POPL		AX

// LTYPE3 rimrem	{ outcode(int($1), &$2); }
	SUBB	$1, AX
	SUBB	$1, foo+4(SB)
	SUBB	BX, AX
	SUBB	BX, foo+4(SB)

// LTYPE4 remrim	{ outcode(int($1), &$2); }
	CMPB	AX, $1
	CMPB	foo+4(SB), $4
	CMPB	BX, AX
	CMPB	foo+4(SB), BX

// LTYPER nonrel	{ outcode(int($1), &$2); }
label:
	JC	label // JCS
	JC	-1(PC) // JCS -1(PC)

// LTYPEC spec3	{ outcode(int($1), &$2); }
	CALL	AX
	JCS	2(PC)

func si1c(c <-chan int64) {
	// amd64:"CMPQ\t.+, [$]256"
	// s390x:"CLGIJ\t[$]12, R[0-9]+, [$]255"
	for x := <-c; x >= 0 && x < 256; x = <-c {
	}
}

func si2c(c <-chan int32) {
	// amd64:"CMPL\t.+, [$]256"
	// s390x:"CLIJ\t[$]12, R[0-9]+, [$]255"
	for x := <-c; x >= 0 && x < 256; x = <-c {
	}
}

func si3c(c <-chan int16) {
	// amd64:"CMPW\t.+, [$]256"
	// s390x:"CLIJ\t[$]12, R[0-9]+, [$]255"

  TFE_DeleteContext(ctx);
}
TEST(CAPI, TensorHandleSilentCopy) {
  TensorHandleSilentCopy(false, TFE_DEVICE_PLACEMENT_SILENT,
                         TFE_DEVICE_PLACEMENT_SILENT, false);
}
TEST(CAPI, TensorHandleSilentCopyAsync) {
  TensorHandleSilentCopy(true, TFE_DEVICE_PLACEMENT_SILENT,
                         TFE_DEVICE_PLACEMENT_SILENT, false);
}
TEST(CAPI, TensorHandleSilentCopyLocalPolicy) {

#else
tail:	// Count the last 0 - 31 bytes.
	CMP	R4, $16
	BLT	tail_8
	MOVD	(R3), R12
	MOVD	8(R3), R14
	CMPB	R12, R5, R12
	CMPB	R14, R5, R14
	POPCNTD	R12, R12
	POPCNTD	R14, R14
	ADD	R12, R18, R18
	ADD	R14, R18, R18
	ADD	$16, R3, R3
	ADD	$-16, R4, R4

tail_8:	// Count the remaining 0 - 15 bytes.
	CMP	R4, $8
	BLT	tail_4
	MOVD	(R3), R12
	CMPB	R12, R5, R12
	POPCNTD	R12, R12
	ADD	R12, R18, R18

	// for large sizes. The cutover is approximately 1K.
tail:
	TESTL	BX, BX
	JEQ	move_0
	CMPL	BX, $2
	JBE	move_1or2
	CMPL	BX, $4
	JB	move_3
	JE	move_4
	CMPL	BX, $8
	JBE	move_5through8
	CMPL	BX, $16
	JBE	move_9through16

/*
 * check and set for backwards
 */
	CMPL	SI, DI
	JLS	back

/*
 * forward copy loop
 */
forward:
	MOVL	BX, CX
	SHRL	$2, CX
	ANDL	$3, BX

	MOVQ	$usleep2<>(SB), AX // to hide from 6l

	// Execute call on m->g0.
	get_tls(R15)
	CMPQ	R15, $0
	JE	noswitch

	MOVQ	g(R15), R13
	CMPQ	R13, $0
	JE	noswitch
	MOVQ	g_m(R13), R13
	CMPQ	R13, $0
	JE	noswitch
	// TODO(aram): do something about the cpu profiler here.

	MOVQ	m_g0(R13), R14
	CMPQ	g(R15), R14
	JNE	switch
	// executing on m->g0 already
	CALL	AX
	RET

switch:

  EXPECT_EQ(22, product[3]);
  TF_DeleteStatus(status);
}
TEST(CAPI, Executor_MatMul_CPU) { Executor_MatMul_CPU(false); }
TEST(CAPI, Executor_MatMul_CPUAsync) { Executor_MatMul_CPU(true); }

void Deleter(void* data, size_t unused, void* tensor_handle) {
  TFE_DeleteTensorHandle(static_cast<TFE_TensorHandle*>(tensor_handle));
}

TEST(CAPI, TensorHandleOnDeviceMemory) {
  TF_Status* status = TF_NewStatus();

//   b in DI
//   count in BX
// Output:
//   result in AX
TEXT memeqbody<>(SB),NOSPLIT,$0-0
	CMPQ	BX, $8
	JB	small
	CMPQ	BX, $64
	JB	bigloop
#ifndef hasAVX2
	CMPB	internal∕cpu·X86+const_offsetX86HasAVX2(SB), $1
	JE	hugeloop_avx2

	// 64 bytes at a time using xmm registers
	PCALIGN $16
hugeloop:
	CMPQ	BX, $64
	JB	bigloop
	MOVOU	(SI), X0
	MOVOU	(DI), X1
	MOVOU	16(SI), X2
	MOVOU	16(DI), X3

  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
  TF_DeleteGraph(graph);
  TF_DeleteStatus(s);
}

TEST(CAPI, Session_Min_CPU) { RunMinTest(/*device=*/"", /*use_XLA=*/false); }

TEST(CAPI, Session_Min_XLA_CPU) { RunMinTest(/*device=*/"", /*use_XLA=*/true); }

TEST(CAPI, Session_Min_GPU) {
  const string gpu_device = GPUDeviceName();
  // Skip this test if no GPU is available.

	/* If there is less than 16 bytes left we are done */
	CMPQ        CX, $16
	JB          finish

	/* Encode 16 bytes */
remain16:
	MOVOU       (SI), X10
	MOVOA       X1, X5
	PCLMULQDQ   $0, X0, X1
	PCLMULQDQ   $0x11, X0, X5
	PXOR        X5, X1
	PXOR        X10, X1
	SUBQ        $16, CX
	ADDQ        $16, SI
	CMPQ        CX, $16
	JGE         remain16

finish: