As discovered by Rainer, an oversight in sparc_rtx_costs causes the function
to return very high costs for multiply operations with -m64 for default V9 or
new processors (when TARGET_DEPRECATED_V8_INSNS is not set to be precise).
Fixed by the attached patch to config/sparc/sparc.c, the config/sparc/sparc.h
hunk being a no-op (because TARGET_V8PLUS => TARGET_DEPRECATED_V8_INSNS).
Tested on SPARC/Solaris and SPARC64/Solaris, applied on the mainline.
2015-02-03 Eric Botcazou <ebotca...@adacore.com>
PR target/62631
* config/sparc/sparc.h (TARGET_HARD_MUL): Remove TARGET_V8PLUS.
(TARGET_HARD_MUL32): Rewrite based on TARGET_HARD_MUL.
* config/sparc/sparc.c (sparc_rtx_costs) <MULT>: Return costs based on
int_mulX for integers in 64-bit mode if TARGET_HARD_MUL is not set.
--
Eric BotcazouIndex: config/sparc/sparc.c
===================================================================
--- config/sparc/sparc.c (revision 220343)
+++ config/sparc/sparc.c (working copy)
@@ -11075,7 +11075,7 @@ sparc_rtx_costs (rtx x, int code, int ou
case MULT:
if (float_mode_p)
*total = sparc_costs->float_mul;
- else if (! TARGET_HARD_MUL)
+ else if (TARGET_ARCH32 && !TARGET_HARD_MUL)
*total = COSTS_N_INSNS (25);
else
{
@@ -11113,7 +11113,7 @@ sparc_rtx_costs (rtx x, int code, int ou
bit_cost = COSTS_N_INSNS (bit_cost);
}
- if (mode == DImode)
+ if (mode == DImode || !TARGET_HARD_MUL)
*total = sparc_costs->int_mulX + bit_cost;
else
*total = sparc_costs->int_mul + bit_cost;
Index: config/sparc/sparc.h
===================================================================
--- config/sparc/sparc.h (revision 220343)
+++ config/sparc/sparc.h (working copy)
@@ -426,22 +426,20 @@ extern enum cmodel sparc_cmodel;
#define WCHAR_TYPE_SIZE 16
/* Mask of all CPU selection flags. */
-#define MASK_ISA \
-(MASK_V8 + MASK_SPARCLITE + MASK_SPARCLET + MASK_V9 + MASK_DEPRECATED_V8_INSNS)
+#define MASK_ISA \
+ (MASK_SPARCLITE + MASK_SPARCLET \
+ + MASK_V8 + MASK_V9 + MASK_DEPRECATED_V8_INSNS)
-/* TARGET_HARD_MUL: Use hardware multiply instructions but not %y.
- TARGET_HARD_MUL32: Use hardware multiply instructions with rd %y
- to get high 32 bits. False in V8+ or V9 because multiply stores
- a 64-bit result in a register. */
-
-#define TARGET_HARD_MUL32 \
- ((TARGET_V8 || TARGET_SPARCLITE \
- || TARGET_SPARCLET || TARGET_DEPRECATED_V8_INSNS) \
- && ! TARGET_V8PLUS && TARGET_ARCH32)
+/* TARGET_HARD_MUL: Use 32-bit hardware multiply instructions but not %y. */
+#define TARGET_HARD_MUL \
+ (TARGET_SPARCLITE || TARGET_SPARCLET \
+ || TARGET_V8 || TARGET_DEPRECATED_V8_INSNS)
-#define TARGET_HARD_MUL \
- (TARGET_V8 || TARGET_SPARCLITE || TARGET_SPARCLET \
- || TARGET_DEPRECATED_V8_INSNS || TARGET_V8PLUS)
+/* TARGET_HARD_MUL32: Use 32-bit hardware multiply instructions with %y
+ to get high 32 bits. False in 64-bit or V8+ because multiply stores
+ a 64-bit result in a register. */
+#define TARGET_HARD_MUL32 \
+ (TARGET_HARD_MUL && TARGET_ARCH32 && !TARGET_V8PLUS)
/* MASK_APP_REGS must always be the default because that's what
FIXED_REGISTERS is set to and -ffixed- is processed before
