ANDROID: x86: entry: work around LLVM_IAS=1 bug in LSL

clang-10 had a bug where it did not accept 64b GPR operands to LSL (not
"logical shift left" but "load segment limit").

It's common within the kernel to work around broken or missing assembler
support by putting raw encodings inline.  We can roughtly approximate
register only operands with the macro:

.macro LSL opd
       REG_TYPE lsl_opd_type \opd
       .if lsl_opd_type == REG_TYPE_R64
       R64_NUM lsl_opd \opd
       PFX_REX 0 0 1
       .else
       R32_NUM lsl_opd \opd
       .endif
       .byte 0x0f, 0x03
       MODRM 0xc0 lsl_opd 0x0
.endm

in arch/x86/asm/inst.h, but this doesn't handle memory operands. Since
this macro is only ever expanded currently with %rax, hardcode the
instruction encoding for now to unblock enabling LLVM_IAS=1 for x86.

Once AOSP LLVM is upgraded to a point where it contains the fix, this
should be reverted.

Bug: 141693040
Bug: 171348143
Link: https://github.com/ClangBuiltLinux/linux/issues/1079
Link: 3c2a56a857
Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>
Change-Id: I8c80f3e7883d4b0774dfd1642e31b0793c5f86ed

arch/x86/entry/calling.h

@@ -366,7 +366,11 @@ For 32-bit we have the following conventions - kernel is built with
  */
 .macro LOAD_CPU_AND_NODE_SEG_LIMIT reg:req
 	movq	$__CPUNODE_SEG, \reg
+#ifdef __clang__
+	.quad 0xc0030f48
+#else
 	lsl	\reg, \reg
+#endif
 .endm
 
 /*