RISC-V: Optimize slli(.uw)? + addw + zext.w into sh[123]add + zext.w
Checks
Commit Message
gcc/ChangeLog:
* config/riscv/bitmanip.md: Handle corner-cases for combine
when chaining slli(.uw)? + addw
gcc/testsuite/ChangeLog:
* gcc.target/riscv/zba-shNadd-04.c: New test.
---
gcc/config/riscv/bitmanip.md | 49 +++++++++++++++++++
gcc/config/riscv/riscv-protos.h | 1 +
gcc/config/riscv/riscv.cc | 7 +++
.../gcc.target/riscv/zba-shNadd-04.c | 23 +++++++++
4 files changed, 80 insertions(+)
create mode 100644 gcc/testsuite/gcc.target/riscv/zba-shNadd-04.c
Comments
On 11/8/22 12:57, Philipp Tomsich wrote:
> gcc/ChangeLog:
>
> * config/riscv/bitmanip.md: Handle corner-cases for combine
> when chaining slli(.uw)? + addw
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/riscv/zba-shNadd-04.c: New test.
OK.
Something to consider. We're gaining a lot of
(subreg:SI (reg:DI) 0) kinds of operands.
Would it make sense to make an operand predicate that accepted
(reg:SI) or (subreg:SI (reg:DI) 0)?
It will reduce my compaints about subregs :-) But the real reason I'm
suggesting we consider adding such a predicate is, AFIACT, it it gives
combine a chance to eliminate the subreg. I haven't actually tested
this, but it seems like it might be worth a quick experiment independent
of these patches (and probably targeted towards gcc-14 rather than gcc-13).
jeff
Applied to master. Thanks.
--Philipp.
On Fri, 18 Nov 2022 at 20:52, Jeff Law <jeffreyalaw@gmail.com> wrote:
>
> On 11/8/22 12:57, Philipp Tomsich wrote:
> > gcc/ChangeLog:
> >
> > * config/riscv/bitmanip.md: Handle corner-cases for combine
> > when chaining slli(.uw)? + addw
> >
> > gcc/testsuite/ChangeLog:
> >
> > * gcc.target/riscv/zba-shNadd-04.c: New test.
>
> OK.
>
> Something to consider. We're gaining a lot of
>
> (subreg:SI (reg:DI) 0) kinds of operands.
>
>
> Would it make sense to make an operand predicate that accepted
>
> (reg:SI) or (subreg:SI (reg:DI) 0)?
>
>
> It will reduce my compaints about subregs :-) But the real reason I'm
> suggesting we consider adding such a predicate is, AFIACT, it it gives
> combine a chance to eliminate the subreg. I haven't actually tested
> this, but it seems like it might be worth a quick experiment independent
> of these patches (and probably targeted towards gcc-14 rather than gcc-13).
>
>
>
> jeff
>
>
On Fri, 18 Nov 2022 at 20:52, Jeff Law <jeffreyalaw@gmail.com> wrote:
> Something to consider. We're gaining a lot of
>
> (subreg:SI (reg:DI) 0) kinds of operands.
>
>
> Would it make sense to make an operand predicate that accepted
>
> (reg:SI) or (subreg:SI (reg:DI) 0)?
>
>
> It will reduce my compaints about subregs :-) But the real reason I'm
> suggesting we consider adding such a predicate is, AFIACT, it it gives
> combine a chance to eliminate the subreg. I haven't actually tested
> this, but it seems like it might be worth a quick experiment independent
> of these patches (and probably targeted towards gcc-14 rather than gcc-13).
>
I like the idea. Definitively something to consider. We'll give this a try.
--Philipp.
@@ -56,6 +56,55 @@
[(set (match_dup 5) (plus:DI (ashift:DI (match_dup 1) (match_dup 2)) (match_dup 3)))
(set (match_dup 0) (sign_extend:DI (div:SI (subreg:SI (match_dup 5) 0) (subreg:SI (match_dup 4) 0))))])
+; Zba does not provide W-forms of sh[123]add(.uw)?, which leads to an
+; interesting irregularity: we can generate a signed 32-bit result
+; using slli(.uw)?+ addw, but a unsigned 32-bit result can be more
+; efficiently be generated as sh[123]add+zext.w (the .uw can be
+; dropped, if we zero-extend the output anyway).
+;
+; To enable this optimization, we split [ slli(.uw)?, addw, zext.w ]
+; into [ sh[123]add, zext.w ] for use during combine.
