This is a follow-up to my previous LLVM13 support patches (the amdgcn
port uses the LLVM assembler) to fix up a corner case.
With this patch one can now enable debug information in LLVM 13 offload
binaries. This was trickier than you'd think because the different LLVM
versions have different attribute interfaces and behaviours and when you
fix one issue another issue pops up in another case. The root of the
problem with debug is that mkoffload has to set the ELF flags on the
early debug binary the same way the assembler will do in all supported
cases, or else it won't link.
The only known remaining problem with LLVM 13 compatibility is an
assembler error with kernels that use mapped variables. It only affects
a few test cases in the testsuite.
LLVMs 10, 11, and 12 remain unsupported.
Andrew
amdgcn: fix up offload debug linking with LLVM 13
Between LLVM 9 and LLVM 13 the attribute works differently in several ways,
and this needs to be allowed for in GCC and mkoffload independently.
This patch fixes up mkoffload when debug info is enabled, which is made more
complicated because the configure tests checks whether the attribute option
is accepted silently, but does not check if the assembler actually sets the
ELF flags for that attribute, and mkoffload needs to mimick that behaviour
exactly. The patch therefore removes some of the conditionals.
gcc/ChangeLog:
* config/gcn/gcn-hsa.h (S_FIJI): Set unconditionally.
(S_900): Likewise.
(S_906): Likewise.
* config/gcn/gcn.c: Hard code SRAM ECC settings for old architectures.
* config/gcn/mkoffload.c (ELFABIVERSION_AMDGPU_HSA): Rename to ...
(ELFABIVERSION_AMDGPU_HSA_V3): ... this.
(ELFABIVERSION_AMDGPU_HSA_V4): New.
(SET_SRAM_ECC_UNSUPPORTED): New.
(copy_early_debug_info): Create elf flags to match the other objects.
(main): Just let the attribute flags pass through.
diff --git a/gcc/config/gcn/gcn-hsa.h b/gcc/config/gcn/gcn-hsa.h
index 6a432d17d99f..4fd2f07b9836 100644
--- a/gcc/config/gcn/gcn-hsa.h
+++ b/gcc/config/gcn/gcn-hsa.h
@@ -96,21 +96,10 @@ extern unsigned int gcn_local_sym_hash (const char *name);
#define X_908 "march=gfx908:;"
#endif
-#ifdef HAVE_GCN_SRAM_ECC_FIJI
-#define S_FIJI
-#else
+/* These targets can't have SRAM-ECC, even if a broken assembler allows it. */
#define S_FIJI "!march=*:;march=fiji:;"
-#endif
-#ifdef HAVE_GCN_SRAM_ECC_GFX900
-#define S_900
-#else
#define S_900 "march=gfx900:;"
-#endif
-#ifdef HAVE_GCN_SRAM_ECC_GFX906
-#define S_906
-#else
#define S_906 "march=gfx906:;"
-#endif
#ifdef HAVE_GCN_SRAM_ECC_GFX908
#define S_908
#else
diff --git a/gcc/config/gcn/gcn.c b/gcc/config/gcn/gcn.c
index 2e90f327c451..75a9c5766947 100644
--- a/gcc/config/gcn/gcn.c
+++ b/gcc/config/gcn/gcn.c
@@ -5226,27 +5226,21 @@ output_file_start (void)
#ifndef HAVE_GCN_XNACK_FIJI
use_xnack_attr = false;
#endif
-#ifndef HAVE_GCN_SRAM_ECC_FIJI
use_sram_attr = false;
-#endif
break;
case PROCESSOR_VEGA10:
cpu = "gfx900";
#ifndef HAVE_GCN_XNACK_GFX900
use_xnack_attr = false;
#endif
-#ifndef HAVE_GCN_SRAM_ECC_GFX900
use_sram_attr = false;
-#endif
break;
case PROCESSOR_VEGA20:
cpu = "gfx906";
#ifndef HAVE_GCN_XNACK_GFX906
use_xnack_attr = false;
#endif
-#ifndef HAVE_GCN_SRAM_ECC_GFX906
