[Mesa-dev] [PATCH 7/7] radeonsi: optionally run the LLVM IR verifier pass

Nicolai Hähnle nhaehnle at gmail.com
Thu Sep 29 13:15:57 UTC 2016


From: Nicolai Hähnle <nicolai.haehnle at amd.com>

This is enabled automatically if shader printing is enabled, or separately
by R600_DEBUG=checkir. Catch mal-formed IR before it crashes in a later
pass.
---
 src/gallium/drivers/radeon/r600_pipe_common.c      |  7 ++++++
 src/gallium/drivers/radeon/r600_pipe_common.h      |  3 +++
 src/gallium/drivers/radeon/radeon_llvm.h           |  3 ++-
 .../drivers/radeon/radeon_setup_tgsi_llvm.c        |  6 ++++-
 src/gallium/drivers/radeonsi/si_shader.c           | 28 ++++++++++++++++------
 5 files changed, 38 insertions(+), 9 deletions(-)

diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index 5b1ce04..e7bf7f2 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -632,20 +632,21 @@ static const struct debug_named_value common_debug_options[] = {
 	{ "vs", DBG_VS, "Print vertex shaders" },
 	{ "gs", DBG_GS, "Print geometry shaders" },
 	{ "ps", DBG_PS, "Print pixel shaders" },
 	{ "cs", DBG_CS, "Print compute shaders" },
 	{ "tcs", DBG_TCS, "Print tessellation control shaders" },
 	{ "tes", DBG_TES, "Print tessellation evaluation shaders" },
 	{ "noir", DBG_NO_IR, "Don't print the LLVM IR"},
 	{ "notgsi", DBG_NO_TGSI, "Don't print the TGSI"},
 	{ "noasm", DBG_NO_ASM, "Don't print disassembled shaders"},
 	{ "preoptir", DBG_PREOPT_IR, "Print the LLVM IR before initial optimizations" },
+	{ "checkir", DBG_CHECK_IR, "Enable additional sanity checks on shader IR" },
 
 	{ "testdma", DBG_TEST_DMA, "Invoke SDMA tests and exit." },
 
 	/* features */
 	{ "nodma", DBG_NO_ASYNC_DMA, "Disable asynchronous DMA" },
 	{ "nohyperz", DBG_NO_HYPERZ, "Disable Hyper-Z" },
 	/* GL uses the word INVALIDATE, gallium uses the word DISCARD */
 	{ "noinvalrange", DBG_NO_DISCARD_RANGE, "Disable handling of INVALIDATE_RANGE map flags" },
 	{ "no2d", DBG_NO_2D_TILING, "Disable 2D tiling" },
 	{ "notiling", DBG_NO_TILING, "Disable tiling" },
@@ -1276,20 +1277,26 @@ bool r600_can_dump_shader(struct r600_common_screen *rscreen,
 		return (rscreen->debug_flags & DBG_GS) != 0;
 	case PIPE_SHADER_FRAGMENT:
 		return (rscreen->debug_flags & DBG_PS) != 0;
 	case PIPE_SHADER_COMPUTE:
 		return (rscreen->debug_flags & DBG_CS) != 0;
 	default:
 		return false;
 	}
 }
 
+bool r600_extra_shader_checks(struct r600_common_screen *rscreen, unsigned processor)
+{
+	return (rscreen->debug_flags & DBG_CHECK_IR) ||
+	       r600_can_dump_shader(rscreen, processor);
+}
+
 void r600_screen_clear_buffer(struct r600_common_screen *rscreen, struct pipe_resource *dst,
 			      uint64_t offset, uint64_t size, unsigned value,
 			      enum r600_coherency coher)
 {
 	struct r600_common_context *rctx = (struct r600_common_context*)rscreen->aux_context;
 
