flang2: An attempt to introduce opaque pointers to flang #1306
Conversation
Signed-off-by: Paul Osmialowski <pawel.osmialowski@arm.com>
|
@pawosm-arm @kiranchandramohan @shivaramaarao Is it important to keep the master branch backward-compatible with LLVM 14 and older releases, perhaps with liberal uses of macros or by duplicating files? IMHO it's probably better to create a new master branch that uses opaque pointers only, and dual-maintain both branches until nobody needs the old master branch any more. What do you think? |
| @@ -272,7 +272,7 @@ write_prototypes(FILE *out, LLVMModuleRef module) | |||
|
|
|||
| if (text_calls) { | |||
| fprintf(out, "declare i64 @llvm.nvvm.texsurf.handle.p1i64(metadata, i64 " | |||
| "addrspace(1)*) nounwind readnone\n"); | |||
| "ptr addrspace(1)) nounwind readnone\n"); | |||
There was a problem hiding this comment.
Please delete i64 in the previous line:
i64 addrspace(1)* -> ptr addrspace(1)
There was a problem hiding this comment.
ah, good cacth. Seems this is something that wouldn't ever be executed in our ecosystem.
| elem_type, AG_TYPENAME(gblsym), AG_NAME(gblsym)); | ||
| "ptr bitcast(ptr getelementptr(" | ||
| "%sptr @%s, i32 0) to ptr)", | ||
| elem_type, AG_NAME(gblsym)); |
There was a problem hiding this comment.
Please delete the bitcast instruction:
"ptr getelementptr(%sptr @%s, i32 0)",
elem_type, AG_NAME(gblsym));
| elem_type, n_elts, name, suffix ? suffix : ""); | ||
| "ptr bitcast(ptr getelementptr(" | ||
| "%sptr @%s%s, i32 0) to ptr)", | ||
| elem_type, name, suffix ? suffix : ""); |
There was a problem hiding this comment.
Please delete the bitcast instruction:
"ptr getelementptr(%sptr @%s%s, i32 0)",
elem_type, name, suffix ? suffix : "");
| initname, initname, ADDRESSG(TYPDEF_INITG(tag))); | ||
| " ptr bitcast(ptr getelementptr(i8, ptr " | ||
| "@%s, i32 %ld) to ptr),\n", | ||
| initname, ADDRESSG(TYPDEF_INITG(tag))); |
There was a problem hiding this comment.
Please delete the bitcast instruction:
" ptr getelementptr(i8, ptr @%s, i32 %ld),\n",
initname, ADDRESSG(TYPDEF_INITG(tag)));
| ",\n", name, name); | ||
| " ptr bitcast(ptr getelementptr(i8, ptr " | ||
| "@%s$parents, i32 0) to ptr)" | ||
| ",\n", name); |
There was a problem hiding this comment.
Please delete the bitcast instruction:
" ptr getelementptr(i8, ptr @%s$parents, i32 0),\n", name);
| @@ -153,7 +153,7 @@ write_proc_pointer(SPTR sptr) | |||
| if (abi) { | |||
| fntype = ll_abi_function_type(abi)->str; | |||
| } | |||
| fprintf(ASMFIL, "i8* bitcast(%s* @%s to i8*)", fntype, getsname(sptr)); | |||
| fprintf(ASMFIL, "ptr bitcast(ptr @%s to ptr)", getsname(sptr)); | |||
There was a problem hiding this comment.
The variable fntype is no longer used. Please delete the related code and the bitcast instruction.
| fprintf(ASMFIL, " }>\n@%s = alias %s%sbitcast (<{%s}>* @%s.%d to %s*)", | ||
| getsname(sptr), alTy, alSep, bare, getsname(sptr), sptr, typed); | ||
| fprintf(ASMFIL, " }>\n@%s = alias %s%sbitcast (ptr @%s.%d to ptr)", | ||
| getsname(sptr), alTy, alSep, getsname(sptr), sptr); |
There was a problem hiding this comment.
Please delete the bitcast instruction:
fprintf(ASMFIL, " }>\n@%s = alias %s%sptr @%s.%d",
getsname(sptr), alTy, alSep, getsname(sptr), sptr);
| "\t%%prof.thisfn = bitcast %s* @%s to i8*\n" | ||
| "\t%%prof.callsite = call i8*(i32) @" PROF_CALLSITE "(i32 0)\n", | ||
| currFnTy->str, currFn); | ||
| "\t%%prof.thisfn = bitcast ptr @%s to ptr\n" |
There was a problem hiding this comment.
The no-op bitcasts are redundant in the IR, please delete the bitcast instruction:
"\t%%prof.thisfn = ptr @%s\n"
| @@ -4100,7 +4100,7 @@ lldbg_function_end(LL_DebugInfo *db, int func) | |||
| sname = getsname(i); // temporary pointer | |||
| buff = (char *)getitem(LLVM_LONGTERM_AREA, strlen(name) + strlen(sname) + | |||
| strlen(type->str) + 25); | |||
| sprintf(buff, "bitcast (%%struct%s* @%s to %s)", sname, name, type->str); | |||
| sprintf(buff, "bitcast (ptr @%s to %s)", name, type->str); | |||
There was a problem hiding this comment.
The variable sname is no longer used. Please delete the related code and the bitcast instruction for the pointer type.
|
@pawosm-arm Would you have time to work on this PR? |
Good question. Sadly, time management is the reason I may need to abandon this patch. As I said, I'm taking no credits on this one, so feel free to build on it if you need it. |
|
Superceded by #1321. |
This is my naïve approach, it's not perfect, uses ptr to ptr bitcasts, but may work as well.
Won't be of any use until we shift to LLVM15.