diff mbox series

RISC-V: Add the missed half floating-point mode patterns of local_pic_load/store when only use zfhmin

Message ID 20230812143643.1511082-1-lehua.ding@rivai.ai
State New
Headers show
Series RISC-V: Add the missed half floating-point mode patterns of local_pic_load/store when only use zfhmin | expand

Commit Message

Lehua Ding Aug. 12, 2023, 2:36 p.m. UTC
Hi,

There is a new failed RISC-V testcase(testsuite/gcc.target/riscv/rvv/autovec/vls/const-4.c)
on the current trunk branch when use medany as default cmodel.
The reason is the load of half floating-point imm is convert from RTL 1 to RTL
2 as the cmodel be changed from medlow to medany. This change let insn 7 be
combineed with @pred_broadcast patterns (insn 8) at combine pass. However,
insn 6 and insn 7 are combined for SF and DF mode, but not for HF mode, and
the fail combined leads to insn 7 and insn 8 be combined. The reason of the
fail combined is the local_pic_loadhf pattern doesn't exist when only enable
zfhmin(implied by zvfh).

Therefore, when only zfhmin but not zfh is enabled, the define_insn of
*local_pic_load<ANYF:mode> must also be able to produce the pattern for
*load_pic_loadhf pattern, since the zfhmin extension also includes a
half floating-point load/store instructions. So, I added an ANFLSF Iterator
and applied it to local_pic_load/store define_insns. I have checked other ANYF
usage scenarios and feel that this is the only place that needs to be corrected.
I may have missed something, please correct. Thanks.

RTL 1:

(insn 6 3 7 2 (set (reg:DI 137)
        (high:DI (symbol_ref/u:DI ("*.LC0") [flags 0x82]))) "/work/home/lding/open-source/riscv-gnu-toolchain-push/gcc/gcc/testsuite/gcc.target/riscv/rvv/autovec/vls/const-4.c":7:1 discrim 3 179 {*movdi_64bit}
     (nil))
(insn 7 6 8 2 (set (reg:HF 136)
        (mem/u/c:HF (lo_sum:DI (reg:DI 137)
                (symbol_ref/u:DI ("*.LC0") [flags 0x82])) [0  S2 A16])) "/work/home/lding/open-source/riscv-gnu-toolchain-push/gcc/gcc/testsuite/gcc.target/riscv/rvv/autovec/vls/const-4.c":7:1 discrim 3 126 {*movhf_hardfloat}
     (expr_list:REG_EQUAL (const_double:HF 8.8828125e+0 [0x0.8e2p+4])
        (nil)))

RTL 2:

(insn 6 3 7 2 (set (reg/f:DI 137)
        (symbol_ref/u:DI ("*.LC0") [flags 0x82])) "/work/home/lding/open-source/riscv-gnu-toolchain-push/gcc/gcc/testsuite/gcc.target/riscv/rvv/autovec/vls/const-4.c":7:1 discrim 3 179 {*movdi_64bit}
     (nil))
(insn 7 6 8 2 (set (reg:HF 136)
        (mem/u/c:HF (reg/f:DI 137) [0  S2 A16])) "/work/home/lding/open-source/riscv-gnu-toolchain-push/gcc/gcc/testsuite/gcc.target/riscv/rvv/autovec/vls/const-4.c":7:1 discrim 3 126 {*movhf_hardfloat}
     (expr_list:REG_EQUAL (const_double:HF 8.8828125e+0 [0x0.8e2p+4])
        (nil)))
(insn 8 7 9 2 (set (reg:V2HF 135)
        (if_then_else:V2HF (unspec:V2BI [
                    (const_vector:V2BI [
                            (const_int 1 [0x1]) repeated x2
                        ])
                    (const_int 2 [0x2]) repeated x3
                    (const_int 0 [0])
                    (reg:SI 66 vl)
                    (reg:SI 67 vtype)
                ] UNSPEC_VPREDICATE)
            (vec_duplicate:V2HF (reg:HF 136))
            (unspec:V2HF [
                    (reg:SI 0 zero)
                ] UNSPEC_VUNDEF))) "/work/home/lding/open-source/riscv-gnu-toolchain-push/gcc/gcc/testsuite/gcc.target/riscv/rvv/autovec/vls/const-4.c":6:1 discrim 3 1389 {*pred_broadcastv2hf}
     (nil))

Best,
Lehua

gcc/ChangeLog:

	* config/riscv/iterators.md (TARGET_HARD_FLOAT || TARGET_ZFINX): New.
	* config/riscv/pic.md (*local_pic_load<ANYF:mode>): Change ANYF.
	(*local_pic_load<ANYLSF:mode>): To ANYLSF.
	(*local_pic_load_32d<ANYF:mode>): Ditto.
	(*local_pic_load_32d<ANYLSF:mode>): Ditto.
	(*local_pic_store<ANYF:mode>): Ditto.
	(*local_pic_store<ANYLSF:mode>): Ditto.
	(*local_pic_store_32d<ANYF:mode>): Ditto.
	(*local_pic_store_32d<ANYLSF:mode>): Ditto.

gcc/testsuite/ChangeLog:

	* gcc.target/riscv/_Float16-zfhmin-4.c: New test.

