ac: replace ac_build_kill with ac_build_kill_if_false

This will be a new LLVM intrinsic and will also work nicely with
llvm.amdgcn.wqm.vote.

Reviewed-by: Nicolai Hähnle <[email protected]>

5 files changed, 49 insertions, 78 deletions

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 4d8f42d572c..5bd04663216 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -1405,20 +1405,13 @@ LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 				  AC_FUNC_ATTR_LEGACY);
 }
 
-/**
- * KILL, AKA discard in GLSL.
- *
- * \param value  kill if value < 0.0 or value == NULL.
- */
-void ac_build_kill(struct ac_llvm_context *ctx, LLVMValueRef value)
+void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
 {
-	if (value) {
-		ac_build_intrinsic(ctx, "llvm.AMDGPU.kill", ctx->voidt,
-				   &value, 1, AC_FUNC_ATTR_LEGACY);
-	} else {
-		ac_build_intrinsic(ctx, "llvm.AMDGPU.kilp", ctx->voidt,
-				   NULL, 0, AC_FUNC_ATTR_LEGACY);
-	}
+	LLVMValueRef value = LLVMBuildSelect(ctx->builder, i1,
+					     LLVMConstReal(ctx->f32, 1),
+					     LLVMConstReal(ctx->f32, -1), "");
+	ac_build_intrinsic(ctx, "llvm.AMDGPU.kill", ctx->voidt,
+			   &value, 1, AC_FUNC_ATTR_LEGACY);
 }
 
 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index aa2a2899ab4..8f7a9abb76c 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -265,7 +265,7 @@ LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 				   struct ac_image_args *a);
 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 				    LLVMValueRef args[2]);
-void ac_build_kill(struct ac_llvm_context *ctx, LLVMValueRef value);
+void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 			  LLVMValueRef offset, LLVMValueRef width,
 			  bool is_signed);
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 2e50e50b129..4f7e8449725 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -3779,14 +3779,10 @@ static void emit_discard_if(struct ac_nir_context *ctx,
 {
 	LLVMValueRef cond;
 
-	cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE,
+	cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
 			     get_src(ctx, instr->src[0]),
 			     ctx->ac.i32_0, "");
-
-	cond = LLVMBuildSelect(ctx->ac.builder, cond,
-			       LLVMConstReal(ctx->ac.f32, -1.0f),
-			       ctx->ac.f32_0, "");
-	ac_build_kill(&ctx->ac, cond);
+	ac_build_kill_if_false(&ctx->ac, cond);
 }
 
 static LLVMValueRef
@@ -4021,7 +4017,7 @@ visit_emit_vertex(struct nir_to_llvm_context *ctx,
 		  const nir_intrinsic_instr *instr)
 {
 	LLVMValueRef gs_next_vertex;
-	LLVMValueRef can_emit, kill;
+	LLVMValueRef can_emit;
 	int idx;
 
 	assert(instr->const_index[0] == 0);
@@ -4037,11 +4033,7 @@ visit_emit_vertex(struct nir_to_llvm_context *ctx,
 	 */
 	can_emit = LLVMBuildICmp(ctx->builder, LLVMIntULT, gs_next_vertex,
 				 LLVMConstInt(ctx->i32, ctx->gs_max_out_vertices, false), "");
-
-	kill = LLVMBuildSelect(ctx->builder, can_emit,
-			       LLVMConstReal(ctx->f32, 1.0f),
-			       LLVMConstReal(ctx->f32, -1.0f), "");
-	ac_build_kill(&ctx->ac, kill);
+	ac_build_kill_if_false(&ctx->ac, can_emit);
 
