i386: Optimization for APX NDD is always zero-uppered for shift

gcc/ChangeLog:

	PR target/113729
	* config/i386/i386.md (*ashlqi3_1_zext<mode><nf_name>):
	New define_insn.
	(*ashlhi3_1_zext<mode><nf_name>): Ditto.
	(*<insn>qi3_1_zext<mode><nf_name>): Ditto.
	(*<insn>hi3_1_zext<mode><nf_name>): Ditto.

gcc/testsuite/ChangeLog:

	* gcc.target/i386/pr113729.c: Add testcase for shift and rotate.

diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index 9f933e908d3d58ff4c04a4a15afbab6015bd16ac..d3ba2425f16e3937cac576acdebff9266d9ac03d 100644 (file)
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -15919,6 +15919,28 @@
   [(set_attr "type" "ishiftx")
    (set_attr "mode" "SI")])
 
+(define_insn "*ashlqi3_1_zext<mode><nf_name>"
+  [(set (match_operand:SWI248x 0 "register_operand" "=r")
+       (zero_extend:SWI248x
+         (ashift:QI (match_operand:QI 1 "nonimmediate_operand" "rm")
+                    (match_operand:QI 2 "nonmemory_operand" "cI"))))]
+  "TARGET_APX_NDD && <nf_condition>"
+  "<nf_prefix>sal{b}\t{%2, %1, %b0|%b0, %1, %2}"
+  [(set_attr "type" "ishiftx")
+   (set_attr "has_nf" "1")
+   (set_attr "mode" "QI")])
+
+(define_insn "*ashlhi3_1_zext<mode><nf_name>"
+  [(set (match_operand:SWI48x 0 "register_operand" "=r")
+       (zero_extend:SWI48x
+         (ashift:HI (match_operand:HI 1 "nonimmediate_operand" "rm")
+                    (match_operand:QI 2 "nonmemory_operand" "cI"))))]
+  "TARGET_APX_NDD && <nf_condition>"
+  "<nf_prefix>sal{w}\t{%2, %1, %w0|%w0, %1, %2}"
+  [(set_attr "type" "ishiftx")
+   (set_attr "has_nf" "1")
+   (set_attr "mode" "HI")])
+
 (define_insn "*ashlsi3_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r,r,r,r")
        (zero_extend:DI
@@ -17403,6 +17425,28 @@
   [(set_attr "type" "ishiftx")
    (set_attr "mode" "SI")])
 
+(define_insn "*<insn>qi3_1_zext<mode><nf_name>"
+  [(set (match_operand:SWI248x 0 "register_operand" "=r")
+       (zero_extend:SWI248x
+         (any_shiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "rm")
+                         (match_operand:QI 2 "nonmemory_operand" "cI"))))]
+  "TARGET_APX_NDD && <nf_condition>"
+  "<nf_prefix><shift>{b}\t{%2, %1, %b0|%b0, %1, %2}"
+  [(set_attr "type" "ishift")
+   (set_attr "has_nf" "1")
+   (set_attr "mode" "QI")])
+
+(define_insn "*<insn>hi3_1_zext<mode><nf_name>"
+  [(set (match_operand:SWI48x 0 "register_operand" "=r")
+       (zero_extend:SWI48x
+         (any_shiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "rm")
+                         (match_operand:QI 2 "nonmemory_operand" "cI"))))]
+  "TARGET_APX_NDD && <nf_condition>"
+  "<nf_prefix><shift>{w}\t{%2, %1, %w0|%w0, %1, %2}"
+  [(set_attr "type" "ishift")
+   (set_attr "has_nf" "1")
+   (set_attr "mode" "HI")])
+
 (define_insn "*<insn>si3_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r,r,r,?k")
        (zero_extend:DI
@@ -18179,6 +18223,28 @@
   [(set_attr "type" "rotatex")
    (set_attr "mode" "SI")])
 
