#include <llvm-c/Support.h>
#include <llvm-c/Transforms/IPO.h>
#include <llvm-c/Transforms/Scalar.h>
-#if HAVE_LLVM >= 0x0700
#include <llvm-c/Transforms/Utils.h>
-#endif
#include "c11/threads.h"
+#include "gallivm/lp_bld_misc.h"
#include "util/u_math.h"
#include <assert.h>
* https://reviews.llvm.org/D26348
*
* "mesa" is the prefix for error messages.
+ *
+ * -global-isel-abort=2 is a no-op unless global isel has been enabled.
+ * This option tells the backend to fall-back to SelectionDAG and print
+ * a diagnostic message if global isel fails.
*/
- const char *argv[2] = { "mesa", "-simplifycfg-sink-common=false" };
- LLVMParseCommandLineOptions(2, argv, NULL);
+ const char *argv[3] = { "mesa", "-simplifycfg-sink-common=false", "-global-isel-abort=2" };
+ LLVMParseCommandLineOptions(3, argv, NULL);
}
static once_flag ac_init_llvm_target_once_flag = ONCE_FLAG_INIT;
call_once(&ac_init_llvm_target_once_flag, ac_init_llvm_target);
}
-LLVMTargetRef ac_get_llvm_target(const char *triple)
+static LLVMTargetRef ac_get_llvm_target(const char *triple)
{
LLVMTargetRef target = NULL;
char *err_message = NULL;
case CHIP_RAVEN:
return "gfx902";
case CHIP_VEGA12:
- return HAVE_LLVM >= 0x0700 ? "gfx904" : "gfx902";
+ return "gfx904";
+ case CHIP_VEGA20:
+ return "gfx906";
+ case CHIP_RAVEN2:
+ return HAVE_LLVM >= 0x0800 ? "gfx909" : "gfx902";
default:
return "";
}
}
-LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family,
- enum ac_target_machine_options tm_options,
- const char **out_triple)
+static LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family,
+ enum ac_target_machine_options tm_options,
+ LLVMCodeGenOptLevel level,
+ const char **out_triple)
{
assert(family >= CHIP_TAHITI);
char features[256];
const char *triple = (tm_options & AC_TM_SUPPORTS_SPILL) ? "amdgcn-mesa-mesa3d" : "amdgcn--";
LLVMTargetRef target = ac_get_llvm_target(triple);
- bool barrier_does_waitcnt = true; /* TODO: not for Vega20 */
snprintf(features, sizeof(features),
- "+DumpCode,+vgpr-spilling,-fp32-denormals,+fp64-denormals%s%s%s%s%s",
+ "+DumpCode,-fp32-denormals,+fp64-denormals%s%s%s%s%s%s",
+ HAVE_LLVM >= 0x0800 ? "" : ",+vgpr-spilling",
tm_options & AC_TM_SISCHED ? ",+si-scheduler" : "",
tm_options & AC_TM_FORCE_ENABLE_XNACK ? ",+xnack" : "",
tm_options & AC_TM_FORCE_DISABLE_XNACK ? ",-xnack" : "",
tm_options & AC_TM_PROMOTE_ALLOCA_TO_SCRATCH ? ",-promote-alloca" : "",
- barrier_does_waitcnt ? ",+auto-waitcnt-before-barrier" : "");
-
+ tm_options & AC_TM_NO_LOAD_STORE_OPT ? ",-load-store-opt" : "");
+
LLVMTargetMachineRef tm = LLVMCreateTargetMachine(
target,
triple,
ac_get_llvm_processor_name(family),
features,
- LLVMCodeGenLevelDefault,
+ level,
LLVMRelocDefault,
LLVMCodeModelDefault);
if (out_triple)
*out_triple = triple;
+ if (tm_options & AC_TM_ENABLE_GLOBAL_ISEL)
+ ac_enable_global_isel(tm);
return tm;
}
-LLVMPassManagerRef ac_create_passmgr(LLVMTargetLibraryInfoRef target_library_info,
- bool check_ir)
+static LLVMPassManagerRef ac_create_passmgr(LLVMTargetLibraryInfoRef target_library_info,
+ bool check_ir)
{
LLVMPassManagerRef passmgr = LLVMCreatePassManager();
if (!passmgr)
if (check_ir)
LLVMAddVerifierPass(passmgr);
LLVMAddAlwaysInlinerPass(passmgr);
+ /* Normally, the pass manager runs all passes on one function before
+ * moving onto another. Adding a barrier no-op pass forces the pass
+ * manager to run the inliner on all functions first, which makes sure
+ * that the following passes are only run on the remaining non-inline
+ * function, so it removes useless work done on dead inline functions.
+ */
+ ac_llvm_add_barrier_noop_pass(passmgr);
/* This pass should eliminate all the load and store instructions. */
LLVMAddPromoteMemoryToRegisterPass(passmgr);
LLVMAddScalarReplAggregatesPass(passmgr);
bool
ac_init_llvm_compiler(struct ac_llvm_compiler *compiler,
- bool okay_to_leak_target_library_info,
enum radeon_family family,
enum ac_target_machine_options tm_options)
{
const char *triple;
memset(compiler, 0, sizeof(*compiler));
- compiler->tm = ac_create_target_machine(family,
- tm_options, &triple);
+ compiler->tm = ac_create_target_machine(family, tm_options,
+ LLVMCodeGenLevelDefault,
+ &triple);
if (!compiler->tm)
return false;
- if (okay_to_leak_target_library_info || (HAVE_LLVM >= 0x0700)) {
- compiler->target_library_info =
- ac_create_target_library_info(triple);
- if (!compiler->target_library_info)
+ if (tm_options & AC_TM_CREATE_LOW_OPT) {
+ compiler->low_opt_tm =
+ ac_create_target_machine(family, tm_options,
+ LLVMCodeGenLevelLess, NULL);
+ if (!compiler->low_opt_tm)
goto fail;
}
+ compiler->target_library_info =
+ ac_create_target_library_info(triple);
+ if (!compiler->target_library_info)
+ goto fail;
+
compiler->passmgr = ac_create_passmgr(compiler->target_library_info,
tm_options & AC_TM_CHECK_IR);
if (!compiler->passmgr)
{
if (compiler->passmgr)
LLVMDisposePassManager(compiler->passmgr);
-#if HAVE_LLVM >= 0x0700
- /* This crashes on LLVM 5.0 and 6.0 and Ubuntu 18.04, so leak it there. */
if (compiler->target_library_info)
ac_dispose_target_library_info(compiler->target_library_info);
-#endif
+ if (compiler->low_opt_tm)
+ LLVMDisposeTargetMachine(compiler->low_opt_tm);
if (compiler->tm)
LLVMDisposeTargetMachine(compiler->tm);
}
projects / mesa.git / blobdiff