Hello,
This patch does some further improvements to the utilization of rotate
insns on SH. Tested on rev 191657 with
make -k check RUNTESTFLAGS="--target_board=sh-sim
\{-m2/-ml,-m2/-mb,-m2a/-mb,-m4/-ml,-m4/-mb,-m4a/-ml,-m4a/-mb}"
and no new failures.
OK?
(BTW, comparing test summaries of rev. 191342 and rev. 191657 shows a
lot of new PCH related failures on my xgcc SH setup)
Cheers,
Oleg
gcc/ChangeLog:
PR target/54089
* config/sh/constraints.md (Jhb): New constraint.
* config/sh/predicates.md (negt_reg_shl31_operand): New
predicate.
* config/sh/sh.md (rotrsi3): New expander.
(rotrsi3_1, *rotrsi3_1, *rotlsi3_1): New insns.
(rotlsi3, rotlhi3): Use const_int_operand predicate instead of
immediate_operand and remove CONST_INT_P checks in expansion
code.
(*rotcr): Cleanup variable usage. Handle preceding nott insn.
Add split with swapped operands.
(*rotcr_neg_t, *movt_msb, *negt_msb): New insns and splits.
testsuite/ChangeLog:
PR target/54089
* gcc.target/sh/pr54089-1.c (test_15, test_16, test_17, test_18,
test_19, test_20, test_21, test_22, test_23): New functions.
* gcc.target/sh/pr54089-4.c: New.
* gcc.target/sh/pr54089-5.c: New.
* gcc.target/sh/pr54089-6.c: New.
* gcc.target/sh/pr54089-7.c: New.
Index: gcc/testsuite/gcc.target/sh/pr54089-4.c
===================================================================
--- gcc/testsuite/gcc.target/sh/pr54089-4.c (revision 0)
+++ gcc/testsuite/gcc.target/sh/pr54089-4.c (revision 0)
@@ -0,0 +1,15 @@
+/* Check that the rotcr instruction is generated when shifting the
+ negated T bit on non-SH2A. */
+/* { dg-do compile { target "sh*-*-*" } } */
+/* { dg-options "-O1" } */
+/* { dg-skip-if "" { "sh*-*-*" } { "-m5*" "-m2a*" } { "" } } */
+/* { dg-final { scan-assembler-times "rotcr" 1 } } */
+/* { dg-final { scan-assembler-times "tst" 1 } } */
+/* { dg-final { scan-assembler-times "movt" 1 } } */
+
+int
+test_00 (int a, int b)
+{
+ int r = a != b;
+ return r << 31;
+}
Index: gcc/testsuite/gcc.target/sh/pr54089-6.c
===================================================================
--- gcc/testsuite/gcc.target/sh/pr54089-6.c (revision 0)
+++ gcc/testsuite/gcc.target/sh/pr54089-6.c (revision 0)
@@ -0,0 +1,30 @@
+/* Check that the rotr and rotl instructions are generated. */
+/* { dg-do compile { target "sh*-*-*" } } */
+/* { dg-options "-O1" } */
+/* { dg-skip-if "" { "sh*-*-*" } { "-m5*" } { "" } } */
+/* { dg-final { scan-assembler-times "rotr" 2 } } */
+/* { dg-final { scan-assembler-times "rotl" 2 } } */
+
+int
+test_00 (int a)
+{
+ return (a << 1) | ((a >> 31) & 1);
+}
+
+int
+test_01 (int a)
+{
+ return (a << 1) | ((unsigned int)a >> 31);
+}
+
+int
+test_02 (int a)
+{
+ return ((unsigned int)a >> 1) | (a << 31);
+}
+
+int
+test_03 (int a)
+{
+ return ((a >> 1) & 0x7FFFFFFF) | (a << 31);
+}
Index: gcc/testsuite/gcc.target/sh/pr54089-1.c
===================================================================
