD",
	"SRADCC",
	"SRDCC",
	"EXTSWSLI",
	"EXTSWSLICC",
	"STDCCC",
	"TD",
	"SETB",
	"DWORD",
	"REMD",
	"REMDU",
	"HRFID",
	"POPCNTD",
	"POPCNTW",
	"POPCNTB",
	"CNTTZW",

// with coefficients determined with a Remez exchange algorithm.

TEXT	·acosAsm(SB), NOSPLIT, $0-16
	FMOVD	x+0(FP), F0
	MOVD	$·acosrodataL13<>+0(SB), R9
	LGDR	F0, R12
	FMOVD	F0, F10
	SRAD	$32, R12
	WORD	$0xC0293FE6	//iilf	%r2,1072079005
	BYTE	$0xA0
	BYTE	$0x9D
	WORD	$0xB917001C	//llgtr	%r1,%r12
	CMPW	R1,R2
	BGT	L2
	FMOVD	192(R9), F8
	FMADD	F0, F0, F8
	FMOVD	184(R9), F1
L3:

	return uint64(v) >> 16
}

func rshConst64Ux64Overflow8(v uint8) uint64 {
	// riscv64:"MOV\t\\$0,",-"SRL"
	return uint64(v) >> 8
}

func rshConst64x64(v int64) int64 {
	// ppc64x:"SRAD"
	// riscv64:"SRAI\t",-"OR",-"SLTIU"
	return v >> uint64(33)
}

func rshConst64x64Overflow32(v int32) int64 {
	// riscv64:"SRAIW",-"SLLI",-"SRAI\t"
	return int64(v) >> 32
}

update even if the grad is zero, but in this sparse implementation, ms
and mom will not update in iterations during which the grad is zero.

mean_square = decay * mean_square + (1-decay) * gradient ** 2
Delta = learning_rate * gradient / sqrt(mean_square + epsilon)

ms <- rho * ms_{t-1} + (1-rho) * grad * grad
mom <- momentum * mom_{t-1} + lr * grad / sqrt(ms + epsilon)
var <- var - mom
  }];

// with coefficients determined with a Remez exchange algorithm.

TEXT	·erfAsm(SB), NOSPLIT, $0-16
	FMOVD	x+0(FP), F0
	MOVD	$·erfrodataL13<>+0(SB), R5
	LGDR	F0, R1
	FMOVD	F0, F6
	SRAD	$48, R1
	MOVH	$16383, R3
	RISBGZ	$49, $63, $0, R1, R2
	MOVW	R2, R6
	MOVW	R3, R7
	CMPBGT	R6, R7, L2
	MOVH	$12287, R1
	MOVW	R1, R7
	CMPBLE	R6, R7 ,L12
	MOVH	$16367, R1
	MOVW	R1, R7

	SLW	$4, R3, R6              // eb63000400df
	SLW	R2, R3, R6              // eb63200000df
	SLD	$4, R3, R6              // eb630004000d
	SLD	R2, R3, R6              // eb632000000d
	SRAD	$4, R5, R8              // eb850004000a
	SRAD	R3, R5, R8              // eb853000000a
	SRAW	$4, R5, R8              // eb85000400dc
	SRAW	R3, R5, R8              // eb85300000dc
	RLL	R1, R2, R3              // eb321000001d