use_sram_attr = false;
-#endif
break;
case PROCESSOR_GFX908:
cpu = "gfx908";
diff --git a/gcc/config/gcn/mkoffload.c b/gcc/config/gcn/mkoffload.c
index a3b22d059b96..b2e71ea5aa00 100644
--- a/gcc/config/gcn/mkoffload.c
+++ b/gcc/config/gcn/mkoffload.c
@@ -42,8 +42,10 @@
#undef ELFOSABI_AMDGPU_HSA
#define ELFOSABI_AMDGPU_HSA 64
-#undef ELFABIVERSION_AMDGPU_HSA
-#define ELFABIVERSION_AMDGPU_HSA 1
+#undef ELFABIVERSION_AMDGPU_HSA_V3
+#define ELFABIVERSION_AMDGPU_HSA_V3 1
+#undef ELFABIVERSION_AMDGPU_HSA_V4
+#define ELFABIVERSION_AMDGPU_HSA_V4 2
#undef EF_AMDGPU_MACH_AMDGCN_GFX803
#define EF_AMDGPU_MACH_AMDGCN_GFX803 0x2a
@@ -77,6 +79,7 @@
#define SET_SRAM_ECC_ON(VAR) VAR |= EF_AMDGPU_SRAM_ECC_V3
#define SET_SRAM_ECC_ANY(VAR) SET_SRAM_ECC_ON (VAR)
#define SET_SRAM_ECC_OFF(VAR) VAR &= ~EF_AMDGPU_SRAM_ECC_V3
+#define SET_SRAM_ECC_UNSUPPORTED(VAR) SET_SRAM_ECC_OFF (VAR)
#define TEST_SRAM_ECC_ANY(VAR) 0 /* Not supported. */
#define TEST_SRAM_ECC_ON(VAR) (VAR & EF_AMDGPU_SRAM_ECC_V3)
#endif
@@ -94,6 +97,9 @@
| EF_AMDGPU_FEATURE_SRAMECC_ANY_V4)
#define SET_SRAM_ECC_OFF(VAR) VAR = ((VAR & ~EF_AMDGPU_FEATURE_SRAMECC_V4) \
| EF_AMDGPU_FEATURE_SRAMECC_OFF_V4)
+#define SET_SRAM_ECC_UNSUPPORTED(VAR) \
+ VAR = ((VAR & ~EF_AMDGPU_FEATURE_SRAMECC_V4) \
+ | EF_AMDGPU_FEATURE_SRAMECC_UNSUPPORTED_V4)
#define TEST_SRAM_ECC_ANY(VAR) ((VAR & EF_AMDGPU_FEATURE_SRAMECC_V4) \
== EF_AMDGPU_FEATURE_SRAMECC_ANY_V4)
#define TEST_SRAM_ECC_ON(VAR) ((VAR & EF_AMDGPU_FEATURE_SRAMECC_V4) \
@@ -346,12 +352,27 @@ copy_early_debug_info (const char *infile, const char
*outfile)
/* We only support host relocations of x86_64, for now. */
gcc_assert (ehdr.e_machine == EM_X86_64);
+ /* Fiji devices use HSACOv3 regardless of the assembler. */
+ uint32_t elf_flags_actual = (elf_arch == EF_AMDGPU_MACH_AMDGCN_GFX803
+ ? 0 : elf_flags);
+ /* GFX900 devices don't support the sramecc attribute even if
+ a buggy assembler thinks it does. This must match gcn-hsa.h */
+ if (elf_arch == EF_AMDGPU_MACH_AMDGCN_GFX900)
+ SET_SRAM_ECC_UNSUPPORTED (elf_flags_actual);
+
/* Patch the correct elf architecture flag into the file. */
ehdr.e_ident[7] = ELFOSABI_AMDGPU_HSA;
- ehdr.e_ident[8] = ELFABIVERSION_AMDGPU_HSA;
+#ifdef HAVE_GCN_ASM_V3_SYNTAX
+ ehdr.e_ident[8] = ELFABIVERSION_AMDGPU_HSA_V3;
+#endif
+#ifdef HAVE_GCN_ASM_V4_SYNTAX
+ ehdr.e_ident[8] = (elf_arch == EF_AMDGPU_MACH_AMDGCN_GFX803
+ ? ELFABIVERSION_AMDGPU_HSA_V3
+ : ELFABIVERSION_AMDGPU_HSA_V4);
+#endif
ehdr.e_type = ET_REL;
ehdr.e_machine = EM_AMDGPU;
- ehdr.e_flags = elf_arch | elf_flags;
+ ehdr.e_flags = elf_arch | elf_flags_actual;
/* Load the section headers so we can walk them later. */
Elf64_Shdr *sections = (Elf64_Shdr *)xmalloc (sizeof (Elf64_Shdr)
@@ -987,13 +1008,6 @@ main (int argc, char **argv)
obstack_ptr_grow (&cc_argv_obstack, "-xlto");
if (fopenmp)
obstack_ptr_grow (&cc_argv_obstack, "-mgomp");
- obstack_ptr_grow (&cc_argv_obstack,
- (TEST_XNACK (elf_flags)
- ? "-mxnack" : "-mno-xnack"));
- obstack_ptr_grow (&cc_argv_obstack,
- (TEST_SRAM_ECC_ON (elf_flags) ? "-msram-ecc=on"
- : TEST_SRAM_ECC_ANY (elf_flags) ? "-msram-ecc=any"
- : "-msram-ecc=off"));
for (int ix = 1; ix != argc; ix++)
{