--- gcc/testsuite/gcc.target/sh/pr54089-1.c (revision 191657)
+++ gcc/testsuite/gcc.target/sh/pr54089-1.c (working copy)
@@ -2,7 +2,7 @@
/* { dg-do compile { target "sh*-*-*" } } */
/* { dg-options "-O1" } */
/* { dg-skip-if "" { "sh*-*-*" } { "-m5*"} { "" } } */
-/* { dg-final { scan-assembler-times "rotcr" 15 } } */
+/* { dg-final { scan-assembler-times "rotcr" 24 } } */
/* { dg-final { scan-assembler-times "shll\t" 1 } } */
typedef char bool;
@@ -109,3 +109,66 @@
bool r = b < 0;
return ((a >> 1) | (r << 31));
}
+
+unsigned int
+test_15 (unsigned int a, int b, int c)
+{
+ bool r = b != c;
+ return ((a >> 1) | (r << 31));
+}
+
+unsigned int
+test_16 (unsigned int a, int b, int c)
+{
+ bool r = b != c;
+ return ((a >> 2) | (r << 31));
+}
+
+unsigned int
+test_17 (unsigned int a, int b, int c)
+{
+ bool r = b != c;
+ return ((a >> 3) | (r << 31));
+}
+
+unsigned int
+test_18 (unsigned int a, int b, int c)
+{
+ bool r = b != c;
+ return ((a >> 4) | (r << 31));
+}
+
+unsigned int
+test_19 (unsigned int a, int b, int c)
+{
+ bool r = b != c;
+ return ((a >> 5) | (r << 31));
+}
+
+unsigned int
+test_20 (unsigned int a, int b, int c)
+{
+ bool r = b != c;
+ return ((a >> 6) | (r << 31));
+}
+
+unsigned int
+test_21 (unsigned int a, int b, int c)
+{
+ bool r = b != c;
+ return ((a >> 7) | (r << 31));
+}
+
+unsigned int
+test_22 (unsigned int a, int b, int c)
+{
+ bool r = b != c;
+ return ((a >> 8) | (r << 31));
+}
+
+unsigned int
+test_23 (unsigned int a, int b, int c)
+{
+ bool r = b != c;
+ return ((a >> 31) | (r << 31));
+}
Index: gcc/testsuite/gcc.target/sh/pr54089-5.c
===================================================================
--- gcc/testsuite/gcc.target/sh/pr54089-5.c (revision 0)
+++ gcc/testsuite/gcc.target/sh/pr54089-5.c (revision 0)
@@ -0,0 +1,14 @@
+/* Check that the movrt rotr instruction sequence is generated when shifting
+ the negated T bit on SH2A. */
+/* { dg-do compile { target "sh*-*-*" } } */
+/* { dg-options "-O1" } */
+/* { dg-skip-if "" { "sh*-*-*" } { "*" } { "-m2a*" } } */
+/* { dg-final { scan-assembler-times "movrt" 1 } } */
+/* { dg-final { scan-assembler-times "rotr" 1 } } */
+
+int
+test_00 (int a, int b)
+{
+ int r = a != b;
+ return r << 31;
+}
Index: gcc/testsuite/gcc.target/sh/pr54089-7.c
===================================================================
--- gcc/testsuite/gcc.target/sh/pr54089-7.c (revision 0)
+++ gcc/testsuite/gcc.target/sh/pr54089-7.c (revision 0)
@@ -0,0 +1,63 @@
+/* Check that the rotcr instruction is generated. */
+/* { dg-do compile { target "sh*-*-*" } } */
+/* { dg-options "-O1" } */
+/* { dg-skip-if "" { "sh*-*-*" } { "-m5*"} { "" } } */
+/* { dg-final { scan-assembler-times "rotcr" 4 } } */
+/* { dg-final { scan-assembler-not "movt" } } */
+/* { dg-final { scan-assembler-not "or\t" } } */
+/* { dg-final { scan-assembler-not "rotr" } } */
+/* { dg-final { scan-assembler-not "and" } } */
+
+typedef char bool;
+
+int
