Remove indirect-load for constants on Xtensa Target to improve performance

[dongsheng28849455]

In XIP mode on Xtensa platform, compared with non XIP, there may be some performance degradation. The Coremark test results on the ESP32/Nuttx platform are:

Non-XIP: 114.94
XIP: 83

According to the handling of constants by the Xtensa backend, even without indrect load, there will be no relocation. Because the backend does not perform fixup processing on immediate numbers when processing jump/branch instructions, Refer to getL32RTargetEncoding
So for the Xtensa platform, the indirct loading is not so necessary, after removing it, the performance is improve as below(Coremark):

Opt XIP: 112.9

[no1wudi]

How about implement this by modify the intrinsic flags instead?

wasm-micro-runtime/core/iwasm/aot/aot_intrinsic.c

Lines 870 to 884 in 8acaaa2

	else if (!strncmp(comp_ctx->target_arch, "xtensa", 6)) {
	/*
	* Note: Use builtin intrinsics since hardware float operation
	* will cause rodata relocation
	*/
	add_f32_common_intrinsics(comp_ctx);
	add_i32_common_intrinsics(comp_ctx);
	add_f64_common_intrinsics(comp_ctx);
	add_i64_common_intrinsics(comp_ctx);
	add_common_float_integer_conversion(comp_ctx);
	add_intrinsic_capability(comp_ctx, AOT_INTRINSIC_FLAG_F32_CONST);
	add_intrinsic_capability(comp_ctx, AOT_INTRINSIC_FLAG_F64_CONST);
	add_intrinsic_capability(comp_ctx, AOT_INTRINSIC_FLAG_I32_CONST);
	add_intrinsic_capability(comp_ctx, AOT_INTRINSIC_FLAG_I64_CONST);
	}

[no1wudi]

How about add more intrinsic flags for conversions?

[lum1n0us]

Perhaps there ought to be a function to eliminate a particular intrinsic capability, such as remove_intrinsic_capability(comp_ctx, AOT_INTRINSIC_FLAG_F32_CMP)?