Merge pull request #844 from pcercuei/update-lightrec-20240902

Update Lightrec 2024-09-02

deps/lightrec/.gitrepo

@@ -6,7 +6,7 @@
 [subrepo]
 	remote = https://github.com/pcercuei/lightrec.git
 	branch = master
-	commit = 601afca8e889bdda7040ff5c64f7bbd20d1d5f2c
-	parent = 459f02ad03fa10b5c403fed724d47fe5adfd5fb1
+	commit = ea20362c9542f12fb6a0f27aa7df66b2af06b84d
+	parent = 8847df50c67c19c605f60a109d30556b74d08eee
 	method = merge
 	cmdver = 0.4.6

deps/lightrec/constprop.c

@@ -59,7 +59,7 @@ static void lightrec_propagate_addi(u32 rs, u32 rd,
 				    const struct constprop_data *d,
 				    struct constprop_data *v)
 {
-	u32 end, bit, sum, min, mask, imm, value;
+	u32 end, bit, sum, min, max, mask, imm, value;
 	struct constprop_data result = {
 		.value = v[rd].value,
 		.known = v[rd].known,
@@ -110,6 +110,15 @@ static void lightrec_propagate_addi(u32 rs, u32 rd,
 					 * sign bits are known. */
 					min = get_min_value(&v[rs])
 						+ get_min_value(d);
+					max = get_max_value(&v[rs])
+						+ get_max_value(d);
+
+					/* The sum may have less sign bits */
+					if ((s32)min < 0)
+						mask &= min & max;
+					else
+						mask &= ~(min | mask);
+
 					result.value = (min & mask)
 						| (result.value & ~mask);
 					result.known |= mask << carry;

deps/lightrec/emitter.c

@@ -1300,7 +1300,7 @@ static void rec_store_memory(struct lightrec_cstate *cstate,
 	struct opcode *op = &block->opcode_list[offset];
 	jit_state_t *_jit = block->_jit;
 	union code c = op->c;
-	u8 rs, rt, tmp = 0, tmp2 = 0, tmp3, addr_reg, addr_reg2;
+	u8 rs, rt, tmp = 0, tmp2 = 0, tmp3, addr_reg, addr_reg2, src_reg;
 	s16 imm = (s16)c.i.imm;
 	s32 simm = (s32)imm << (1 - lut_is_32bit(state));
 	s32 lut_offt = lightrec_offset(code_lut);
@@ -1342,25 +1342,23 @@ static void rec_store_memory(struct lightrec_cstate *cstate,
 	}
 
 	rt = lightrec_alloc_reg_in(reg_cache, _jit, in_reg, 0);
+	src_reg = rt;
 
 	if (is_big_endian() && swap_code && in_reg) {
 		tmp3 = lightrec_alloc_reg_temp(reg_cache, _jit);
 
 		jit_new_node_ww(swap_code, tmp3, rt);
 
-		if (c.i.op == OP_META_SWU)
-			jit_unstr(addr_reg2, tmp3, LIGHTNING_UNALIGNED_32BIT);
-		else
-			jit_new_node_www(code, imm, addr_reg2, tmp3);
-
-		lightrec_free_reg(reg_cache, tmp3);
-	} else if (c.i.op == OP_META_SWU) {
-		jit_unstr(addr_reg2, rt, LIGHTNING_UNALIGNED_32BIT);
-	} else {
-		jit_new_node_www(code, imm, addr_reg2, rt);
+		lightrec_free_reg(reg_cache, rt);
+		src_reg = tmp3;
 	}
 
