===================================================================
@@ -2604,25 +2604,141 @@
;; Floating point move instructions.
-(define_expand "movsf"
- [(set (match_operand:SF 0 "nonimmediate_operand" "")
- (match_operand:SF 1 "general_operand" ""))]
+(define_expand "movtf"
+ [(set (match_operand:TF 0 "nonimmediate_operand" "")
+ (match_operand:TF 1 "nonimmediate_operand" ""))]
+ "TARGET_SSE2"
+{
+ ix86_expand_move (TFmode, operands);
+ DONE;
+})
+
+(define_expand "mov<mode>"
+ [(set (match_operand:X87MODEF 0 "nonimmediate_operand" "")
+ (match_operand:X87MODEF 1 "general_operand" ""))]
""
- "ix86_expand_move (SFmode, operands); DONE;")
+ "ix86_expand_move (<MODE>mode, operands); DONE;")
-(define_insn "*pushsf"
- [(set (match_operand:SF 0 "push_operand" "=<,<,<")
- (match_operand:SF 1 "general_no_elim_operand" "f,rFm,x"))]
- "!TARGET_64BIT"
+(define_insn "*pushtf"
+ [(set (match_operand:TF 0 "push_operand" "=<,<,<")
+ (match_operand:TF 1 "general_no_elim_operand" "x,Fo,*r"))]
+ "TARGET_SSE2"
{
- /* Anything else should be already split before reg-stack. */
- gcc_assert (which_alternative == 1);
- return "push{l}\t%1";
+ /* This insn should be already split before reg-stack. */
+ gcc_unreachable ();
}
- [(set_attr "type" "multi,push,multi")
+ [(set_attr "type" "multi")
+ (set_attr "unit" "sse,*,*")
+ (set_attr "mode" "TF,SI,SI")])
+
+(define_split
+ [(set (match_operand:TF 0 "push_operand" "")
+ (match_operand:TF 1 "general_operand" ""))]
+ "TARGET_SSE2 && reload_completed
+ && !SSE_REG_P (operands[1])"
+ [(const_int 0)]
+ "ix86_split_long_move (operands); DONE;")
+
+(define_split
+ [(set (match_operand:TF 0 "push_operand" "")
+ (match_operand:TF 1 "any_fp_register_operand" ""))]
+ "TARGET_SSE2"
+ [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (const_int -16)))
+ (set (mem:TF (reg:P SP_REG)) (match_dup 1))]
+ "")
+
+(define_insn "*pushxf"
+ [(set (match_operand:XF 0 "push_operand" "=<,<")
+ (match_operand:XF 1 "general_no_elim_operand" "f,ro"))]
+ "optimize_function_for_speed_p (cfun)"
+{
+ /* This insn should be already split before reg-stack. */
+ gcc_unreachable ();
+}
+ [(set_attr "type" "multi")
+ (set_attr "unit" "i387,*")
+ (set_attr "mode" "XF,SI")])
+
+;; Size of pushxf is 3 (for sub) + 2 (for fstp) + memory operand size.
+;; Size of pushxf using integer instructions is 3+3*memory operand size
+;; Pushing using integer instructions is longer except for constants
+;; and direct memory references (assuming that any given constant is pushed
+;; only once, but this ought to be handled elsewhere).
+
+(define_insn "*pushxf_nointeger"
+ [(set (match_operand:XF 0 "push_operand" "=X,X,X")
+ (match_operand:XF 1 "general_no_elim_operand" "f,Fo,*r"))]
+ "optimize_function_for_size_p (cfun)"
+{
+ /* This insn should be already split before reg-stack. */
+ gcc_unreachable ();
+}
+ [(set_attr "type" "multi")
(set_attr "unit" "i387,*,*")
- (set_attr "mode" "SF,SI,SF")])
+ (set_attr "mode" "XF,SI,SI")])
+(define_split
+ [(set (match_operand:XF 0 "push_operand" "")
+ (match_operand:XF 1 "any_fp_register_operand" ""))]
+ "reload_completed"
+ [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (match_dup 2)))
+ (set (mem:XF (reg:P SP_REG)) (match_dup 1))]
+ "operands[2] = GEN_INT (TARGET_128BIT_LONG_DOUBLE ? -16 : -12);")
+
+(define_split
+ [(set (match_operand:XF 0 "push_operand" "")
+ (match_operand:XF 1 "general_operand" ""))]
+ "reload_completed
+ && !ANY_FP_REG_P (operands[1])"
+ [(const_int 0)]
+ "ix86_split_long_move (operands); DONE;")
+
+(define_insn "*pushdf"
+ [(set (match_operand:DF 0 "push_operand" "=<,<,<")
+ (match_operand:DF 1 "general_no_elim_operand" "f,rFo,Y2"))]
+ "TARGET_64BIT || TARGET_INTEGER_DFMODE_MOVES"
+{
+ /* This insn should be already split before reg-stack. */
+ gcc_unreachable ();
+}
+ [(set_attr "type" "multi")
+ (set_attr "unit" "i387,*,*")
+ (set_attr "mode" "DF,SI,DF")])
+
+;; Size of pushdf is 3 (for sub) + 2 (for fstp) + memory operand size.
+;; Size of pushdf using integer instructions is 2+2*memory operand size
+;; On the average, pushdf using integers can be still shorter. Allow this
+;; pattern for optimize_size too.
+
+(define_insn "*pushdf_nointeger"
+ [(set (match_operand:DF 0 "push_operand" "=<,<,<,<")
+ (match_operand:DF 1 "general_no_elim_operand" "f,Fo,*r,Y2"))]
+ "!(TARGET_64BIT || TARGET_INTEGER_DFMODE_MOVES)"
+{
+ /* This insn should be already split before reg-stack. */
+ gcc_unreachable ();
+}
+ [(set_attr "type" "multi")
+ (set_attr "unit" "i387,*,*,*")
+ (set_attr "mode" "DF,SI,SI,DF")])
+
+;; %%% Kill this when call knows how to work this out.