+test_00 (int* a, int* b)
+{
+ int i;
+ unsigned int r = 0;
+ for (i = 0; i < 16; ++i)
+ {
+ bool t = a[i] == b[i];
+ r = (t << 31) | (r >> 1);
+ }
+ return r;
+}
+
+int
+test_01 (int* a, int* b)
+{
+ int i;
+ unsigned int r = 0;
+ for (i = 0; i < 16; ++i)
+ {
+ bool t = a[i] == b[i];
+ r = (t << 31) | (r >> 2);
+ }
+ return r;
+}
+
+int
+test_02 (int* a, int* b)
+{
+ int i;
+ unsigned int r = 0;
+ for (i = 0; i < 16; ++i)
+ {
+ bool t = a[i] == b[i];
+ r = (t << 31) | (r >> 3);
+ }
+ return r;
+}
+
+unsigned int
+test_03 (const bool* a)
+{
+ int i;
+ unsigned int r = 0;
+ for (i = 0; i < 32; ++i)
+ {
+ bool t = a[i];
+ r = (t << 31) | (r >> 1);
+ }
+ return r;
+}
Index: gcc/config/sh/sh.md
===================================================================
--- gcc/config/sh/sh.md (revision 191657)
+++ gcc/config/sh/sh.md (working copy)
@@ -3817,6 +3817,42 @@
GEN_INT (56), GEN_INT (8));
})
+(define_expand "rotrsi3"
+ [(set (match_operand:SI 0 "arith_reg_dest")
+ (rotatert:SI (match_operand:SI 1 "arith_reg_operand")
+ (match_operand:SI 2 "const_int_operand")))]
+ "TARGET_SH1"
+{
+ HOST_WIDE_INT ival = INTVAL (operands[2]);
+ if (ival == 1)
+ {
+ emit_insn (gen_rotrsi3_1 (operands[0], operands[1]));
+ DONE;
+ }
+
+ FAIL;
+})
+
+(define_insn "rotrsi3_1"
+ [(set (match_operand:SI 0 "arith_reg_dest" "=r")
+ (rotatert:SI (match_operand:SI 1 "arith_reg_operand" "0")
+ (const_int 1)))
+ (set (reg:SI T_REG)
+ (and:SI (match_dup 1) (const_int 1)))]
+ "TARGET_SH1"
+ "rotr %0"
+ [(set_attr "type" "arith")])
+
+;; A slimplified version of rotr for combine.
+(define_insn "*rotrsi3_1"
+ [(set (match_operand:SI 0 "arith_reg_dest" "=r")
+ (rotatert:SI (match_operand:SI 1 "arith_reg_operand" "0")
+ (const_int 1)))
+ (clobber (reg:SI T_REG))]
+ "TARGET_SH1"
+ "rotr %0"
+ [(set_attr "type" "arith")])
+
(define_insn "rotlsi3_1"
[(set (match_operand:SI 0 "arith_reg_dest" "=r")
(rotate:SI (match_operand:SI 1 "arith_reg_operand" "0")
@@ -3827,6 +3863,16 @@
"rotl %0"
[(set_attr "type" "arith")])
+;; A simplified version of rotl for combine.
+(define_insn "*rotlsi3_1"
+ [(set (match_operand:SI 0 "arith_reg_dest" "=r")
+ (rotate:SI (match_operand:SI 1 "arith_reg_operand" "0")
+ (const_int 1)))
+ (clobber (reg:SI T_REG))]
+ "TARGET_SH1"
+ "rotl %0"
+ [(set_attr "type" "arith")])
+
(define_insn "rotlsi3_31"
[(set (match_operand:SI 0 "arith_reg_dest" "=r")
(rotate:SI (match_operand:SI 1 "arith_reg_operand" "0")
@@ -3845,9 +3891,9 @@
[(set_attr "type" "arith")])
(define_expand "rotlsi3"
- [(set (match_operand:SI 0 "arith_reg_dest" "")
- (rotate:SI (match_operand:SI 1 "arith_reg_operand" "")
- (match_operand:SI 2 "immediate_operand" "")))]
+ [(set (match_operand:SI 0 "arith_reg_dest")
+ (rotate:SI (match_operand:SI 1 "arith_reg_operand")
+ (match_operand:SI 2 "const_int_operand")))]
"TARGET_SH1"
{
static const char rot_tab[] = {
@@ -3857,12 +3903,8 @@
002, 002, 010, 000, 000, 000, 000, 000,
};
- int count, choice;
