- do not use imul mem, imm32 on newer AMD cpu's

[r19288]

diff --git a/ir/be/ia32/ia32_architecture.c b/ir/be/ia32/ia32_architecture.c
index 32e4871..ff4be15 100644 (file)
--- a/ir/be/ia32/ia32_architecture.c
+++ b/ir/be/ia32/ia32_architecture.c
@@ -443,6 +443,7 @@ void ia32_setup_cg_config(void)
                                              IS_P6_ARCH(opt_arch) || IS_NETBURST_ARCH(opt_arch) ||
                                              (opt_arch == arch_core2) || (opt_arch == arch_generic) ||
                                              (opt_arch == arch_i386) || (opt_arch == arch_i486);
+       ia32_cg_config.use_imul_mem_imm32   = !(opt_arch == arch_opteron || opt_arch == arch_k10);
        ia32_cg_config.optimize_cc          = opt_cc;
        ia32_cg_config.use_unsafe_floatconv = opt_unsafe_floatconv;
 

diff --git a/ir/be/ia32/ia32_architecture.h b/ir/be/ia32/ia32_architecture.h
index ba2277f..af54faa 100644 (file)
--- a/ir/be/ia32/ia32_architecture.h
+++ b/ir/be/ia32/ia32_architecture.h
@@ -53,6 +53,8 @@ typedef struct {
        unsigned use_sub_esp_4:1;
        /** use sub esp, 8 instead of 2 push's */
        unsigned use_sub_esp_8:1;
+       /** use imul mem, imm32 instruction (slow on some cpu's */
+       unsigned use_imul_mem_imm32:1;
        /** optimize calling convention where possible */
        unsigned optimize_cc:1;
        /**

diff --git a/ir/be/ia32/ia32_transform.c b/ir/be/ia32/ia32_transform.c
index 41ae349..74c29cd 100644 (file)
--- a/ir/be/ia32/ia32_transform.c
+++ b/ir/be/ia32/ia32_transform.c
@@ -1152,6 +1152,7 @@ static ir_node *gen_Mul(ir_node *node) {
        ir_node *op1  = get_Mul_left(node);
        ir_node *op2  = get_Mul_right(node);
        ir_mode *mode = get_irn_mode(node);
+       unsigned flags;
 
        if (mode_is_float(mode)) {
                if (ia32_cg_config.use_sse2)
@@ -1165,9 +1166,10 @@ static ir_node *gen_Mul(ir_node *node) {
        /* for the lower 32bit of the result it doesn't matter whether we use
         * signed or unsigned multiplication so we use IMul as it has fewer
         * constraints */
-       return gen_binop(node, op1, op2, new_rd_ia32_IMul,
-                        match_commutative | match_am | match_mode_neutral |
-                        match_immediate | match_am_and_immediates);
+       flags = match_commutative | match_am | match_mode_neutral | match_immediate;
+       if (ia32_cg_config.use_imul_mem_imm32)
+               flags |= match_am_and_immediates;
+       return gen_binop(node, op1, op2, new_rd_ia32_IMul, flags);
 }
 
 /**