When using the unpredicated immediate forms of MUL, LSL, LSR and ASR,
the rtl patterns would still have the predicate operand we created for
the other forms. This patch splits the patterns after reload in order
to get rid of the predicate, like we already do for WHILE.
Tested on aarch64-linux-gnu and applied.
Richard
2018-12-07 Richard Sandiford <richard.sandif...@arm.com>
gcc/
* config/aarch64/aarch64-sve.md (*mul<mode>3, *v<optab><mode>3):
Split the patterns after reload if we don't need the predicate
operand.
(*post_ra_mul<mode>3, *post_ra_v<optab><mode>3): New patterns.
gcc/testsuite/
* gcc.target/aarch64/sve/pred_elim_2.c: New test.
Index: gcc/config/aarch64/aarch64-sve.md
===================================================================
--- gcc/config/aarch64/aarch64-sve.md 2018-12-07 15:01:42.602176516 +0000
+++ gcc/config/aarch64/aarch64-sve.md 2018-12-07 15:02:00.230028176 +0000
@@ -936,7 +936,7 @@ (define_expand "mul<mode>3"
;; predicate for the first alternative, but using Upa or X isn't likely
;; to gain much and would make the instruction seem less uniform to the
;; register allocator.
-(define_insn "*mul<mode>3"
+(define_insn_and_split "*mul<mode>3"
[(set (match_operand:SVE_I 0 "register_operand" "=w, w, ?&w")
(unspec:SVE_I
[(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
@@ -946,12 +946,30 @@ (define_insn "*mul<mode>3"
UNSPEC_MERGE_PTRUE))]
"TARGET_SVE"
"@
- mul\t%0.<Vetype>, %0.<Vetype>, #%3
+ #
mul\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
movprfx\t%0, %2\;mul\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
+ ; Split the unpredicated form after reload, so that we don't have
+ ; the unnecessary PTRUE.
+ "&& reload_completed
+ && !register_operand (operands[3], <MODE>mode)"
+ [(set (match_dup 0) (mult:SVE_I (match_dup 2) (match_dup 3)))]
+ ""
[(set_attr "movprfx" "*,*,yes")]
)
+;; Unpredicated multiplications by a constant (post-RA only).
+;; These are generated by splitting a predicated instruction whose
+;; predicate is unused.
+(define_insn "*post_ra_mul<mode>3"
+ [(set (match_operand:SVE_I 0 "register_operand" "=w")
+ (mult:SVE_I
+ (match_operand:SVE_I 1 "register_operand" "0")
+ (match_operand:SVE_I 2 "aarch64_sve_mul_immediate")))]
+ "TARGET_SVE && reload_completed"
+ "mul\t%0.<Vetype>, %0.<Vetype>, #%2"
+)
+
(define_insn "*madd<mode>"
[(set (match_operand:SVE_I 0 "register_operand" "=w, w, ?&w")
(plus:SVE_I
@@ -1232,7 +1250,7 @@ (define_expand "v<optab><mode>3"
;; actually need the predicate for the first alternative, but using Upa
;; or X isn't likely to gain much and would make the instruction seem
;; less uniform to the register allocator.
-(define_insn "*v<optab><mode>3"
+(define_insn_and_split "*v<optab><mode>3"
[(set (match_operand:SVE_I 0 "register_operand" "=w, w, ?&w")
(unspec:SVE_I
[(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
@@ -1242,12 +1260,28 @@ (define_insn "*v<optab><mode>3"
UNSPEC_MERGE_PTRUE))]
"TARGET_SVE"
"@
- <shift>\t%0.<Vetype>, %2.<Vetype>, #%3
+ #
<shift>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
movprfx\t%0, %2\;<shift>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
+ "&& reload_completed
+ && !register_operand (operands[3], <MODE>mode)"
+ [(set (match_dup 0) (ASHIFT:SVE_I (match_dup 2) (match_dup 3)))]
+ ""
[(set_attr "movprfx" "*,*,yes")]
)
+;; Unpredicated shift operations by a constant (post-RA only).
+;; These are generated by splitting a predicated instruction whose
+;; predicate is unused.
+(define_insn "*post_ra_v<optab><mode>3"
+ [(set (match_operand:SVE_I 0 "register_operand" "=w")
+ (ASHIFT:SVE_I
+ (match_operand:SVE_I 1 "register_operand" "w")
+ (match_operand:SVE_I 2 "aarch64_simd_<lr>shift_imm")))]
+ "TARGET_SVE && reload_completed"
+ "<shift>\t%0.<Vetype>, %1.<Vetype>, #%2"
+)
+
;; LSL, LSR and ASR by a scalar, which expands into one of the vector
;; shifts above.
(define_expand "<ASHIFT:optab><mode>3"
Index: gcc/testsuite/gcc.target/aarch64/sve/pred_elim_2.c
===================================================================
--- /dev/null 2018-11-29 13:15:04.463550658 +0000
+++ gcc/testsuite/gcc.target/aarch64/sve/pred_elim_2.c 2018-12-07
15:02:00.230028176 +0000
@@ -0,0 +1,31 @@
+/* { dg-options "-O2 -ftree-vectorize" } */
+
+#include <stdint.h>
+
+#define TEST_OP(NAME, TYPE, OP) \
+ void \
+ NAME##_##TYPE (TYPE *restrict a, TYPE *restrict b, int n) \
+ { \
+ for (int i = 0; i < n; ++i) \
+ a[i] = b[i] OP; \
+ }
+
+#define TEST_TYPE(TYPE) \
+ TEST_OP (shl, TYPE, << 6) \
+ TEST_OP (shr, TYPE, >> 6) \
+ TEST_OP (mult, TYPE, * 0x2b)
+
+TEST_TYPE (int8_t)
+TEST_TYPE (int16_t)
+TEST_TYPE (int32_t)
+TEST_TYPE (int64_t)
+TEST_TYPE (uint8_t)
+TEST_TYPE (uint16_t)
+TEST_TYPE (uint32_t)
+TEST_TYPE (uint64_t)
+
+/* { dg-final { scan-assembler-times {\tlsl\t} 8 } } */
+/* { dg-final { scan-assembler-times {\tlsr\t} 4 } } */
+/* { dg-final { scan-assembler-times {\tasr\t} 4 } } */
+/* { dg-final { scan-assembler-times {\tmul\t} 8 } } */
+/* { dg-final { scan-assembler-not {\tptrue\t} } } */