-
- if (!CONST_INT_P (operands[2]))
- FAIL;
- count = INTVAL (operands[2]);
- choice = rot_tab[count];
+ int count = INTVAL (operands[2]);
+ int choice = rot_tab[count];
if (choice & 010 && SH_DYNAMIC_SHIFT_COST <= 1)
FAIL;
choice &= 7;
@@ -3908,12 +3950,12 @@
[(set_attr "type" "arith")])
(define_expand "rotlhi3"
- [(set (match_operand:HI 0 "arith_reg_operand" "")
- (rotate:HI (match_operand:HI 1 "arith_reg_operand" "")
- (match_operand:HI 2 "immediate_operand" "")))]
+ [(set (match_operand:HI 0 "arith_reg_operand")
+ (rotate:HI (match_operand:HI 1 "arith_reg_operand")
+ (match_operand:HI 2 "const_int_operand")))]
"TARGET_SH1"
{
- if (!CONST_INT_P (operands[2]) || INTVAL (operands[2]) != 8)
+ if (INTVAL (operands[2]) != 8)
FAIL;
})
@@ -3950,11 +3992,7 @@
{
if (INTVAL (operands[2]) > 1)
{
- /* use plus_constant function ?? */
- const int shift_count = INTVAL (operands[2]) - 1;
- const rtx shift_count_rtx = GEN_INT (shift_count);
- rtx shift_res = gen_reg_rtx (SImode);
-
+ const rtx shift_count = GEN_INT (INTVAL (operands[2]) - 1);
rtx prev_set_t_insn = NULL_RTX;
rtx tmp_t_reg = NULL_RTX;
@@ -3963,10 +4001,24 @@
shift insn before that insn, to remove the T_REG dependency.
If the insn that sets the T_REG cannot be found, store the T_REG
in a temporary reg and restore it after the shift. */
- if (sh_lshrsi_clobbers_t_reg_p (shift_count_rtx)
- && ! sh_dynamicalize_shift_p (shift_count_rtx))
+ if (sh_lshrsi_clobbers_t_reg_p (shift_count)
+ && ! sh_dynamicalize_shift_p (shift_count))
{
prev_set_t_insn = prev_nonnote_insn_bb (curr_insn);
+
+ /* Skip the nott insn, which was probably inserted by the splitter
+ of *rotcr_neg_t. Don't use one of the recog functions
+ here during insn splitting, since that causes problems in later
+ passes. */
+ if (prev_set_t_insn != NULL_RTX)
+ {
+ rtx pat = PATTERN (prev_set_t_insn);
+ if (GET_CODE (pat) == SET
+ && t_reg_operand (XEXP (pat, 0), SImode)
+ && negt_reg_operand (XEXP (pat, 1), SImode))
+ prev_set_t_insn = prev_nonnote_insn_bb (prev_set_t_insn);
+ }
+
if (! (prev_set_t_insn != NULL_RTX
&& reg_set_p (get_t_reg_rtx (), prev_set_t_insn)
&& ! reg_referenced_p (get_t_reg_rtx (),
@@ -3978,14 +4030,15 @@
}
}
- rtx shift_rtx = gen_lshrsi3 (shift_res, operands[1], shift_count_rtx);
- operands[1] = shift_res;
+ rtx shift_result = gen_reg_rtx (SImode);
+ rtx shift_insn = gen_lshrsi3 (shift_result, operands[1], shift_count);
+ operands[1] = shift_result;
/* Emit the shift insn before the insn that sets T_REG, if possible. */
if (prev_set_t_insn != NULL_RTX)
- emit_insn_before (shift_rtx, prev_set_t_insn);
+ emit_insn_before (shift_insn, prev_set_t_insn);
else
- emit_insn (shift_rtx);
+ emit_insn (shift_insn);
/* Restore T_REG if it has been saved before. */
if (tmp_t_reg != NULL_RTX)
@@ -4007,6 +4060,20 @@
DONE;
})
+;; If combine tries the same as above but with swapped operands, split
+;; it so that it will try the pattern above.