+(define_split
+ [(set (match_operand:DF 0 "push_operand" "")
+ (match_operand:DF 1 "any_fp_register_operand" ""))]
+ "reload_completed"
+ [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (const_int -8)))
+ (set (mem:DF (reg:P SP_REG)) (match_dup 1))]
+ "")
+
+(define_split
+ [(set (match_operand:DF 0 "push_operand" "")
+ (match_operand:DF 1 "general_operand" ""))]
+ "reload_completed
+ && !ANY_FP_REG_P (operands[1])"
+ [(const_int 0)]
+ "ix86_split_long_move (operands); DONE;")
+
(define_insn "*pushsf_rex64"
[(set (match_operand:SF 0 "push_operand" "=X,X,X")
(match_operand:SF 1 "nonmemory_no_elim_operand" "f,rF,x"))]
@@ -2636,6 +2752,19 @@
(set_attr "unit" "i387,*,*")
(set_attr "mode" "SF,DI,SF")])
+(define_insn "*pushsf"
+ [(set (match_operand:SF 0 "push_operand" "=<,<,<")
+ (match_operand:SF 1 "general_no_elim_operand" "f,rFm,x"))]
+ "!TARGET_64BIT"
+{
+ /* Anything else should be already split before reg-stack. */
+ gcc_assert (which_alternative == 1);
+ return "push{l}\t%1";
+}
+ [(set_attr "type" "multi,push,multi")
+ (set_attr "unit" "i387,*,*")
+ (set_attr "mode" "SF,SI,SF")])
+
(define_split
[(set (match_operand:SF 0 "push_operand" "")
(match_operand:SF 1 "memory_operand" ""))]
@@ -2649,202 +2778,148 @@
(define_split
[(set (match_operand:SF 0 "push_operand" "")
(match_operand:SF 1 "any_fp_register_operand" ""))]
- "!TARGET_64BIT"
- [(set (reg:SI SP_REG) (plus:SI (reg:SI SP_REG) (const_int -4)))
- (set (mem:SF (reg:SI SP_REG)) (match_dup 1))])
+ "reload_completed"
+ [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (match_dup 2)))
+ (set (mem:SF (reg:P SP_REG)) (match_dup 1))]
+ "operands[2] = GEN_INT (-GET_MODE_SIZE (<MODE>mode));")
-(define_split
- [(set (match_operand:SF 0 "push_operand" "")
- (match_operand:SF 1 "any_fp_register_operand" ""))]
- "TARGET_64BIT"
- [(set (reg:DI SP_REG) (plus:DI (reg:DI SP_REG) (const_int -8)))
- (set (mem:SF (reg:DI SP_REG)) (match_dup 1))])
-
-(define_insn "*movsf_1"
- [(set (match_operand:SF 0 "nonimmediate_operand"
- "=f,m,f,r ,m ,x,x,x ,m,!*y,!m,!*y,?Yi,?r,!*Ym,!r")
- (match_operand:SF 1 "general_operand"
- "fm,f,G,rmF,Fr,C,x,xm,x,m ,*y,*y ,r ,Yi,r ,*Ym"))]
- "!(MEM_P (operands[0]) && MEM_P (operands[1]))
- && (reload_in_progress || reload_completed
- || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
- || (!TARGET_SSE_MATH && optimize_function_for_size_p (cfun)
- && standard_80387_constant_p (operands[1]))
- || GET_CODE (operands[1]) != CONST_DOUBLE
- || memory_operand (operands[0], SFmode))"
+(define_insn "*movtf_internal"
+ [(set (match_operand:TF 0 "nonimmediate_operand" "=x,m,x,?r,?o")
+ (match_operand:TF 1 "general_operand" "xm,x,C,roF,Fr"))]
+ "TARGET_SSE2
+ && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
{
switch (which_alternative)
{
case 0:
case 1:
- return output_387_reg_move (insn, operands);
-
- case 2:
- return standard_80387_constant_opcode (operands[1]);
-
- case 3:
- case 4:
- return "mov{l}\t{%1, %0|%0, %1}";
- case 5:
- if (get_attr_mode (insn) == MODE_TI)
- return "%vpxor\t%0, %d0";
- else
- return "%vxorps\t%0, %d0";
- case 6:
if (get_attr_mode (insn) == MODE_V4SF)
return "%vmovaps\t{%1, %0|%0, %1}";
else
- return "%vmovss\t{%1, %d0|%d0, %1}";
- case 7:
- if (TARGET_AVX)
- return REG_P (operands[1]) ? "vmovss\t{%1, %0, %0|%0, %0, %1}"
- : "vmovss\t{%1, %0|%0, %1}";
+ return "%vmovdqa\t{%1, %0|%0, %1}";
+ case 2:
+ if (get_attr_mode (insn) == MODE_V4SF)
+ return "%vxorps\t%0, %d0";
else
- return "movss\t{%1, %0|%0, %1}";
- case 8:
- return "%vmovss\t{%1, %0|%0, %1}";
-
- case 9: case 10: case 14: case 15:
- return "movd\t{%1, %0|%0, %1}";
- case 12: case 13:
- return "%vmovd\t{%1, %0|%0, %1}";
-
- case 11:
- return "movq\t{%1, %0|%0, %1}";
-
+ return "%vpxor\t%0, %d0";
+ case 3:
+ case 4:
+ return "#";
default:
gcc_unreachable ();
}
}
- [(set_attr "type" "fmov,fmov,fmov,imov,imov,sselog1,ssemov,ssemov,ssemov,mmxmov,mmxmov,mmxmov,ssemov,ssemov,mmxmov,mmxmov")
- (set (attr "prefix")
- (if_then_else (eq_attr "alternative" "5,6,7,8,12,13")
- (const_string "maybe_vex")
- (const_string "orig")))
+ [(set_attr "type" "ssemov,ssemov,sselog1,*,*")
+ (set_attr "prefix" "maybe_vex,maybe_vex,maybe_vex,*,*")
(set (attr "mode")
- (cond [(eq_attr "alternative" "3,4,9,10")
- (const_string "SI")
- (eq_attr "alternative" "5")
+ (cond [(eq_attr "alternative" "0,2")
(if_then_else
- (and (and (ne (symbol_ref "TARGET_SSE_LOAD0_BY_PXOR")
- (const_int 0))
- (ne (symbol_ref "TARGET_SSE2")
- (const_int 0)))
- (eq (symbol_ref "optimize_function_for_size_p (cfun)")
- (const_int 0)))
- (const_string "TI")
- (const_string "V4SF"))
- /* For architectures resolving dependencies on
- whole SSE registers use APS move to break dependency
- chains, otherwise use short move to avoid extra work.