 	/* loop num outputs */
 	idx = 0;
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index c3fe13deeaa..8abacac8c8d 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -2272,22 +2272,24 @@ static void si_alpha_test(struct lp_build_tgsi_context *bld_base,
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 
 	if (ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_NEVER) {
+		static LLVMRealPredicate cond_map[PIPE_FUNC_ALWAYS + 1] = {
+			[PIPE_FUNC_LESS] = LLVMRealOLT,
+			[PIPE_FUNC_EQUAL] = LLVMRealOEQ,
+			[PIPE_FUNC_LEQUAL] = LLVMRealOLE,
+			[PIPE_FUNC_GREATER] = LLVMRealOGT,
+			[PIPE_FUNC_NOTEQUAL] = LLVMRealONE,
+			[PIPE_FUNC_GEQUAL] = LLVMRealOGE,
+		};
+		LLVMRealPredicate cond = cond_map[ctx->shader->key.part.ps.epilog.alpha_func];
+		assert(cond);
+
 		LLVMValueRef alpha_ref = LLVMGetParam(ctx->main_fn,
 				SI_PARAM_ALPHA_REF);
-
 		LLVMValueRef alpha_pass =
-			lp_build_cmp(&bld_base->base,
-				     ctx->shader->key.part.ps.epilog.alpha_func,
-				     alpha, alpha_ref);
-		LLVMValueRef arg =
-			lp_build_select(&bld_base->base,
-					alpha_pass,
-					LLVMConstReal(ctx->f32, 1.0f),
-					LLVMConstReal(ctx->f32, -1.0f));
-
-		ac_build_kill(&ctx->ac, arg);
+			LLVMBuildFCmp(ctx->ac.builder, cond, alpha, alpha_ref, "");
+		ac_build_kill_if_false(&ctx->ac, alpha_pass);
 	} else {
-		ac_build_kill(&ctx->ac, NULL);
+		ac_build_kill_if_false(&ctx->ac, LLVMConstInt(ctx->i1, 0, 0));
 	}
 }
 
@@ -3573,7 +3575,7 @@ static void si_llvm_return_fs_outputs(struct ac_shader_abi *abi,
 	LLVMValueRef ret;
 
 	if (ctx->postponed_kill)
-		ac_build_kill(&ctx->ac, LLVMBuildLoad(builder, ctx->postponed_kill, ""));
+		ac_build_kill_if_false(&ctx->ac, LLVMBuildLoad(builder, ctx->postponed_kill, ""));
 
 	/* Read the output values. */
 	for (i = 0; i < info->num_outputs; i++) {
@@ -4056,7 +4058,7 @@ static void si_llvm_emit_vertex(
 	LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
 					    ctx->param_gs2vs_offset);
 	LLVMValueRef gs_next_vertex;
-	LLVMValueRef can_emit, kill;
+	LLVMValueRef can_emit;
 	unsigned chan, offset;
 	int i;
 	unsigned stream;
@@ -4082,11 +4084,7 @@ static void si_llvm_emit_vertex(
 
 	bool use_kill = !info->writes_memory;
 	if (use_kill) {
-		kill = lp_build_select(&bld_base->base, can_emit,
-				       LLVMConstReal(ctx->f32, 1.0f),
-				       LLVMConstReal(ctx->f32, -1.0f));
-
-		ac_build_kill(&ctx->ac, kill);
+		ac_build_kill_if_false(&ctx->ac, can_emit);
 	} else {
 		lp_build_if(&if_state, &ctx->gallivm, can_emit);
 	}
@@ -4881,11 +4879,7 @@ static void si_llvm_emit_polygon_stipple(struct si_shader_context *ctx,
 	row = ac_to_integer(&ctx->ac, row);
 	bit = LLVMBuildLShr(builder, row, address[0], "");
 	bit = LLVMBuildTrunc(builder, bit, ctx->i1, "");
-
-	/* The intrinsic kills the thread if arg < 0. */
-	bit = LLVMBuildSelect(builder, bit, LLVMConstReal(ctx->f32, 0),
-			      LLVMConstReal(ctx->f32, -1), "");
-	ac_build_kill(&ctx->ac, bit);
+	ac_build_kill_if_false(&ctx->ac, bit);
 }
 