+(define_split
+ [(set (match_operand:SI 0 "arith_reg_dest")
+ (ior:SI (ashift:SI (match_operand:SI 1 "arith_reg_or_t_reg_operand")
+ (const_int 31))
+ (lshiftrt:SI (match_operand:SI 2 "arith_reg_operand")
+ (match_operand:SI 3 "const_int_operand"))))]
+ "TARGET_SH1 && can_create_pseudo_p ()"
+ [(parallel [(set (match_dup 0)
+ (ior:SI (lshiftrt:SI (match_dup 2) (match_dup 3))
+ (ashift:SI (match_dup 1) (const_int 31))))
+ (clobber (reg:SI T_REG))])])
+
;; rotcr combine bridge pattern which will make combine try out more
;; complex patterns.
(define_insn_and_split "*rotcr"
@@ -4040,6 +4107,41 @@
DONE;
})
+;; rotcr combine patterns for rotating in the negated T_REG value.
+(define_insn_and_split "*rotcr_neg_t"
+ [(set (match_operand:SI 0 "arith_reg_dest")
+ (ior:SI (match_operand:SI 1 "negt_reg_shl31_operand")
+ (lshiftrt:SI (match_operand:SI 2 "arith_reg_operand")
+ (match_operand:SI 3 "const_int_operand"))))
+ (clobber (reg:SI T_REG))]
+ "TARGET_SH1"
+ "#"
+ "&& can_create_pseudo_p ()"
+ [(parallel [(set (match_dup 0)
+ (ior:SI (lshiftrt:SI (match_dup 2) (match_dup 3))
+ (ashift:SI (reg:SI T_REG) (const_int 31))))
+ (clobber (reg:SI T_REG))])]
+{
+ emit_insn (gen_nott (get_t_reg_rtx ()));
+})
+
+(define_insn_and_split "*rotcr_neg_t"
+ [(set (match_operand:SI 0 "arith_reg_dest")
+ (ior:SI (lshiftrt:SI (match_operand:SI 1 "arith_reg_operand")
+ (match_operand:SI 2 "const_int_operand"))
+ (match_operand:SI 3 "negt_reg_shl31_operand")))
+ (clobber (reg:SI T_REG))]
+ "TARGET_SH1"
+ "#"
+ "&& can_create_pseudo_p ()"
+ [(parallel [(set (match_dup 0)
+ (ior:SI (lshiftrt:SI (match_dup 1) (match_dup 2))
+ (ashift:SI (reg:SI T_REG) (const_int 31))))
+ (clobber (reg:SI T_REG))])]
+{
+ emit_insn (gen_nott (get_t_reg_rtx ()));
+})
+
;; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
;; SImode shift left
@@ -10720,6 +10822,53 @@
operands[0] = gen_reg_rtx (SImode);
})
+;; Store T bit as MSB in a reg.
+;; T = 0: 0x00000000 -> reg
+;; T = 1: 0x80000000 -> reg
+(define_insn_and_split "*movt_msb"
+ [(set (match_operand:SI 0 "arith_reg_dest")
+ (mult:SI (match_operand:SI 1 "t_reg_operand")
+ (const_int -2147483648))) ;; 0xffffffff80000000
+ (clobber (reg:SI T_REG))]
+ "TARGET_SH1"
+ "#"
+ "&& 1"
+ [(set (match_dup 0) (ashift:SI (reg:SI T_REG) (const_int 31)))])
+
+;; Store inverted T bit as MSB in a reg.
+;; T = 0: 0x80000000 -> reg
+;; T = 1: 0x00000000 -> reg
+;; On SH2A we can get away without clobbering the T_REG.
