ac,radeonsi: reduce optimizations for complex compute shaders on older APUs (v2)

To make dEQP-GLES31.functional.ssbo.layout.random.all_shared_buffer.23
finish sooner on the older CPUs. (otherwise it gets killed and we fail
the test)

Acked-by: Dave Airlie <[email protected]>

2 files changed, 25 insertions, 4 deletions

diff --git a/src/amd/common/ac_llvm_util.c b/src/amd/common/ac_llvm_util.c
index 678bc34e6f8..10e1ca99d41 100644
--- a/src/amd/common/ac_llvm_util.c
+++ b/src/amd/common/ac_llvm_util.c
@@ -142,6 +142,7 @@ const char *ac_get_llvm_processor_name(enum radeon_family family)
 
 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);
@@ -163,7 +164,7 @@ static LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family,
 	                             triple,
 	                             ac_get_llvm_processor_name(family),
 				     features,
-	                             LLVMCodeGenLevelDefault,
+	                             level,
 	                             LLVMRelocDefault,
 	                             LLVMCodeModelDefault);
 
@@ -308,11 +309,20 @@ ac_init_llvm_compiler(struct ac_llvm_compiler *compiler,
 	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 (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;
+	}
+
 	if (okay_to_leak_target_library_info || (HAVE_LLVM >= 0x0700)) {
 		compiler->target_library_info =
 			ac_create_target_library_info(triple);
@@ -341,6 +351,8 @@ ac_destroy_llvm_compiler(struct ac_llvm_compiler *compiler)
 	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);
 }
diff --git a/src/amd/common/ac_llvm_util.h b/src/amd/common/ac_llvm_util.h
index d4dea4dfde6..eaf5f21876b 100644
--- a/src/amd/common/ac_llvm_util.h
+++ b/src/amd/common/ac_llvm_util.h
@@ -64,6 +64,7 @@ enum ac_target_machine_options {
 	AC_TM_PROMOTE_ALLOCA_TO_SCRATCH = (1 << 4),
 	AC_TM_CHECK_IR = (1 << 5),
 	AC_TM_ENABLE_GLOBAL_ISEL = (1 << 6),
+	AC_TM_CREATE_LOW_OPT = (1 << 7),
 };
 
 enum ac_float_mode {
@@ -74,10 +75,18 @@ enum ac_float_mode {
 
 /* Per-thread persistent LLVM objects. */
 struct ac_llvm_compiler {
-	LLVMTargetMachineRef		tm;
 	LLVMTargetLibraryInfoRef	target_library_info;
 	LLVMPassManagerRef		passmgr;
+
+	/* Default compiler. */
+	LLVMTargetMachineRef		tm;
 	struct ac_compiler_passes	*passes;
+
+	/* Optional compiler for faster compilation with fewer optimizations.
+	 * LLVM modules can be created with "tm" too. There is no difference.
+	 */
+	LLVMTargetMachineRef		low_opt_tm; /* uses -O1 instead of -O2 */
+	struct ac_compiler_passes	*low_opt_passes;
 };
 
 const char *ac_get_llvm_processor_name(enum radeon_family family);