 void si_shader_binary_read_config(struct ac_shader_binary *binary,
@@ -5854,8 +5848,10 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx,
 
 	if (ctx->type == PIPE_SHADER_FRAGMENT && sel->info.uses_kill &&
 	    ctx->screen->b.debug_flags & DBG(FS_CORRECT_DERIVS_AFTER_KILL)) {
-		/* This is initialized to 0.0 = not kill. */
-		ctx->postponed_kill = lp_build_alloca(&ctx->gallivm, ctx->f32, "");
+		ctx->postponed_kill = lp_build_alloca_undef(&ctx->gallivm, ctx->i1, "");
+		/* true = don't kill. */
+		LLVMBuildStore(ctx->ac.builder, LLVMConstInt(ctx->i1, 1, 0),
+			       ctx->postponed_kill);
 	}
 
 	if (sel->tokens) {
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
index fc705c376e5..ad7a42f7b84 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
@@ -39,20 +39,18 @@ static void kill_if_fetch_args(struct lp_build_tgsi_context *bld_base,
 
 	for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
 		LLVMValueRef value = lp_build_emit_fetch(bld_base, inst, 0, i);
-		conds[i] = LLVMBuildFCmp(builder, LLVMRealOLT, value,
+		conds[i] = LLVMBuildFCmp(builder, LLVMRealOGE, value,
 					ctx->ac.f32_0, "");
 	}
 
-	/* Or the conditions together */
+	/* And the conditions together */
 	for (i = TGSI_NUM_CHANNELS - 1; i > 0; i--) {
-		conds[i - 1] = LLVMBuildOr(builder, conds[i], conds[i - 1], "");
+		conds[i - 1] = LLVMBuildAnd(builder, conds[i], conds[i - 1], "");
 	}
 
 	emit_data->dst_type = ctx->voidt;
 	emit_data->arg_count = 1;
-	emit_data->args[0] = LLVMBuildSelect(builder, conds[0],
-					LLVMConstReal(ctx->f32, -1.0f),
-					ctx->ac.f32_0, "");
+	emit_data->args[0] = conds[0];
 }
 
 static void kil_emit(const struct lp_build_tgsi_action *action,
@@ -61,31 +59,23 @@ static void kil_emit(const struct lp_build_tgsi_action *action,
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	LLVMBuilderRef builder = ctx->ac.builder;
+	LLVMValueRef visible;
+
+	if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_KILL_IF) {
+		visible = emit_data->args[0];
+	} else {
+		assert(emit_data->inst->Instruction.Opcode == TGSI_OPCODE_KILL);
+		visible = LLVMConstInt(ctx->i1, false, 0);
+	}
 
 	if (ctx->postponed_kill) {
-		if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_KILL_IF) {
-			LLVMValueRef val;
-
-			/* Take the minimum kill value. This is the same as OR
-			 * between 2 kill values. If the value is negative,
-			 * the pixel will be killed.
-			 */
-			val = LLVMBuildLoad(builder, ctx->postponed_kill, "");
-			val = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_MIN,
-							val, emit_data->args[0]);
-			LLVMBuildStore(builder, val, ctx->postponed_kill);
-		} else {
-			LLVMBuildStore(builder,
-				       LLVMConstReal(ctx->f32, -1),
-				       ctx->postponed_kill);
-		}
+		LLVMValueRef mask = LLVMBuildLoad(builder, ctx->postponed_kill, "");
+		mask = LLVMBuildAnd(builder, mask, visible, "");
+		LLVMBuildStore(builder, mask, ctx->postponed_kill);
 		return;
 	}
 
-	if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_KILL_IF)
-		ac_build_kill(&ctx->ac, emit_data->args[0]);
-	else
-		ac_build_kill(&ctx->ac, NULL);
+	ac_build_kill_if_false(&ctx->ac, visible);
 }
 
 static void emit_icmp(const struct lp_build_tgsi_action *action,