 	pipe_mutex_lock(rscreen->aux_context_lock);
 	rctx->clear_buffer(&rctx->b, dst, offset, size, value, coher);
 	rscreen->aux_context->flush(rscreen->aux_context, NULL, 0);
 	pipe_mutex_unlock(rscreen->aux_context_lock);
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h b/src/gallium/drivers/radeon/r600_pipe_common.h
index f23f1c4..c836ab1 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -71,20 +71,21 @@
 #define DBG_VS			(1 << 6)
 #define DBG_GS			(1 << 7)
 #define DBG_PS			(1 << 8)
 #define DBG_CS			(1 << 9)
 #define DBG_TCS			(1 << 10)
 #define DBG_TES			(1 << 11)
 #define DBG_NO_IR		(1 << 12)
 #define DBG_NO_TGSI		(1 << 13)
 #define DBG_NO_ASM		(1 << 14)
 #define DBG_PREOPT_IR		(1 << 15)
+#define DBG_CHECK_IR		(1 << 16)
 /* gaps */
 #define DBG_TEST_DMA		(1 << 20)
 /* Bits 21-31 are reserved for the r600g driver. */
 /* features */
 #define DBG_NO_ASYNC_DMA	(1llu << 32)
 #define DBG_NO_HYPERZ		(1llu << 33)
 #define DBG_NO_DISCARD_RANGE	(1llu << 34)
 #define DBG_NO_2D_TILING	(1llu << 35)
 #define DBG_NO_TILING		(1llu << 36)
 #define DBG_SWITCH_ON_EOP	(1llu << 37)
@@ -714,20 +715,22 @@ bool r600_common_screen_init(struct r600_common_screen *rscreen,
 void r600_destroy_common_screen(struct r600_common_screen *rscreen);
 void r600_preflush_suspend_features(struct r600_common_context *ctx);
 void r600_postflush_resume_features(struct r600_common_context *ctx);
 bool r600_common_context_init(struct r600_common_context *rctx,
 			      struct r600_common_screen *rscreen,
 			      unsigned context_flags);
 void r600_common_context_cleanup(struct r600_common_context *rctx);
 void r600_context_add_resource_size(struct pipe_context *ctx, struct pipe_resource *r);
 bool r600_can_dump_shader(struct r600_common_screen *rscreen,
 			  unsigned processor);
+bool r600_extra_shader_checks(struct r600_common_screen *rscreen,
+			      unsigned processor);
 void r600_screen_clear_buffer(struct r600_common_screen *rscreen, struct pipe_resource *dst,
 			      uint64_t offset, uint64_t size, unsigned value,
 			      enum r600_coherency coher);
 struct pipe_resource *r600_resource_create_common(struct pipe_screen *screen,
 						  const struct pipe_resource *templ);
 const char *r600_get_llvm_processor_name(enum radeon_family family);
 void r600_need_dma_space(struct r600_common_context *ctx, unsigned num_dw,
 			 struct r600_resource *dst, struct r600_resource *src);
 void r600_dma_emit_wait_idle(struct r600_common_context *rctx);
 void radeon_save_cs(struct radeon_winsys *ws, struct radeon_winsys_cs *cs,
diff --git a/src/gallium/drivers/radeon/radeon_llvm.h b/src/gallium/drivers/radeon/radeon_llvm.h
index 2f9572a..6010254 100644
--- a/src/gallium/drivers/radeon/radeon_llvm.h
+++ b/src/gallium/drivers/radeon/radeon_llvm.h
@@ -121,21 +121,22 @@ void radeon_llvm_context_init(struct radeon_llvm_context *ctx,
 			      const struct tgsi_token *tokens);
 
 void radeon_llvm_create_func(struct radeon_llvm_context *ctx,
 			     LLVMTypeRef *return_types, unsigned num_return_elems,
 			     LLVMTypeRef *ParamTypes, unsigned ParamCount);
 
 void radeon_llvm_dispose(struct radeon_llvm_context *ctx);
 
 unsigned radeon_llvm_reg_index_soa(unsigned index, unsigned chan);
 
-void radeon_llvm_finalize_module(struct radeon_llvm_context *ctx);
+void radeon_llvm_finalize_module(struct radeon_llvm_context *ctx,
+				 bool run_verifier);
 
 void build_tgsi_intrinsic_nomem(const struct lp_build_tgsi_action *action,
 				struct lp_build_tgsi_context *bld_base,
 				struct lp_build_emit_data *emit_data);
 
 LLVMValueRef radeon_llvm_emit_fetch_64bit(struct lp_build_tgsi_context *bld_base,
 					  enum tgsi_opcode_type type,
 					  LLVMValueRef ptr,
 					  LLVMValueRef ptr2);
 
diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
index 80e9707..8e364c9 100644
--- a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
+++ b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
@@ -2107,33 +2107,37 @@ void radeon_llvm_create_func(struct radeon_llvm_context *ctx,
 