-
- Do the same for architectures resolving dependencies on
- the parts. While in DF mode it is better to always handle
- just register parts, the SF mode is different due to lack
- of instructions to load just part of the register. It is
- better to maintain the whole registers in single format
- to avoid problems on using packed logical operations. */
- (eq_attr "alternative" "6")
+ (ne (symbol_ref "optimize_function_for_size_p (cfun)")
+ (const_int 0))
+ (const_string "V4SF")
+ (const_string "TI"))
+ (eq_attr "alternative" "1")
(if_then_else
- (ior (ne (symbol_ref "TARGET_SSE_PARTIAL_REG_DEPENDENCY")
+ (ior (ne (symbol_ref "TARGET_SSE_TYPELESS_STORES")
(const_int 0))
- (ne (symbol_ref "TARGET_SSE_SPLIT_REGS")
+ (ne (symbol_ref "optimize_function_for_size_p (cfun)")
(const_int 0)))
(const_string "V4SF")
- (const_string "SF"))
- (eq_attr "alternative" "11")
- (const_string "DI")]
- (const_string "SF")))])
+ (const_string "TI"))]
+ (const_string "DI")))])
-(define_insn "*swapsf"
- [(set (match_operand:SF 0 "fp_register_operand" "+f")
- (match_operand:SF 1 "fp_register_operand" "+f"))
- (set (match_dup 1)
- (match_dup 0))]
- "reload_completed || TARGET_80387"
+(define_split
+ [(set (match_operand:TF 0 "nonimmediate_operand" "")
+ (match_operand:TF 1 "general_operand" ""))]
+ "reload_completed
+ && !(SSE_REG_P (operands[0]) || SSE_REG_P (operands[1]))"
+ [(const_int 0)]
+ "ix86_split_long_move (operands); DONE;")
+
+(define_insn "*movxf_internal"
+ [(set (match_operand:XF 0 "nonimmediate_operand" "=f,m,f,r,o")
+ (match_operand:XF 1 "general_operand" "fm,f,G,roF,Fr"))]
+ "optimize_function_for_speed_p (cfun)
+ && !(MEM_P (operands[0]) && MEM_P (operands[1]))
+ && (reload_in_progress || reload_completed
+ || GET_CODE (operands[1]) != CONST_DOUBLE
+ || memory_operand (operands[0], XFmode))"
{
- if (STACK_TOP_P (operands[0]))
- return "fxch\t%1";
- else
- return "fxch\t%0";
-}
- [(set_attr "type" "fxch")
- (set_attr "mode" "SF")])
+ switch (which_alternative)
+ {
+ case 0:
+ case 1:
+ return output_387_reg_move (insn, operands);
-(define_expand "movdf"
- [(set (match_operand:DF 0 "nonimmediate_operand" "")
- (match_operand:DF 1 "general_operand" ""))]
- ""
- "ix86_expand_move (DFmode, operands); DONE;")
+ case 2:
+ return standard_80387_constant_opcode (operands[1]);
-;; Size of pushdf is 3 (for sub) + 2 (for fstp) + memory operand size.
-;; Size of pushdf using integer instructions is 2+2*memory operand size
-;; On the average, pushdf using integers can be still shorter. Allow this
-;; pattern for optimize_size too.
+ case 3: case 4:
+ return "#";
-(define_insn "*pushdf_nointeger"
- [(set (match_operand:DF 0 "push_operand" "=<,<,<,<")
- (match_operand:DF 1 "general_no_elim_operand" "f,Fo,*r,Y2"))]
- "!TARGET_64BIT && !TARGET_INTEGER_DFMODE_MOVES"
-{
- /* This insn should be already split before reg-stack. */
- gcc_unreachable ();
+ default:
+ gcc_unreachable ();
+ }
}
- [(set_attr "type" "multi")
- (set_attr "unit" "i387,*,*,*")
- (set_attr "mode" "DF,SI,SI,DF")])
+ [(set_attr "type" "fmov,fmov,fmov,multi,multi")
+ (set_attr "mode" "XF,XF,XF,SI,SI")])
-(define_insn "*pushdf_integer"
- [(set (match_operand:DF 0 "push_operand" "=<,<,<")
- (match_operand:DF 1 "general_no_elim_operand" "f,rFo,Y2"))]
- "TARGET_64BIT || TARGET_INTEGER_DFMODE_MOVES"
+;; Do not use integer registers when optimizing for size
+(define_insn "*movxf_internal_nointeger"
+ [(set (match_operand:XF 0 "nonimmediate_operand" "=f,m,f,*r,o")
+ (match_operand:XF 1 "general_operand" "fm,f,G,*roF,F*r"))]
+ "optimize_function_for_size_p (cfun)
+ && !(MEM_P (operands[0]) && MEM_P (operands[1]))
+ && (reload_in_progress || reload_completed
+ || standard_80387_constant_p (operands[1])
+ || GET_CODE (operands[1]) != CONST_DOUBLE
+ || memory_operand (operands[0], XFmode))"
{
- /* This insn should be already split before reg-stack. */
- gcc_unreachable ();
+ switch (which_alternative)
+ {
+ case 0:
+ case 1:
+ return output_387_reg_move (insn, operands);
+
+ case 2:
+ return standard_80387_constant_opcode (operands[1]);
+
+ case 3: case 4:
+ return "#";
+ default:
+ gcc_unreachable ();
+ }
}
- [(set_attr "type" "multi")
- (set_attr "unit" "i387,*,*")
- (set_attr "mode" "DF,SI,DF")])
+ [(set_attr "type" "fmov,fmov,fmov,multi,multi")
+ (set_attr "mode" "XF,XF,XF,SI,SI")])
-;; %%% Kill this when call knows how to work this out.
