From a3e195ae70c569e1a140e22175220d28d8be35c0 Mon Sep 17 00:00:00 2001 From: Moritz Angermann Date: Wed, 25 Nov 2020 09:32:00 +0800 Subject: CmmToLlvm: Declare signature for memcmp Otherwise `opt` fails with: error: use of undefined value '@memcmp$def' --- compiler/GHC/CmmToLlvm/Base.hs | 10 ++++++++-- rts/linker/Elf.c | 11 +++++++++-- rts/linker/elf_reloc_aarch64.c | 3 ++- testsuite/driver/testlib.py | 10 ++++++++++ 4 files changed, 29 insertions(+), 5 deletions(-) diff --git a/compiler/GHC/CmmToLlvm/Base.hs b/compiler/GHC/CmmToLlvm/Base.hs index ead3572a79..9fde251e4b 100644 --- a/compiler/GHC/CmmToLlvm/Base.hs +++ b/compiler/GHC/CmmToLlvm/Base.hs @@ -475,14 +475,17 @@ getUniqMeta s = getEnv (flip lookupUFM s . envUniqMeta) ghcInternalFunctions :: LlvmM () ghcInternalFunctions = do platform <- getPlatform + dflags <- getDynFlags let w = llvmWord platform + cint = LMInt $ widthInBits $ cIntWidth dflags + mk "memcmp" cint [i8Ptr, i8Ptr, w] mk "memcpy" i8Ptr [i8Ptr, i8Ptr, w] mk "memmove" i8Ptr [i8Ptr, i8Ptr, w] mk "memset" i8Ptr [i8Ptr, w, w] mk "newSpark" w [i8Ptr, i8Ptr] where mk n ret args = do - let n' = llvmDefLabel $ fsLit n + let n' = fsLit n decl = LlvmFunctionDecl n' ExternallyVisible CC_Ccc ret FixedArgs (tysToParams args) Nothing renderLlvm $ ppLlvmFunctionDecl decl @@ -515,7 +518,10 @@ getGlobalPtr llvmLbl = do let mkGlbVar lbl ty = LMGlobalVar lbl (LMPointer ty) Private Nothing Nothing case m_ty of -- Directly reference if we have seen it already - Just ty -> return $ mkGlbVar (llvmDefLabel llvmLbl) ty Global + Just ty -> do + if llvmLbl `elem` (map fsLit ["newSpark", "memmove", "memcpy", "memcmp", "memset"]) + then return $ mkGlbVar (llvmLbl) ty Global + else return $ mkGlbVar (llvmDefLabel llvmLbl) ty Global -- Otherwise use a forward alias of it Nothing -> do saveAlias llvmLbl diff --git a/rts/linker/Elf.c b/rts/linker/Elf.c index 3c0ace3a4b..0a02fb452a 100644 --- a/rts/linker/Elf.c +++ b/rts/linker/Elf.c @@ -780,7 +780,12 @@ ocGetNames_ELF ( ObjectCode* oc ) else if (!oc->imageMapped || size < getPageSize() / 3) { bool executable = kind == SECTIONKIND_CODE_OR_RODATA; m32_allocator *allocator = executable ? oc->rx_m32 : oc->rw_m32; - start = m32_alloc(allocator, size, 8); + // align on 16 bytes. The reason being that llvm will emit see + // paddq statements for x86_64 under optimisation and load from + // RODATA sections. Specifically .rodata.cst16. However we don't + // handle the cst part in any way what so ever, so 16 seems + // better than 8. + start = m32_alloc(allocator, size, 16); if (start == NULL) goto fail; memcpy(start, oc->image + offset, size); alloc = SECTION_M32; @@ -915,7 +920,7 @@ ocGetNames_ELF ( ObjectCode* oc ) symbol->addr = (SymbolAddr*)( (intptr_t) oc->sections[secno].start + (intptr_t) symbol->elf_sym->st_value); - + ASSERT(symbol->addr != 0x0); if (ELF_ST_BIND(symbol->elf_sym->st_info) == STB_LOCAL) { isLocal = true; isWeak = false; @@ -1842,6 +1847,7 @@ ocResolve_ELF ( ObjectCode* oc ) #endif ASSERT(symbol->elf_sym->st_name == 0); ASSERT(symbol->elf_sym->st_value == 0); + ASSERT(0x0 != oc->sections[ secno ].start); symbol->addr = oc->sections[ secno ].start; } } @@ -1915,6 +1921,7 @@ int ocRunInit_ELF( ObjectCode *oc ) init_start = (init_t*)init_startC; init_end = (init_t*)(init_startC + shdr[i].sh_size); for (init = init_start; init < init_end; init++) { + ASSERT(0x0 != *init); (*init)(argc, argv, envv); } } diff --git a/rts/linker/elf_reloc_aarch64.c b/rts/linker/elf_reloc_aarch64.c index cdaaa67711..8cffb9f03e 100644 --- a/rts/linker/elf_reloc_aarch64.c +++ b/rts/linker/elf_reloc_aarch64.c @@ -297,7 +297,7 @@ relocateObjectCodeAarch64(ObjectCode * oc) { relTab->sectionHeader->sh_link, ELF64_R_SYM((Elf64_Xword)rel->r_info)); - assert(symbol != NULL); + assert(0x0 != symbol); /* decode implicit addend */ int64_t addend = decodeAddendAarch64(targetSection, rel); @@ -324,6 +324,7 @@ relocateObjectCodeAarch64(ObjectCode * oc) { ELF64_R_SYM((Elf64_Xword)rel->r_info)); assert(0x0 != symbol); + assert(0x0 != symbol->addr); /* take explicit addend */ int64_t addend = rel->r_addend; diff --git a/testsuite/driver/testlib.py b/testsuite/driver/testlib.py index 5812a361c4..d5d690bba5 100644 --- a/testsuite/driver/testlib.py +++ b/testsuite/driver/testlib.py @@ -2189,6 +2189,13 @@ def normalise_errmsg(s: str) -> str: s = re.sub('Failed to remove file (.*); error= (.*)$', '', s) s = re.sub('DeleteFile "(.+)": permission denied \(Access is denied\.\)(.*)$', '', s) + # filter out unsupported GNU_PROPERTY_TYPE (5), which is emitted by LLVM10 + # and not understood by older binutils (ar, ranlib, ...) + s = modify_lines(s, lambda l: re.sub('^(.+)warning: (.+): unsupported GNU_PROPERTY_TYPE \(5\) type: 0xc000000(.*)$', '', l)) + + # filter out nix garbage, that just keeps on showing up as errors on darwin + s = modify_lines(s, lambda l: re.sub('^(.+)\.dylib, ignoring unexpected dylib file$','', l)) + return s # normalise a .prof file, so that we can reasonably compare it against @@ -2259,6 +2266,9 @@ def normalise_output( s: str ) -> str: s = re.sub('([^\\s])\\.exe', '\\1', s) s = normalise_callstacks(s) s = normalise_type_reps(s) + # ghci outputs are pretty unstable with -fexternal-dynamic-refs, which is + # requires for -fPIC + s = re.sub(' -fexternal-dynamic-refs\n','',s) return s def normalise_asm( s: str ) -> str: -- cgit v1.2.1