+(define_insn "*<insn>qi3_1_zext<mode><nf_name>"
+  [(set (match_operand:SWI248x 0 "register_operand" "=r")
+       (zero_extend:SWI248x
+         (any_rotate:QI (match_operand:QI 1 "nonimmediate_operand" "rm")
+                        (match_operand:QI 2 "nonmemory_operand" "cI"))))]
+  "TARGET_APX_NDD && <nf_condition>"
+  "<nf_prefix><rotate>{b}\t{%2, %1, %b0|%b0, %1, %2}"
+  [(set_attr "type" "rotate")
+   (set_attr "has_nf" "1")
+   (set_attr "mode" "QI")])
+
+(define_insn "*<insn>hi3_1_zext<mode><nf_name>"
+  [(set (match_operand:SWI48x 0 "register_operand" "=r")
+       (zero_extend:SWI48x
+         (any_rotate:HI (match_operand:HI 1 "nonimmediate_operand" "rm")
+                        (match_operand:QI 2 "nonmemory_operand" "cI"))))]
+  "TARGET_APX_NDD && <nf_condition>"
+  "<nf_prefix><rotate>{w}\t{%2, %1, %w0|%w0, %1, %2}"
+  [(set_attr "type" "rotate")
+   (set_attr "has_nf" "1")
+   (set_attr "mode" "HI")])
+
 (define_insn "*<insn>si3_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r,r,r")
        (zero_extend:DI

diff --git a/gcc/testsuite/gcc.target/i386/pr113729.c b/gcc/testsuite/gcc.target/i386/pr113729.c
index d17e0999e6c0981888c9277aabfcf01c2d597767..53b6df78c26b6a00fe4829a5e4298ef72270e660 100644 (file)
--- a/gcc/testsuite/gcc.target/i386/pr113729.c
+++ b/gcc/testsuite/gcc.target/i386/pr113729.c
@@ -71,3 +71,65 @@ F (int, char, xor, ^)
 F (int64_t, char, xor, ^)
 F (int, short, xor, ^)
 F (int64_t, short, xor, ^)
+
+#define F2(TYPE1,TYPE2, OP_NAME, OP, IMM)  \
+TYPE1                                      \
+__attribute__ ((noipa))                    \
+f2_##OP_NAME##_##TYPE1##_##TYPE2 (TYPE2 a) \
+{                                          \
+  unsigned TYPE2 b = a OP IMM;             \
+  return b;                                \
+}                      
+
+/* ashlqi3_1_zext<mode> */
+F2 (short, char, shl, <<, 7)
+F2 (int, char, shl, <<, 6)
+F2 (int64_t, char, shl, <<, 7)
+F2 (int, short, shl, <<, 6)
+F2 (int64_t, short, shl, <<, 3)
+
+/* ashrqi3_1_zext<mode> */
+F2 (short, char, sar, >>, 7)
+F2 (int, char, sar, >>, 6)
+F2 (int64_t, char, sar, >>, 7)
+F2 (int, short, sar, >>, 6)
+F2 (int64_t, short, sar, >>, 3)
+
+#define F3(TYPE1,TYPE2, OP_NAME, OP, IMM)   \
+TYPE1                                       \
+__attribute__ ((noipa))                     \
+f3_##OP_NAME##_##TYPE1##_##TYPE2 (TYPE2 a)  \
+{                                           \
+  TYPE2 b = a OP IMM;                       \
+  return b;                                 \
+}                      
+
+/* lshrhi3_1_zext<mode> */
+F3 (short, uint8_t, shr, >>, 7)
+F3 (int, uint8_t, shr, >>, 6)
+F3 (int64_t, uint8_t, shr, >>, 7)
+F3 (int, uint16_t, shr, >>, 6)
+F3 (int64_t, uint16_t, shr, >>, 3)
+
+#define F4(TYPE1,TYPE2, OP_NAME, OP1, OP2, IMM1)                      \
+TYPE1                                                                 \
+__attribute__ ((noipa))                                               \
+foo4_##OP_NAME##_##TYPE1##_##TYPE2 (unsigned TYPE2 a)                 \
+{                                                                     \
+  unsigned TYPE2 b = (a OP1 IMM1 | a OP2 (8 * sizeof(TYPE2) - IMM1)); \
+  return b;                                                           \
+}
+
+/* rotrqi3_1_zext<mode> */
+F4 (short, char, ror, >>, <<, 1)
+F4 (int, char, ror, >>, <<, 1)
+F4 (long, char, ror, >>, <<, 1)
+F4 (int, short, ror, >>, <<, 1)
+F4 (long, short, ror, >>, <<, 1)
+
+/* rotlqi3_1_zext<mode> */
+F4 (short, char, rol, <<, >>, 1)
+F4 (int, char, rol, <<, >>, 1)
+F4 (long, char, rol, <<, >>, 1)
+F4 (int, short, rol, <<, >>, 1)
+F4 (long, short, rol, <<, >>, 1)