(define_split
- [(set (match_operand:DF 0 "push_operand" "")
- (match_operand:DF 1 "any_fp_register_operand" ""))]
- "reload_completed"
- [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (const_int -8)))
- (set (mem:DF (reg:P SP_REG)) (match_dup 1))]
- "")
-
-(define_split
- [(set (match_operand:DF 0 "push_operand" "")
- (match_operand:DF 1 "general_operand" ""))]
- "reload_completed"
+ [(set (match_operand:XF 0 "nonimmediate_operand" "")
+ (match_operand:XF 1 "general_operand" ""))]
+ "reload_completed
+ && !(MEM_P (operands[0]) && MEM_P (operands[1]))
+ && ! (ANY_FP_REG_P (operands[0]) ||
+ (GET_CODE (operands[0]) == SUBREG
+ && ANY_FP_REG_P (SUBREG_REG (operands[0]))))
+ && ! (ANY_FP_REG_P (operands[1]) ||
+ (GET_CODE (operands[1]) == SUBREG
+ && ANY_FP_REG_P (SUBREG_REG (operands[1]))))"
[(const_int 0)]
"ix86_split_long_move (operands); DONE;")
-;; Moving is usually shorter when only FP registers are used. This separate
-;; movdf pattern avoids the use of integer registers for FP operations
-;; when optimizing for size.
-
-(define_insn "*movdf_nointeger"
+(define_insn "*movdf_internal_rex64"
[(set (match_operand:DF 0 "nonimmediate_operand"
- "=f,m,f,*r ,o ,Y2*x,Y2*x,Y2*x ,m ")
+ "=f,m,f,r ,m ,Y2*x,Y2*x,Y2*x,m ,Yi,r ")
(match_operand:DF 1 "general_operand"
- "fm,f,G,*roF,*Fr,C ,Y2*x,mY2*x,Y2*x"))]
- "!(MEM_P (operands[0]) && MEM_P (operands[1]))
- && ((optimize_function_for_size_p (cfun)
- || !TARGET_INTEGER_DFMODE_MOVES) && !TARGET_64BIT)
+ "fm,f,G,rmF,Fr,C ,Y2*x,m ,Y2*x,r ,Yi"))]
+ "TARGET_64BIT && !(MEM_P (operands[0]) && MEM_P (operands[1]))
&& (reload_in_progress || reload_completed
|| (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
|| (!(TARGET_SSE2 && TARGET_SSE_MATH)
&& optimize_function_for_size_p (cfun)
- && !memory_operand (operands[0], DFmode)
&& standard_80387_constant_p (operands[1]))
|| GET_CODE (operands[1]) != CONST_DOUBLE
- || ((optimize_function_for_size_p (cfun)
- || !TARGET_MEMORY_MISMATCH_STALL
- || reload_in_progress || reload_completed)
- && memory_operand (operands[0], DFmode)))"
+ || memory_operand (operands[0], DFmode))"
{
switch (which_alternative)
{
@@ -2858,6 +2933,7 @@
case 3:
case 4:
return "#";
+
case 5:
switch (get_attr_mode (insn))
{
@@ -2906,34 +2982,22 @@
else
return "movsd\t{%1, %0|%0, %1}";
case MODE_V1DF:
- if (TARGET_AVX)
- {
- if (REG_P (operands[0]))
- return "vmovlpd\t{%1, %0, %0|%0, %0, %1}";
- else
- return "vmovlpd\t{%1, %0|%0, %1}";
- }
- else
- return "movlpd\t{%1, %0|%0, %1}";
+ return "%vmovlpd\t{%1, %d0|%d0, %1}";
case MODE_V2SF:
- if (TARGET_AVX)
- {
- if (REG_P (operands[0]))
- return "vmovlps\t{%1, %0, %0|%0, %0, %1}";
- else
- return "vmovlps\t{%1, %0|%0, %1}";
- }
- else
- return "movlps\t{%1, %0|%0, %1}";
+ return "%vmovlps\t{%1, %d0|%d0, %1}";
default:
gcc_unreachable ();
}
+ case 9:
+ case 10:
+ return "%vmovd\t{%1, %0|%0, %1}";
+
default:
- gcc_unreachable ();
+ gcc_unreachable();
}
}
- [(set_attr "type" "fmov,fmov,fmov,multi,multi,sselog1,ssemov,ssemov,ssemov")
+ [(set_attr "type" "fmov,fmov,fmov,multi,multi,sselog1,ssemov,ssemov,ssemov,ssemov,ssemov")
(set (attr "prefix")
(if_then_else (eq_attr "alternative" "0,1,2,3,4")
(const_string "orig")
@@ -2945,8 +3009,8 @@
(set (attr "mode")
(cond [(eq_attr "alternative" "0,1,2")
(const_string "DF")
- (eq_attr "alternative" "3,4")
- (const_string "SI")
+ (eq_attr "alternative" "3,4,9,10")
+ (const_string "DI")
/* For SSE1, we have many fewer alternatives. */
(eq (symbol_ref "TARGET_SSE2") (const_int 0))
@@ -2993,12 +3057,14 @@
]
(const_string "DF")))])
-(define_insn "*movdf_integer_rex64"
+(define_insn "*movdf_internal"
[(set (match_operand:DF 0 "nonimmediate_operand"
- "=f,m,f,r ,m ,Y2*x,Y2*x,Y2*x,m ,Yi,r ")
+ "=f,m,f,r ,o ,Y2*x,Y2*x,Y2*x,m ")
(match_operand:DF 1 "general_operand"
- "fm,f,G,rmF,Fr,C ,Y2*x,m ,Y2*x,r ,Yi"))]
- "TARGET_64BIT && !(MEM_P (operands[0]) && MEM_P (operands[1]))
+ "fm,f,G,roF,Fr,C ,Y2*x,m ,Y2*x"))]
+ "!(MEM_P (operands[0]) && MEM_P (operands[1]))
+ && optimize_function_for_speed_p (cfun)
+ && TARGET_INTEGER_DFMODE_MOVES
&& (reload_in_progress || reload_completed
|| (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
|| (!(TARGET_SSE2 && TARGET_SSE_MATH)
@@ -3024,17 +3090,17 @@
switch (get_attr_mode (insn))
{
case MODE_V4SF:
- return "%vxorps\t%0, %d0";
