On 8/31/22 13:39, Vineet Gupta wrote:
On 8/31/22 07:57, Palmer Dabbelt wrote:
if (flag_pic)
- riscv_cmodel = CM_PIC;
+ riscv_cmodel = CM_MEDANY;
/* We get better code with explicit relocs for CM_MEDLOW, but
worse code for the others (for now). Pick the best default. */
I'm fine either way on this one: having CM_PIC gone makes it a bit
more likely to confuse CM_MEDANY with PIC, but flag_pic is overriding
riscv_cmodel anyway so this isn't really used and deleting code is
always a plus.
Indeed this was the most contentious part of removing CM_PIC, but it
seems this is the way fwd. I'll add Kito's comment from [1] in code to
make it more explicit.
[1]https://github.com/riscv-non-isa/riscv-c-api-doc/pull/11#issuecomment-686385585
I think I'll punt on this one, in the short-term. The reason being it
affects USE_LOAD_ADDRESS_MACRO.
#define USE_LOAD_ADDRESS_MACRO(sym) \
(!TARGET_EXPLICIT_RELOCS && \
((flag_pic \
&& ((SYMBOL_REF_P (sym) && SYMBOL_REF_LOCAL_P (sym)) \
|| ((GET_CODE (sym) == CONST) \
&& SYMBOL_REF_P (XEXP (XEXP (sym, 0),0)) \
&& SYMBOL_REF_LOCAL_P (XEXP (XEXP (sym, 0),0))))) \
|| riscv_cmodel == CM_MEDANY))
With the patch, PIC implies CM_MEDANY and thus will change codegen for
pic non-local symbols to also use the load address macro. I think we
want to go in the opposite direction, i.e. wean away from the asm macros
and have gcc codegen natively. It seems there are bugs in that area so
once we flush them out (after creating a few as I don't know of any
existing documented ones) this will get cleaned out.
Thx
-Vineet