 	/* Setup the function */
 	ctx->return_type = ret_type;
 	main_fn_type = LLVMFunctionType(ret_type, ParamTypes, ParamCount, 0);
 	ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, "main", main_fn_type);
 	main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
 			ctx->main_fn, "main_body");
 	LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
 }
 
-void radeon_llvm_finalize_module(struct radeon_llvm_context *ctx)
+void radeon_llvm_finalize_module(struct radeon_llvm_context *ctx,
+				 bool run_verifier)
 {
 	struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
 	const char *triple = LLVMGetTarget(gallivm->module);
 	LLVMTargetLibraryInfoRef target_library_info;
 
 	/* Create the pass manager */
 	gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(
 							gallivm->module);
 
 	target_library_info = gallivm_create_target_library_info(triple);
 	LLVMAddTargetLibraryInfo(target_library_info, gallivm->passmgr);
 
+	if (run_verifier)
+		LLVMAddVerifierPass(gallivm->passmgr);
+
 	/* This pass should eliminate all the load and store instructions */
 	LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 
 	/* Add some optimization passes */
 	LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 	LLVMAddLICMPass(gallivm->passmgr);
 	LLVMAddAggressiveDCEPass(gallivm->passmgr);
 	LLVMAddCFGSimplificationPass(gallivm->passmgr);
 	LLVMAddInstructionCombiningPass(gallivm->passmgr);
 
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 3ccff7a..2cabff6 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -6424,21 +6424,23 @@ static int si_generate_gs_copy_shader(struct si_screen *sscreen,
 
 	si_llvm_export_vs(bld_base, outputs, gsinfo->num_outputs);
 
 	LLVMBuildRetVoid(gallivm->builder);
 
 	/* Dump LLVM IR before any optimization passes */
 	if (sscreen->b.debug_flags & DBG_PREOPT_IR &&
 	    r600_can_dump_shader(&sscreen->b, PIPE_SHADER_GEOMETRY))
 		LLVMDumpModule(bld_base->base.gallivm->module);
 
-	radeon_llvm_finalize_module(&ctx->radeon_bld);
+	radeon_llvm_finalize_module(
+		&ctx->radeon_bld,
+		r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_GEOMETRY));
 
 	r = si_compile_llvm(sscreen, &ctx->shader->binary,
 			    &ctx->shader->config, ctx->tm,
 			    bld_base->base.gallivm->module,
 			    debug, PIPE_SHADER_GEOMETRY,
 			    "GS Copy Shader");
 	if (!r) {
 		if (r600_can_dump_shader(&sscreen->b, PIPE_SHADER_GEOMETRY))
 			fprintf(stderr, "GS Copy Shader:\n");
 		si_shader_dump(sscreen, ctx->shader, debug,
@@ -6708,21 +6710,23 @@ int si_compile_tgsi_shader(struct si_screen *sscreen,
 	}
 
 	si_llvm_build_ret(&ctx, ctx.return_value);
 	mod = bld_base->base.gallivm->module;
 
 	/* Dump LLVM IR before any optimization passes */
 	if (sscreen->b.debug_flags & DBG_PREOPT_IR &&
 	    r600_can_dump_shader(&sscreen->b, ctx.type))
 		LLVMDumpModule(mod);
 
-	radeon_llvm_finalize_module(&ctx.radeon_bld);
+	radeon_llvm_finalize_module(
+		&ctx.radeon_bld,
+		r600_extra_shader_checks(&sscreen->b, ctx.type));
 
 	r = si_compile_llvm(sscreen, &shader->binary, &shader->config, tm,
 			    mod, debug, ctx.type, "TGSI shader");
 	if (r) {
 		fprintf(stderr, "LLVM failed to compile shader\n");
 		goto out;
 	}
 
 	radeon_llvm_dispose(&ctx.radeon_bld);
 
@@ -6969,21 +6973,23 @@ static bool si_compile_vs_prolog(struct si_screen *sscreen,
 					     LLVMGetParam(func, SI_SGPR_BASE_VERTEX), "");
 		}
 
 		index = LLVMBuildBitCast(gallivm->builder, index, ctx.f32, "");
 		ret = LLVMBuildInsertValue(gallivm->builder, ret, index,
 					   num_params++, "");
 	}
 