+ return "xorps\t%0, %0";
case MODE_V2DF:
if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
- return "%vxorps\t%0, %d0";
+ return "xorps\t%0, %0";
else
- return "%vxorpd\t%0, %d0";
+ return "xorpd\t%0, %0";
case MODE_TI:
if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
- return "%vxorps\t%0, %d0";
+ return "xorps\t%0, %0";
else
- return "%vpxor\t%0, %d0";
+ return "pxor\t%0, %0";
default:
gcc_unreachable ();
}
@@ -3044,50 +3110,34 @@
switch (get_attr_mode (insn))
{
case MODE_V4SF:
- return "%vmovaps\t{%1, %0|%0, %1}";
+ return "movaps\t{%1, %0|%0, %1}";
case MODE_V2DF:
if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
- return "%vmovaps\t{%1, %0|%0, %1}";
+ return "movaps\t{%1, %0|%0, %1}";
else
- return "%vmovapd\t{%1, %0|%0, %1}";
+ return "movapd\t{%1, %0|%0, %1}";
case MODE_TI:
if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
- return "%vmovaps\t{%1, %0|%0, %1}";
+ return "movaps\t{%1, %0|%0, %1}";
else
- return "%vmovdqa\t{%1, %0|%0, %1}";
+ return "movdqa\t{%1, %0|%0, %1}";
case MODE_DI:
- return "%vmovq\t{%1, %0|%0, %1}";
+ return "movq\t{%1, %0|%0, %1}";
case MODE_DF:
- if (TARGET_AVX)
- {
- if (REG_P (operands[0]) && REG_P (operands[1]))
- return "vmovsd\t{%1, %0, %0|%0, %0, %1}";
- else
- return "vmovsd\t{%1, %0|%0, %1}";
- }
- else
- return "movsd\t{%1, %0|%0, %1}";
+ return "movsd\t{%1, %0|%0, %1}";
case MODE_V1DF:
- return "%vmovlpd\t{%1, %d0|%d0, %1}";
+ return "movlpd\t{%1, %0|%0, %1}";
case MODE_V2SF:
- return "%vmovlps\t{%1, %d0|%d0, %1}";
+ return "movlps\t{%1, %0|%0, %1}";
default:
gcc_unreachable ();
}
- case 9:
- case 10:
- return "%vmovd\t{%1, %0|%0, %1}";
-
default:
gcc_unreachable();
}
}
- [(set_attr "type" "fmov,fmov,fmov,multi,multi,sselog1,ssemov,ssemov,ssemov,ssemov,ssemov")
- (set (attr "prefix")
- (if_then_else (eq_attr "alternative" "0,1,2,3,4")
- (const_string "orig")
- (const_string "maybe_vex")))
+ [(set_attr "type" "fmov,fmov,fmov,multi,multi,sselog1,ssemov,ssemov,ssemov")
(set (attr "prefix_data16")
(if_then_else (eq_attr "mode" "V1DF")
(const_string "1")
@@ -3095,8 +3145,8 @@
(set (attr "mode")
(cond [(eq_attr "alternative" "0,1,2")
(const_string "DF")
- (eq_attr "alternative" "3,4,9,10")
- (const_string "DI")
+ (eq_attr "alternative" "3,4")
+ (const_string "SI")
/* For SSE1, we have many fewer alternatives. */
(eq (symbol_ref "TARGET_SSE2") (const_int 0))
@@ -3143,21 +3193,29 @@
]
(const_string "DF")))])
-(define_insn "*movdf_integer"
+;; Moving is usually shorter when only FP registers are used. This separate
+;; movdf pattern avoids the use of integer registers for FP operations
+;; when optimizing for size.
+
+(define_insn "*movdf_internal_nointeger"
[(set (match_operand:DF 0 "nonimmediate_operand"
- "=f,m,f,r ,o ,Y2*x,Y2*x,Y2*x,m ")
+ "=f,m,f,*r ,o ,Y2*x,Y2*x,Y2*x ,m ")
(match_operand:DF 1 "general_operand"
- "fm,f,G,roF,Fr,C ,Y2*x,m ,Y2*x"))]
+ "fm,f,G,*roF,*Fr,C ,Y2*x,mY2*x,Y2*x"))]
"!(MEM_P (operands[0]) && MEM_P (operands[1]))
- && optimize_function_for_speed_p (cfun)
- && TARGET_INTEGER_DFMODE_MOVES
+ && ((optimize_function_for_size_p (cfun)
+ || !TARGET_INTEGER_DFMODE_MOVES) && !TARGET_64BIT)
&& (reload_in_progress || reload_completed
|| (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
|| (!(TARGET_SSE2 && TARGET_SSE_MATH)
&& optimize_function_for_size_p (cfun)
+ && !memory_operand (operands[0], DFmode)
&& standard_80387_constant_p (operands[1]))
|| GET_CODE (operands[1]) != CONST_DOUBLE
- || memory_operand (operands[0], DFmode))"
+ || ((optimize_function_for_size_p (cfun)
+ || !TARGET_MEMORY_MISMATCH_STALL
+ || reload_in_progress || reload_completed)
+ && memory_operand (operands[0], DFmode)))"
{
switch (which_alternative)
{
@@ -3171,22 +3229,21 @@
case 3:
case 4:
return "#";
-
case 5:
switch (get_attr_mode (insn))
{
case MODE_V4SF:
- return "xorps\t%0, %0";
+ return "%vxorps\t%0, %d0";
case MODE_V2DF:
if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
- return "xorps\t%0, %0";
+ return "%vxorps\t%0, %d0";
else
- return "xorpd\t%0, %0";
+ return "%vxorpd\t%0, %d0";
case MODE_TI:
if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
- return "xorps\t%0, %0";
+ return "%vxorps\t%0, %d0";
else
- return "pxor\t%0, %0";
+ return "%vpxor\t%0, %d0";
default:
gcc_unreachable ();
}