+(define_split
+ [(set (match_operand:DI 0 "register_operand")
+ (zero_extend:DI (plus:SI (ashift:SI (subreg:SI (match_operand:DI 1 "register_operand") 0)
+ (match_operand:QI 2 "imm123_operand"))
+ (subreg:SI (match_operand:DI 3 "register_operand") 0))))]
+ "TARGET_64BIT && TARGET_ZBA"
+ [(set (match_dup 0) (plus:DI (ashift:DI (match_dup 1) (match_dup 2)) (match_dup 3)))
+ (set (match_dup 0) (zero_extend:DI (subreg:SI (match_dup 0) 0)))])
+
+(define_split
+ [(set (match_operand:DI 0 "register_operand")
+ (zero_extend:DI (plus:SI (subreg:SI (and:DI (ashift:DI (match_operand:DI 1 "register_operand")
+ (match_operand:QI 2 "imm123_operand"))
+ (match_operand:DI 3 "consecutive_bits_operand")) 0)
+ (subreg:SI (match_operand:DI 4 "register_operand") 0))))]
+ "TARGET_64BIT && TARGET_ZBA
+ && riscv_shamt_matches_mask_p (INTVAL (operands[2]), INTVAL (operands[3]))"
+ [(set (match_dup 0) (plus:DI (ashift:DI (match_dup 1) (match_dup 2)) (match_dup 4)))
+ (set (match_dup 0) (zero_extend:DI (subreg:SI (match_dup 0) 0)))])
+
+; Make sure that an andi followed by a sh[123]add remains a two instruction
+; sequence--and is not torn apart into slli, slri, add.
+(define_insn_and_split "*andi_add.uw"
+ [(set (match_operand:DI 0 "register_operand" "=r")
+ (plus:DI (and:DI (ashift:DI (match_operand:DI 1 "register_operand" "r")
+ (match_operand:QI 2 "imm123_operand" "Ds3"))
+ (match_operand:DI 3 "consecutive_bits_operand" ""))
+ (match_operand:DI 4 "register_operand" "r")))
+ (clobber (match_scratch:DI 5 "=&r"))]
+ "TARGET_64BIT && TARGET_ZBA
+ && riscv_shamt_matches_mask_p (INTVAL (operands[2]), INTVAL (operands[3]))
+ && SMALL_OPERAND (INTVAL (operands[3]) >> INTVAL (operands[2]))"
+ "#"
+ "&& reload_completed"
+ [(set (match_dup 5) (and:DI (match_dup 1) (match_dup 3)))
+ (set (match_dup 0) (plus:DI (ashift:DI (match_dup 5) (match_dup 2))
+ (match_dup 4)))]
+{
+ operands[3] = GEN_INT (INTVAL (operands[3]) >> INTVAL (operands[2]));
+})
+
(define_insn "*shNadduw"
[(set (match_operand:DI 0 "register_operand" "=r")
(plus:DI
@@ -77,6 +77,7 @@ extern bool riscv_gpr_save_operation_p (rtx);
extern void riscv_reinit (void);
extern poly_uint64 riscv_regmode_natural_size (machine_mode);
extern bool riscv_v_ext_vector_mode_p (machine_mode);
+extern bool riscv_shamt_matches_mask_p (int, HOST_WIDE_INT);
/* Routines implemented in riscv-c.cc. */
void riscv_cpu_cpp_builtins (cpp_reader *);
@@ -6497,6 +6497,13 @@ riscv_regmode_natural_size (machine_mode mode)
return UNITS_PER_WORD;
}
+/* Return true if a shift-amount matches the trailing cleared bits on a bitmask */
+bool
+riscv_shamt_matches_mask_p (int shamt, HOST_WIDE_INT mask)
+{
+ return shamt == ctz_hwi (mask);
+}
+
/* Initialize the GCC target structure. */
#undef TARGET_ASM_ALIGNED_HI_OP
#define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
new file mode 100644
@@ -0,0 +1,23 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gc_zba -mabi=lp64" } */
+/* { dg-skip-if "" { *-*-* } { "-O0" "-Og" } } */
+
+long long sub1(unsigned long long a, unsigned long long b)
+{
+ b = (b << 32) >> 31;
+ unsigned int x = a + b;
+ return x;
+}
+
+long long sub2(unsigned long long a, unsigned long long b)
+{
+ return (unsigned int)(a + (b << 1));
+}
+
+long long sub3(unsigned long long a, unsigned long long b)
+{
+ return (a + (b << 1)) & ~0u;
+}
+
+/* { dg-final { scan-assembler-times "sh1add" 3 } } */
+/* { dg-final { scan-assembler-times "zext.w\t" 3 } } */