 	/* Compile. */
 	si_llvm_build_ret(&ctx, ret);
-	radeon_llvm_finalize_module(&ctx.radeon_bld);
+	radeon_llvm_finalize_module(
+		&ctx.radeon_bld,
+		r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_VERTEX));
 
 	if (si_compile_llvm(sscreen, &out->binary, &out->config, tm,
 			    gallivm->module, debug, ctx.type,
 			    "Vertex Shader Prolog"))
 		status = false;
 
 	radeon_llvm_dispose(&ctx.radeon_bld);
 	return status;
 }
 
@@ -7041,21 +7047,23 @@ static bool si_compile_vs_epilog(struct si_screen *sscreen,
 		args[7] = uint->undef; /* Z */
 		args[8] = uint->undef; /* W */
 
 		lp_build_intrinsic(base->gallivm->builder, "llvm.SI.export",
 				   LLVMVoidTypeInContext(base->gallivm->context),
 				   args, 9, 0);
 	}
 
 	/* Compile. */
 	LLVMBuildRetVoid(gallivm->builder);
-	radeon_llvm_finalize_module(&ctx.radeon_bld);
+	radeon_llvm_finalize_module(
+		&ctx.radeon_bld,
+		r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_VERTEX));
 
 	if (si_compile_llvm(sscreen, &out->binary, &out->config, tm,
 			    gallivm->module, debug, ctx.type,
 			    "Vertex Shader Epilog"))
 		status = false;
 
 	radeon_llvm_dispose(&ctx.radeon_bld);
 	return status;
 }
 
@@ -7194,21 +7202,23 @@ static bool si_compile_tcs_epilog(struct si_screen *sscreen,
 	declare_tess_lds(&ctx);
 	func = ctx.radeon_bld.main_fn;
 
 	si_write_tess_factors(bld_base,
 			      LLVMGetParam(func, last_sgpr + 1),
 			      LLVMGetParam(func, last_sgpr + 2),
 			      LLVMGetParam(func, last_sgpr + 3));
 
 	/* Compile. */
 	LLVMBuildRetVoid(gallivm->builder);
-	radeon_llvm_finalize_module(&ctx.radeon_bld);
+	radeon_llvm_finalize_module(
+		&ctx.radeon_bld,
+		r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_TESS_CTRL));
 
 	if (si_compile_llvm(sscreen, &out->binary, &out->config, tm,
 			    gallivm->module, debug, ctx.type,
 			    "Tessellation Control Shader Epilog"))
 		status = false;
 
 	radeon_llvm_dispose(&ctx.radeon_bld);
 	return status;
 }
 
@@ -7478,21 +7488,23 @@ static bool si_compile_ps_prolog(struct si_screen *sscreen,
 	}
 
 	/* Tell LLVM to insert WQM instruction sequence when needed. */
 	if (key->ps_prolog.wqm) {
 		LLVMAddTargetDependentFunctionAttr(func,
 						   "amdgpu-ps-wqm-outputs", "");
 	}
 
 	/* Compile. */
 	si_llvm_build_ret(&ctx, ret);
-	radeon_llvm_finalize_module(&ctx.radeon_bld);
+	radeon_llvm_finalize_module(
+		&ctx.radeon_bld,
+		r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_FRAGMENT));
 
 	if (si_compile_llvm(sscreen, &out->binary, &out->config, tm,
 			    gallivm->module, debug, ctx.type,
 			    "Fragment Shader Prolog"))
 		status = false;
 
 	radeon_llvm_dispose(&ctx.radeon_bld);
 	return status;
 }
 
@@ -7598,21 +7610,23 @@ static bool si_compile_ps_epilog(struct si_screen *sscreen,
 	if (depth || stencil || samplemask)
 		si_export_mrt_z(bld_base, depth, stencil, samplemask, &exp);
 	else if (last_color_export == -1)
 		si_export_null(bld_base);
 
 	if (exp.num)
 		si_emit_ps_exports(&ctx, &exp);
 
 	/* Compile. */
 	LLVMBuildRetVoid(gallivm->builder);
-	radeon_llvm_finalize_module(&ctx.radeon_bld);
+	radeon_llvm_finalize_module(
+		&ctx.radeon_bld,
+		r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_FRAGMENT));
 
 	if (si_compile_llvm(sscreen, &out->binary, &out->config, tm,
 			    gallivm->module, debug, ctx.type,
 			    "Fragment Shader Epilog"))
 		status = false;
 
 	radeon_llvm_dispose(&ctx.radeon_bld);
 	return status;
 }
 
-- 
2.7.4



More information about the mesa-dev mailing list