@@ -3196,34 +3253,62 @@
switch (get_attr_mode (insn))
{
case MODE_V4SF:
- return "movaps\t{%1, %0|%0, %1}";
+ return "%vmovaps\t{%1, %0|%0, %1}";
case MODE_V2DF:
if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
- return "movaps\t{%1, %0|%0, %1}";
+ return "%vmovaps\t{%1, %0|%0, %1}";
else
- return "movapd\t{%1, %0|%0, %1}";
+ return "%vmovapd\t{%1, %0|%0, %1}";
case MODE_TI:
if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
- return "movaps\t{%1, %0|%0, %1}";
+ return "%vmovaps\t{%1, %0|%0, %1}";
else
- return "movdqa\t{%1, %0|%0, %1}";
+ return "%vmovdqa\t{%1, %0|%0, %1}";
case MODE_DI:
- return "movq\t{%1, %0|%0, %1}";
+ return "%vmovq\t{%1, %0|%0, %1}";
case MODE_DF:
- return "movsd\t{%1, %0|%0, %1}";
+ if (TARGET_AVX)
+ {
+ if (REG_P (operands[0]) && REG_P (operands[1]))
+ return "vmovsd\t{%1, %0, %0|%0, %0, %1}";
+ else
+ return "vmovsd\t{%1, %0|%0, %1}";
+ }
+ else
+ return "movsd\t{%1, %0|%0, %1}";
case MODE_V1DF:
- return "movlpd\t{%1, %0|%0, %1}";
+ if (TARGET_AVX)
+ {
+ if (REG_P (operands[0]))
+ return "vmovlpd\t{%1, %0, %0|%0, %0, %1}";
+ else
+ return "vmovlpd\t{%1, %0|%0, %1}";
+ }
+ else
+ return "movlpd\t{%1, %0|%0, %1}";
case MODE_V2SF:
- return "movlps\t{%1, %0|%0, %1}";
+ if (TARGET_AVX)
+ {
+ if (REG_P (operands[0]))
+ return "vmovlps\t{%1, %0, %0|%0, %0, %1}";
+ else
+ return "vmovlps\t{%1, %0|%0, %1}";
+ }
+ else
+ return "movlps\t{%1, %0|%0, %1}";
default:
gcc_unreachable ();
}
default:
- gcc_unreachable();
+ gcc_unreachable ();
}
}
[(set_attr "type" "fmov,fmov,fmov,multi,multi,sselog1,ssemov,ssemov,ssemov")
+ (set (attr "prefix")
+ (if_then_else (eq_attr "alternative" "0,1,2,3,4")
+ (const_string "orig")
+ (const_string "maybe_vex")))
(set (attr "prefix_data16")
(if_then_else (eq_attr "mode" "V1DF")
(const_string "1")
@@ -3293,86 +3378,18 @@
[(const_int 0)]
"ix86_split_long_move (operands); DONE;")
-(define_insn "*swapdf"
- [(set (match_operand:DF 0 "fp_register_operand" "+f")
- (match_operand:DF 1 "fp_register_operand" "+f"))
- (set (match_dup 1)
- (match_dup 0))]
- "reload_completed || TARGET_80387"
-{
- if (STACK_TOP_P (operands[0]))
- return "fxch\t%1";
- else
- return "fxch\t%0";
-}
- [(set_attr "type" "fxch")
- (set_attr "mode" "DF")])
-
-(define_expand "movxf"
- [(set (match_operand:XF 0 "nonimmediate_operand" "")
- (match_operand:XF 1 "general_operand" ""))]
- ""
- "ix86_expand_move (XFmode, operands); DONE;")
-
-;; Size of pushdf is 3 (for sub) + 2 (for fstp) + memory operand size.
-;; Size of pushdf using integer instructions is 3+3*memory operand size
-;; Pushing using integer instructions is longer except for constants
-;; and direct memory references.
-;; (assuming that any given constant is pushed only once, but this ought to be
-;; handled elsewhere).
-
-(define_insn "*pushxf_nointeger"
- [(set (match_operand:XF 0 "push_operand" "=X,X,X")
- (match_operand:XF 1 "general_no_elim_operand" "f,Fo,*r"))]
- "optimize_function_for_size_p (cfun)"
-{
- /* This insn should be already split before reg-stack. */
- gcc_unreachable ();
-}
- [(set_attr "type" "multi")
- (set_attr "unit" "i387,*,*")
- (set_attr "mode" "XF,SI,SI")])
-
-(define_insn "*pushxf_integer"
- [(set (match_operand:XF 0 "push_operand" "=<,<")
- (match_operand:XF 1 "general_no_elim_operand" "f,ro"))]
- "optimize_function_for_speed_p (cfun)"
-{
- /* This insn should be already split before reg-stack. */
- gcc_unreachable ();
-}
- [(set_attr "type" "multi")
- (set_attr "unit" "i387,*")
- (set_attr "mode" "XF,SI")])
-
-(define_split
- [(set (match_operand 0 "push_operand" "")
- (match_operand 1 "general_operand" ""))]
- "reload_completed
- && (GET_MODE (operands[0]) == XFmode
- || GET_MODE (operands[0]) == DFmode)
- && !ANY_FP_REG_P (operands[1])"
- [(const_int 0)]
- "ix86_split_long_move (operands); DONE;")
-
-(define_split
- [(set (match_operand:XF 0 "push_operand" "")
- (match_operand:XF 1 "any_fp_register_operand" ""))]
- ""
- [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (match_dup 2)))
- (set (mem:XF (reg:P SP_REG)) (match_dup 1))]
- "operands[2] = GEN_INT (TARGET_128BIT_LONG_DOUBLE ? -16 : -12);")
-
-;; Do not use integer registers when optimizing for size
-(define_insn "*movxf_nointeger"
- [(set (match_operand:XF 0 "nonimmediate_operand" "=f,m,f,*r,o")