+(define_insn_and_split "*negt_msb"
+ [(set (match_operand:SI 0 "arith_reg_dest")
+ (match_operand:SI 1 "negt_reg_shl31_operand"))]
+ "TARGET_SH2A"
+ "#"
+ "&& can_create_pseudo_p ()"
+ [(const_int 0)]
+{
+ rtx tmp = gen_reg_rtx (SImode);
+ emit_insn (gen_movrt (tmp, get_t_reg_rtx ()));
+ emit_insn (gen_rotrsi3 (operands[0], tmp, const1_rtx));
+ DONE;
+})
+
+(define_insn_and_split "*negt_msb"
+ [(set (match_operand:SI 0 "arith_reg_dest")
+ (match_operand:SI 1 "negt_reg_shl31_operand"))
+ (clobber (reg:SI T_REG))]
+ "TARGET_SH1 && !TARGET_SH2A"
+ "#"
+ "&& can_create_pseudo_p ()"
+ [(const_int 0)]
+{
+ rtx tmp = gen_reg_rtx (SImode);
+ emit_move_insn (tmp, get_t_reg_rtx ());
+ emit_insn (gen_cmpeqsi_t (tmp, const0_rtx));
+ emit_insn (gen_rotcr (operands[0], tmp, get_t_reg_rtx ()));
+ DONE;
+})
+
;; The *cset_zero patterns convert optimizations such as
;; "if (test) x = 0;" to "x &= -(test == 0);"
;; back to conditional branch sequences if zero-displacement branches
Index: gcc/config/sh/predicates.md
===================================================================
--- gcc/config/sh/predicates.md (revision 191657)
+++ gcc/config/sh/predicates.md (working copy)
@@ -1033,3 +1033,41 @@
(define_predicate "arith_reg_or_t_reg_operand"
(ior (match_operand 0 "arith_reg_operand")
(match_operand 0 "t_reg_operand")))
+
+;; A predicate describing the negated value of the T bit register shifted
+;; left by 31.
+(define_predicate "negt_reg_shl31_operand"
+ (match_code "plus,minus,if_then_else")
+{
+ /* (plus:SI (mult:SI (match_operand:SI 1 "t_reg_operand")
+ (const_int -2147483648)) ;; 0xffffffff80000000
+ (const_int -2147483648))
+ */
+ if (GET_CODE (op) == PLUS && satisfies_constraint_Jhb (XEXP (op, 1))
+ && GET_CODE (XEXP (op, 0)) == MULT
+ && t_reg_operand (XEXP (XEXP (op, 0), 0), SImode)
+ && satisfies_constraint_Jhb (XEXP (XEXP (op, 0), 1)))
+ return true;
+
+ /* (minus:SI (const_int -2147483648) ;; 0xffffffff80000000
+ (mult:SI (match_operand:SI 1 "t_reg_operand")
+ (const_int -2147483648)))
+ */
+ if (GET_CODE (op) == MINUS
+ && satisfies_constraint_Jhb (XEXP (op, 0))
+ && GET_CODE (XEXP (op, 1)) == MULT
+ && t_reg_operand (XEXP (XEXP (op, 1), 0), SImode)
+ && satisfies_constraint_Jhb (XEXP (XEXP (op, 1), 1)))
+ return true;
+
+ /* (if_then_else:SI (match_operand:SI 1 "t_reg_operand")
+ (const_int 0)
+ (const_int -2147483648)) ;; 0xffffffff80000000
+ */
+ if (GET_CODE (op) == IF_THEN_ELSE && t_reg_operand (XEXP (op, 0), SImode)
+ && satisfies_constraint_Z (XEXP (op, 1))
+ && satisfies_constraint_Jhb (XEXP (op, 2)))
+ return true;
+
+ return false;
+})
Index: gcc/config/sh/constraints.md
===================================================================
--- gcc/config/sh/constraints.md (revision 191657)
+++ gcc/config/sh/constraints.md (working copy)
@@ -33,6 +33,7 @@
;; J16: 0xffffffff00000000 | 0x00000000ffffffff
;; Jmb: 0x000000FF
;; Jmw: 0x0000FFFF
+;; Jhb: 0x80000000
;; Kxx: unsigned xx bit
;; M: 1
;; N: 0
@@ -148,6 +149,11 @@
(and (match_code "const_int")
(match_test "ival == 0xFFFF")))
+(define_constraint "Jhb"
+ "Highest bit constant"
+ (and (match_code "const_int")
+ (match_test "(ival & 0xFFFFFFFF) == 0x80000000")))
+
(define_constraint "K03"
"An unsigned 3-bit constant, as used in SH2A bclr, bset, etc."
(and (match_code "const_int")
