ac: rework ac_build_waitcnt for gfx10

Acked-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

src/amd/common/ac_llvm_build.c

@@ -2882,13 +2882,49 @@ LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 			     LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
 }
 
-void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16)
+void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags)
 {
+	if (!wait_flags)
+		return;
+
+	unsigned lgkmcnt = 63;
+	unsigned expcnt = 7;
+	unsigned vmcnt = ctx->chip_class >= GFX9 ? 63 : 15;
+	unsigned vscnt = 63;
+
+	if (wait_flags & AC_WAIT_LGKM)
+		lgkmcnt = 0;
+	if (wait_flags & AC_WAIT_EXP)
+		expcnt = 0;
+	if (wait_flags & AC_WAIT_VLOAD)
+		vmcnt = 0;
+
+	if (wait_flags & AC_WAIT_VSTORE) {
+		if (ctx->chip_class >= GFX10)
+			vscnt = 0;
+		else
+			vmcnt = 0;
+	}
+
+	unsigned simm16 = (lgkmcnt << 8) |
+			  (expcnt << 4) |
+			  (vmcnt & 0xf) |
+			  ((vmcnt >> 4) << 14);
+
 	LLVMValueRef args[1] = {
 		LLVMConstInt(ctx->i32, simm16, false),
 	};
 	ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
 			   ctx->voidt, args, 1, 0);
+
+	/* TODO: add llvm.amdgcn.s.waitcnt.vscnt into LLVM: */
+	if (0 && ctx->chip_class >= GFX10 && vscnt == 0) {
+		LLVMValueRef args[1] = {
+			LLVMConstInt(ctx->i32, vscnt, false),
+		};
+		ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt.vscnt",
+				   ctx->voidt, args, 1, 0);
+	}
 }
 
 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,

src/amd/common/ac_llvm_build.h

@@ -43,11 +43,10 @@ enum {
 	AC_ADDR_SPACE_CONST_32BIT = 6, /* same as CONST, but the pointer type has 32 bits */
 };
 
-/* Combine these with & instead of |. */
+#define AC_WAIT_LGKM	(1 << 0) /* LDS, GDS, constant, message */
-#define NOOP_WAITCNT	0xcf7f
+#define AC_WAIT_EXP	(1 << 1) /* exports */
-#define LGKM_CNT	0xc07f
+#define AC_WAIT_VLOAD	(1 << 2) /* VMEM load/sample instructions */
-#define EXP_CNT		0xcf0f
+#define AC_WAIT_VSTORE	(1 << 3) /* VMEM store instructions */
-#define VM_CNT		0x0f70 /* On GFX9, vmcnt has 6 bits in [0:3] and [14:15] */
 
 struct ac_llvm_flow;
 
@@ -575,7 +574,7 @@ LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 			   LLVMValueRef s1, LLVMValueRef s2);
 
-void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16);
+void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags);
 
 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 			   unsigned bitsize);

src/amd/common/ac_nir_to_llvm.c

@@ -2741,26 +2741,26 @@ static LLVMValueRef visit_image_size(struct ac_nir_context *ctx,
 static void emit_membar(struct ac_llvm_context *ac,
 			const nir_intrinsic_instr *instr)
 {
-	unsigned waitcnt = NOOP_WAITCNT;
+	unsigned wait_flags = 0;
 
 	switch (instr->intrinsic) {
 	case nir_intrinsic_memory_barrier:
 	case nir_intrinsic_group_memory_barrier:
-		waitcnt &= VM_CNT & LGKM_CNT;
+		wait_flags = AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE;
 		break;
 	case nir_intrinsic_memory_barrier_atomic_counter:
 	case nir_intrinsic_memory_barrier_buffer:
 	case nir_intrinsic_memory_barrier_image:
-		waitcnt &= VM_CNT;
+		wait_flags = AC_WAIT_VLOAD | AC_WAIT_VSTORE;
 		break;
 	case nir_intrinsic_memory_barrier_shared:
-		waitcnt &= LGKM_CNT;
+		wait_flags = AC_WAIT_LGKM;
 		break;
 	default:
 		break;
 	}
-	if (waitcnt != NOOP_WAITCNT)
+
-		ac_build_waitcnt(ac, waitcnt);
+	ac_build_waitcnt(ac, wait_flags);
 }
 
 void ac_emit_barrier(struct ac_llvm_context *ac, gl_shader_stage stage)
@@ -2770,7 +2770,7 @@ void ac_emit_barrier(struct ac_llvm_context *ac, gl_shader_stage stage)
 	 * always fits into a single wave.
 	 */
 	if (ac->chip_class == GFX6 && stage == MESA_SHADER_TESS_CTRL) {
-		ac_build_waitcnt(ac, LGKM_CNT & VM_CNT);
+		ac_build_waitcnt(ac, AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE);
 		return;
 	}
 	ac_build_s_barrier(ac);

src/gallium/drivers/radeonsi/si_shader.c

@@ -3908,21 +3908,20 @@ static void membar_emit(
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	LLVMValueRef src0 = lp_build_emit_fetch(bld_base, emit_data->inst, 0, 0);
 	unsigned flags = LLVMConstIntGetZExtValue(src0);
-	unsigned waitcnt = NOOP_WAITCNT;
+	unsigned wait_flags = 0;
 
 	if (flags & TGSI_MEMBAR_THREAD_GROUP)
-		waitcnt &= VM_CNT & LGKM_CNT;
+		wait_flags |= AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE;
 
 	if (flags & (TGSI_MEMBAR_ATOMIC_BUFFER |
 		     TGSI_MEMBAR_SHADER_BUFFER |
 		     TGSI_MEMBAR_SHADER_IMAGE))
-		waitcnt &= VM_CNT;
+		wait_flags |= AC_WAIT_VLOAD | AC_WAIT_VSTORE;
 
 	if (flags & TGSI_MEMBAR_SHARED)
-		waitcnt &= LGKM_CNT;
+		wait_flags |= AC_WAIT_LGKM;
 
-	if (waitcnt != NOOP_WAITCNT)
+	ac_build_waitcnt(&ctx->ac, wait_flags);
-		ac_build_waitcnt(&ctx->ac, waitcnt);
 }
 
 static void clock_emit(
@@ -4372,7 +4371,7 @@ static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
 	 */
 	if (ctx->screen->info.chip_class == GFX6 &&
 	    ctx->type == PIPE_SHADER_TESS_CTRL) {
-		ac_build_waitcnt(&ctx->ac, LGKM_CNT & VM_CNT);
+		ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE);
 		return;
 	}
 

src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c

@@ -520,7 +520,7 @@ static void load_emit(
 	}
 
 	if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
-		ac_build_waitcnt(&ctx->ac, VM_CNT);
+		ac_build_waitcnt(&ctx->ac, AC_WAIT_VLOAD | AC_WAIT_VSTORE);
 
 	can_speculate = !(inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE) &&
 			  is_oneway_access_only(inst, info,
@@ -709,7 +709,7 @@ static void store_emit(
 	}
 
 	if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
-		ac_build_waitcnt(&ctx->ac, VM_CNT);
+		ac_build_waitcnt(&ctx->ac, AC_WAIT_VLOAD | AC_WAIT_VSTORE);
 
 	bool is_image = inst->Dst[0].Register.File != TGSI_FILE_BUFFER;
 	args.cache_policy = get_cache_policy(ctx, inst,