- (match_operand:XF 1 "general_operand" "fm,f,G,*roF,F*r"))]
- "optimize_function_for_size_p (cfun)
- && !(MEM_P (operands[0]) && MEM_P (operands[1]))
+(define_insn "*movsf_internal"
+ [(set (match_operand:SF 0 "nonimmediate_operand"
+ "=f,m,f,r ,m ,x,x,x ,m,!*y,!m,!*y,?Yi,?r,!*Ym,!r")
+ (match_operand:SF 1 "general_operand"
+ "fm,f,G,rmF,Fr,C,x,xm,x,m ,*y,*y ,r ,Yi,r ,*Ym"))]
+ "!(MEM_P (operands[0]) && MEM_P (operands[1]))
&& (reload_in_progress || reload_completed
- || standard_80387_constant_p (operands[1])
+ || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
+ || (!TARGET_SSE_MATH && optimize_function_for_size_p (cfun)
+ && standard_80387_constant_p (operands[1]))
|| GET_CODE (operands[1]) != CONST_DOUBLE
- || memory_operand (operands[0], XFmode))"
+ || memory_operand (operands[0], SFmode))"
{
switch (which_alternative)
{
@@ -3383,149 +3400,89 @@
case 2:
return standard_80387_constant_opcode (operands[1]);
- case 3: case 4:
- return "#";
- default:
- gcc_unreachable ();
- }
-}
- [(set_attr "type" "fmov,fmov,fmov,multi,multi")
- (set_attr "mode" "XF,XF,XF,SI,SI")])
-
-(define_insn "*movxf_integer"
- [(set (match_operand:XF 0 "nonimmediate_operand" "=f,m,f,r,o")
- (match_operand:XF 1 "general_operand" "fm,f,G,roF,Fr"))]
- "optimize_function_for_speed_p (cfun)
- && !(MEM_P (operands[0]) && MEM_P (operands[1]))
- && (reload_in_progress || reload_completed
- || GET_CODE (operands[1]) != CONST_DOUBLE
- || memory_operand (operands[0], XFmode))"
-{
- switch (which_alternative)
- {
- case 0:
- case 1:
- return output_387_reg_move (insn, operands);
-
- case 2:
- return standard_80387_constant_opcode (operands[1]);
-
- case 3: case 4:
- return "#";
-
- default:
- gcc_unreachable ();
- }
-}
- [(set_attr "type" "fmov,fmov,fmov,multi,multi")
- (set_attr "mode" "XF,XF,XF,SI,SI")])
-
-(define_expand "movtf"
- [(set (match_operand:TF 0 "nonimmediate_operand" "")
- (match_operand:TF 1 "nonimmediate_operand" ""))]
- "TARGET_SSE2"
-{
- ix86_expand_move (TFmode, operands);
- DONE;
-})
-
-(define_insn "*movtf_internal"
- [(set (match_operand:TF 0 "nonimmediate_operand" "=x,m,x,?r,?o")
- (match_operand:TF 1 "general_operand" "xm,x,C,roF,Fr"))]
- "TARGET_SSE2
- && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-{
- switch (which_alternative)
- {
- case 0:
- case 1:
+ case 3:
+ case 4:
+ return "mov{l}\t{%1, %0|%0, %1}";
+ case 5:
+ if (get_attr_mode (insn) == MODE_TI)
+ return "%vpxor\t%0, %d0";
+ else
+ return "%vxorps\t%0, %d0";
+ case 6:
if (get_attr_mode (insn) == MODE_V4SF)
return "%vmovaps\t{%1, %0|%0, %1}";
else
- return "%vmovdqa\t{%1, %0|%0, %1}";
- case 2:
- if (get_attr_mode (insn) == MODE_V4SF)
- return "%vxorps\t%0, %d0";
+ return "%vmovss\t{%1, %d0|%d0, %1}";
+ case 7:
+ if (TARGET_AVX)
+ return REG_P (operands[1]) ? "vmovss\t{%1, %0, %0|%0, %0, %1}"
+ : "vmovss\t{%1, %0|%0, %1}";
else
- return "%vpxor\t%0, %d0";
- case 3:
- case 4:
- return "#";
+ return "movss\t{%1, %0|%0, %1}";
+ case 8:
+ return "%vmovss\t{%1, %0|%0, %1}";
+
+ case 9: case 10: case 14: case 15:
+ return "movd\t{%1, %0|%0, %1}";
+ case 12: case 13:
+ return "%vmovd\t{%1, %0|%0, %1}";
+
+ case 11:
+ return "movq\t{%1, %0|%0, %1}";
+
default:
gcc_unreachable ();
}
}
- [(set_attr "type" "ssemov,ssemov,sselog1,*,*")
- (set_attr "prefix" "maybe_vex,maybe_vex,maybe_vex,*,*")
+ [(set_attr "type" "fmov,fmov,fmov,imov,imov,sselog1,ssemov,ssemov,ssemov,mmxmov,mmxmov,mmxmov,ssemov,ssemov,mmxmov,mmxmov")
+ (set (attr "prefix")
+ (if_then_else (eq_attr "alternative" "5,6,7,8,12,13")
+ (const_string "maybe_vex")
+ (const_string "orig")))
(set (attr "mode")
- (cond [(eq_attr "alternative" "0,2")
+ (cond [(eq_attr "alternative" "3,4,9,10")
+ (const_string "SI")
+ (eq_attr "alternative" "5")
(if_then_else
- (ne (symbol_ref "optimize_function_for_size_p (cfun)")
- (const_int 0))
- (const_string "V4SF")
- (const_string "TI"))
- (eq_attr "alternative" "1")
+ (and (and (ne (symbol_ref "TARGET_SSE_LOAD0_BY_PXOR")
+ (const_int 0))
+ (ne (symbol_ref "TARGET_SSE2")
+ (const_int 0)))
+ (eq (symbol_ref "optimize_function_for_size_p (cfun)")
+ (const_int 0)))
+ (const_string "TI")
+ (const_string "V4SF"))
+ /* For architectures resolving dependencies on
+ whole SSE registers use APS move to break dependency
+ chains, otherwise use short move to avoid extra work.