-	lightrec_free_reg(reg_cache, rt);
+	if (c.i.op == OP_META_SWU)
+		jit_unstr(addr_reg2, src_reg, LIGHTNING_UNALIGNED_32BIT);
+	else
+		jit_new_node_www(code, imm, addr_reg2, src_reg);
+
+	lightrec_free_reg(reg_cache, src_reg);
 
 	if (invalidate) {
 		tmp3 = lightrec_alloc_reg_in(reg_cache, _jit, 0, 0);
@@ -1445,7 +1443,7 @@ static void rec_store_direct_no_invalidate(struct lightrec_cstate *cstate,
 	jit_state_t *_jit = block->_jit;
 	jit_node_t *to_not_ram, *to_end;
 	bool swc2 = c.i.op == OP_SWC2;
-	u8 addr_reg, tmp, tmp2 = 0, rs, rt, in_reg = swc2 ? REG_TEMP : c.i.rt;
+	u8 addr_reg, tmp, tmp2 = 0, rs, rt, src_reg, in_reg = swc2 ? REG_TEMP : c.i.rt;
 	s16 imm;
 
 	jit_note(__FILE__, __LINE__);
@@ -1489,25 +1487,23 @@ static void rec_store_direct_no_invalidate(struct lightrec_cstate *cstate,
 	}
 
 	rt = lightrec_alloc_reg_in(reg_cache, _jit, in_reg, 0);
+	src_reg = rt;
 
 	if (is_big_endian() && swap_code && in_reg) {
 		tmp2 = lightrec_alloc_reg_temp(reg_cache, _jit);
 
 		jit_new_node_ww(swap_code, tmp2, rt);
+		src_reg = tmp2;
 
-		if (c.i.op == OP_META_SWU)
-			jit_unstr(tmp, tmp2, LIGHTNING_UNALIGNED_32BIT);
-		else
-			jit_new_node_www(code, imm, tmp, tmp2);
-
-		lightrec_free_reg(reg_cache, tmp2);
-	} else if (c.i.op == OP_META_SWU) {
-		jit_unstr(tmp, rt, LIGHTNING_UNALIGNED_32BIT);
-	} else {
-		jit_new_node_www(code, imm, tmp, rt);
+		lightrec_free_reg(reg_cache, rt);
 	}
 
-	lightrec_free_reg(reg_cache, rt);
+	if (c.i.op == OP_META_SWU)
+		jit_unstr(tmp, src_reg, LIGHTNING_UNALIGNED_32BIT);
+	else
+		jit_new_node_www(code, imm, tmp, src_reg);
+
+	lightrec_free_reg(reg_cache, src_reg);
 	lightrec_free_reg(reg_cache, tmp);
 }
 
@@ -1521,7 +1517,7 @@ static void rec_store_direct(struct lightrec_cstate *cstate, const struct block
 	jit_state_t *_jit = block->_jit;
 	jit_node_t *to_not_ram, *to_end;
 	bool swc2 = c.i.op == OP_SWC2;
-	u8 addr_reg, tmp, tmp2, tmp3, rs, rt, reg_imm;
+	u8 src_reg, addr_reg, tmp, tmp2, tmp3, rs, rt, reg_imm;
 	u8 in_reg = swc2 ? REG_TEMP : c.i.rt;
 	u32 mask;
 	bool different_offsets = state->offset_ram != state->offset_scratch;
@@ -1602,25 +1598,23 @@ static void rec_store_direct(struct lightrec_cstate *cstate, const struct block
 	lightrec_free_reg(reg_cache, reg_imm);
 
 	rt = lightrec_alloc_reg_in(reg_cache, _jit, in_reg, 0);
+	src_reg = rt;
 
 	if (is_big_endian() && swap_code && in_reg) {
 		tmp = lightrec_alloc_reg_temp(reg_cache, _jit);
 
 		jit_new_node_ww(swap_code, tmp, rt);
+		src_reg = tmp;
 
-		if (c.i.op == OP_META_SWU)
-			jit_unstr(tmp2, tmp, LIGHTNING_UNALIGNED_32BIT);
-		else
-			jit_new_node_www(code, 0, tmp2, tmp);
-
-		lightrec_free_reg(reg_cache, tmp);
-	} else if (c.i.op == OP_META_SWU) {
-		jit_unstr(tmp2, rt, LIGHTNING_UNALIGNED_32BIT);
-	} else {
-		jit_new_node_www(code, 0, tmp2, rt);
+		lightrec_free_reg(reg_cache, rt);
 	}
 
-	lightrec_free_reg(reg_cache, rt);
+	if (c.i.op == OP_META_SWU)
+		jit_unstr(tmp2, src_reg, LIGHTNING_UNALIGNED_32BIT);
+	else
+		jit_new_node_www(code, 0, tmp2, src_reg);
+
+	lightrec_free_reg(reg_cache, src_reg);
 	lightrec_free_reg(reg_cache, tmp2);
 }
 
@@ -1882,19 +1876,19 @@ static void rec_load_direct(struct lightrec_cstate *cstate,
 		else
 			addr_mask = 0x1fffffff;
 
-		reg_imm = lightrec_alloc_reg_temp_with_value(reg_cache, _jit,
-							     addr_mask);
 		if (!state->mirrors_mapped) {
+			reg_imm = lightrec_alloc_reg_temp_with_value(reg_cache, _jit,
+								     addr_mask);
 			jit_andi(tmp, addr_reg, BIT(28));
 			jit_rshi_u(tmp, tmp, 28 - 22);
 			jit_orr(tmp, tmp, reg_imm);
 			jit_andr(rt, addr_reg, tmp);
+
+			lightrec_free_reg(reg_cache, reg_imm);
 		} else {
-			jit_andr(rt, addr_reg, reg_imm);
+			rec_and_mask(cstate, _jit, rt, addr_reg, addr_mask);
 		}
 
-		lightrec_free_reg(reg_cache, reg_imm);
-
 		if (state->offset_ram) {
 			offt_reg = lightrec_get_reg_with_value(reg_cache,
 							       state->offset_ram);

deps/lightrec/lightrec.c

@@ -3,6 +3,7 @@
  * Copyright (C) 2014-2021 Paul Cercueil <paul@crapouillou.net>
  */
 
+#include "arch.h"
 #include "blockcache.h"
 #include "debug.h"
 #include "disassembler.h"
@@ -1138,6 +1139,9 @@ static struct block * generate_dispatcher(struct lightrec_state *state)
 
 	loop = jit_label();
 
+	if (!arch_has_fast_mask())
+		jit_movi(JIT_R1, 0x1fffffff);
+
 	/* Call the block's code */
 	jit_jmpr(JIT_V1);
 
@@ -1582,6 +1586,9 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 	if (OPT_PRELOAD_PC && (block->flags & BLOCK_PRELOAD_PC))
 		lightrec_preload_pc(cstate->reg_cache, _jit);
 
+	if (!arch_has_fast_mask())
+		lightrec_preload_imm(cstate->reg_cache, _jit, JIT_R1, 0x1fffffff);
+
 	cstate->cycles = 0;
 	cstate->nb_local_branches = 0;
 	cstate->nb_targets = 0;

deps/lightrec/regcache.c

@@ -699,6 +699,18 @@ void lightrec_preload_pc(struct regcache *cache, jit_state_t *_jit)
 	jit_live(JIT_V0);
 }
 
+void lightrec_preload_imm(struct regcache *cache, jit_state_t *_jit,
+			  u8 jit_reg, u32 imm)
+{
+	struct native_register *nreg;
+
+	nreg = lightning_reg_to_lightrec(cache, jit_reg);
+	nreg->prio = REG_IS_TEMP_VALUE;
+	nreg->value = imm;
+
+	jit_live(jit_reg);
+}
+
 struct regcache * lightrec_regcache_init(struct lightrec_state *state)
 {
 	struct regcache *cache;

deps/lightrec/regcache.h

@@ -70,6 +70,8 @@ void lightrec_set_reg_out_flags(struct regcache *cache, u8 jit_reg, u8 flags);
 
 void lightrec_regcache_reset(struct regcache *cache);
 void lightrec_preload_pc(struct regcache *cache, jit_state_t *_jit);
+void lightrec_preload_imm(struct regcache *cache, jit_state_t *_jit,
+			  u8 jit_reg, u32 imm);
 
 void lightrec_free_reg(struct regcache *cache, u8 jit_reg);
 void lightrec_free_regs(struct regcache *cache);