+
+ Do the same for architectures resolving dependencies on
+ the parts. While in DF mode it is better to always handle
+ just register parts, the SF mode is different due to lack
+ of instructions to load just part of the register. It is
+ better to maintain the whole registers in single format
+ to avoid problems on using packed logical operations. */
+ (eq_attr "alternative" "6")
(if_then_else
- (ior (ne (symbol_ref "TARGET_SSE_TYPELESS_STORES")
+ (ior (ne (symbol_ref "TARGET_SSE_PARTIAL_REG_DEPENDENCY")
(const_int 0))
- (ne (symbol_ref "optimize_function_for_size_p (cfun)")
+ (ne (symbol_ref "TARGET_SSE_SPLIT_REGS")
(const_int 0)))
(const_string "V4SF")
- (const_string "TI"))]
- (const_string "DI")))])
+ (const_string "SF"))
+ (eq_attr "alternative" "11")
+ (const_string "DI")]
+ (const_string "SF")))])
-(define_insn "*pushtf_sse"
- [(set (match_operand:TF 0 "push_operand" "=<,<,<")
- (match_operand:TF 1 "general_no_elim_operand" "x,Fo,*r"))]
- "TARGET_SSE2"
-{
- /* This insn should be already split before reg-stack. */
- gcc_unreachable ();
-}
- [(set_attr "type" "multi")
- (set_attr "unit" "sse,*,*")
- (set_attr "mode" "TF,SI,SI")])
-
(define_split
- [(set (match_operand:TF 0 "push_operand" "")
- (match_operand:TF 1 "general_operand" ""))]
- "TARGET_SSE2 && reload_completed
- && !SSE_REG_P (operands[1])"
- [(const_int 0)]
- "ix86_split_long_move (operands); DONE;")
-
-(define_split
- [(set (match_operand:TF 0 "push_operand" "")
- (match_operand:TF 1 "any_fp_register_operand" ""))]
- "TARGET_SSE2"
- [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (const_int -16)))
- (set (mem:TF (reg:P SP_REG)) (match_dup 1))]
- "")
-
-(define_split
- [(set (match_operand 0 "nonimmediate_operand" "")
- (match_operand 1 "general_operand" ""))]
- "reload_completed
- && !(MEM_P (operands[0]) && MEM_P (operands[1]))
- && GET_MODE (operands[0]) == XFmode
- && ! (ANY_FP_REG_P (operands[0]) ||
- (GET_CODE (operands[0]) == SUBREG
- && ANY_FP_REG_P (SUBREG_REG (operands[0]))))
- && ! (ANY_FP_REG_P (operands[1]) ||
- (GET_CODE (operands[1]) == SUBREG
- && ANY_FP_REG_P (SUBREG_REG (operands[1]))))"
- [(const_int 0)]
- "ix86_split_long_move (operands); DONE;")
-
-(define_split
[(set (match_operand 0 "register_operand" "")
(match_operand 1 "memory_operand" ""))]
"reload_completed
&& MEM_P (operands[1])
&& (GET_MODE (operands[0]) == TFmode
|| GET_MODE (operands[0]) == XFmode
- || GET_MODE (operands[0]) == SFmode
- || GET_MODE (operands[0]) == DFmode)
+ || GET_MODE (operands[0]) == DFmode
+ || GET_MODE (operands[0]) == SFmode)
&& (operands[2] = find_constant_src (insn))"
[(set (match_dup 0) (match_dup 2))]
{
@@ -3556,8 +3513,8 @@
&& MEM_P (operands[1])
&& (GET_MODE (operands[0]) == TFmode
|| GET_MODE (operands[0]) == XFmode
- || GET_MODE (operands[0]) == SFmode
- || GET_MODE (operands[0]) == DFmode)
+ || GET_MODE (operands[0]) == DFmode
+ || GET_MODE (operands[0]) == SFmode)
&& (operands[2] = find_constant_src (insn))"
[(set (match_dup 0) (match_dup 2))]
{
@@ -3581,21 +3538,6 @@
FAIL;
})
-(define_insn "swapxf"
- [(set (match_operand:XF 0 "register_operand" "+f")
- (match_operand:XF 1 "register_operand" "+f"))
- (set (match_dup 1)
- (match_dup 0))]
- "TARGET_80387"
-{
- if (STACK_TOP_P (operands[0]))
- return "fxch\t%1";
- else
- return "fxch\t%0";
-}
- [(set_attr "type" "fxch")
- (set_attr "mode" "XF")])
-
;; Split the load of -0.0 or -1.0 into fldz;fchs or fld1;fchs sequence
(define_split
[(set (match_operand:X87MODEF 0 "register_operand" "")
@@ -3616,13 +3558,35 @@
operands[1] = CONST1_RTX (<MODE>mode);
})
-(define_split
- [(set (match_operand:TF 0 "nonimmediate_operand" "")
- (match_operand:TF 1 "general_operand" ""))]
- "reload_completed
- && !(SSE_REG_P (operands[0]) || SSE_REG_P (operands[1]))"
- [(const_int 0)]
- "ix86_split_long_move (operands); DONE;")
+(define_insn "swapxf"
+ [(set (match_operand:XF 0 "register_operand" "+f")
+ (match_operand:XF 1 "register_operand" "+f"))
+ (set (match_dup 1)
+ (match_dup 0))]
+ "TARGET_80387"
+{
+ if (STACK_TOP_P (operands[0]))
+ return "fxch\t%1";
+ else
+ return "fxch\t%0";
+}
+ [(set_attr "type" "fxch")
+ (set_attr "mode" "XF")])
+
+(define_insn "*swap<mode>"
+ [(set (match_operand:MODEF 0 "fp_register_operand" "+f")
+ (match_operand:MODEF 1 "fp_register_operand" "+f"))
+ (set (match_dup 1)
+ (match_dup 0))]
+ "TARGET_80387 || reload_completed"
+{
+ if (STACK_TOP_P (operands[0]))
+ return "fxch\t%1";
+ else
+ return "fxch\t%0";
+}
+ [(set_attr "type" "fxch")
+ (set_attr "mode" "<MODE>")])
;; Zero extension instructions
Hello! Attached patch macroizes and reorganizes FP move patterns. 2010-06-23 Uros Bizjak <ubizjak@gmail.com> (mov<mode>): Macroize expander from mov{sf,df,xf} using X87MODEF mode iterator. (pushsf splitter): Macroize splitter using P mode iterator. (*swap<mode>): Macroize insn from *swap{sf,df} using MODEF mode iterator. (*movxf_internal): Rename from *movxf_integer. (*movxf_internal_nointeger): Rename from *movxf_nointeger. (*movdf_internal_rex64): Rename from *movdf_integer_rex64. (*movdf_internal): Rename from *movdf_integer. (*movdf_internal_nointeger): Rename from *movdf_nointeger. (*movsf_internal): Rename from *movdf_1. Patch was bootstrapped and regression tested on x86_64-pc-linux-gnu{,-m32}. Patch was committed to mainline. Uros,.