---
 gcc/config/riscv/iterators.md                 |  5 +++
 gcc/config/riscv/pic.md                       | 34 +++++++++----------
 .../gcc.target/riscv/_Float16-zfhmin-4.c      | 11 ++++++
 3 files changed, 33 insertions(+), 17 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/riscv/_Float16-zfhmin-4.c

Comments

Robin Dapp Aug. 17, 2023, 11:48 a.m. UTC | #1
Hi Lehua,

thanks for fixing this.  Looks like the same reason we have the
separation of zvfh and zvfhmin for vector loads/stores.

> +;; Iterator for hardware-supported load/store floating-point modes.
> +(define_mode_iterator ANYLSF [(SF "TARGET_HARD_FLOAT || TARGET_ZFINX")
> +			      (DF "TARGET_DOUBLE_FLOAT || TARGET_ZDINX")
> +			      (HF "TARGET_ZFHMIN || TARGET_ZHINX")])
> +

I first thought we needed TARGET_ZFH here as well but it appears that
TARGET_ZFH implies TARGET_ZFHMIN via riscv_implied_info.  We're lacking
that on the vector side and this should be addressed separately.

You likely want TARGET_ZHINXMIN instead of ZHINX though?  I mean the
hardware support is obviously always there but the patterns should
be available for the min extension already.  Please double check as
I haven't worked with that extension before.
Our test coverage for the *inx extensions is honestly a bit sparse,
maybe you would also want to add a testcase for a similar scenario?

> -;; We can support ANYF loads into X register if there is no double support
> +;; We can support ANYLSF loads into X register if there is no double support
>  ;; or if the target is 64-bit> -(define_insn "*local_pic_load<ANYF:mode>"
> -  [(set (match_operand:ANYF 0 "register_operand" "=f,*r")
> -	(mem:ANYF (match_operand 1 "absolute_symbolic_operand" "")))
> +(define_insn "*local_pic_load<ANYLSF:mode>"
> +  [(set (match_operand:ANYLSF 0 "register_operand" "=f,*r")
> +	(mem:ANYLSF (match_operand 1 "absolute_symbolic_operand" "")))
>     (clobber (match_scratch:P 2 "=r,X"))]
>    "TARGET_HARD_FLOAT && USE_LOAD_ADDRESS_MACRO (operands[1])
>     && (!TARGET_DOUBLE_FLOAT || TARGET_64BIT)"
>    "@
> -   <ANYF:load>\t%0,%1,%2
> +   <ANYLSF:load>\t%0,%1,%2
>     <softload>\t%0,%1"
>    [(set (attr "length") (const_int 8))])

Unrelated to your patch - but from a quick glimpse here I didn't see
why we require TARGET_HARD_FLOAT for the softload alternatives.  Aren't
zdinx, zfinx, zhinx a bit of a SOFT_FLOAT thing?  Well probably just
semantics... 

Apart from that LGTM.

Regards
 Robin
Lehua Ding Aug. 17, 2023, 12:17 p.m. UTC | #2
Hi Robin,


&gt; You likely want TARGET_ZHINXMIN instead of ZHINX though?&nbsp; I mean the
&gt; hardware support is obviously always there but the patterns should
&gt; be available for the min extension already.&nbsp; Please double check as
&gt; I haven't worked with that extension before.
&gt; Our test coverage for the *inx extensions is honestly a bit sparse,
&gt; maybe you would also want to add a testcase for a similar scenario?


Indeed, thanks for the reminder. I'll add the missing ones and add V2 patch.


&gt; Unrelated to your patch - but from a quick glimpse here I didn't see
&gt; why we require TARGET_HARD_FLOAT for the softload alternatives.&nbsp; Aren't
&gt; zdinx, zfinx, zhinx a bit of a SOFT_FLOAT thing?&nbsp; Well probably just
&gt; semantics...


Looking closely at this condition is a bit odd for me too.




Best,
Lehua


------------------&nbsp;Original&nbsp;------------------
From:                                                                                                                        "Robin Dapp"                                                                                    <rdapp.gcc@gmail.com&gt;;
Date:&nbsp;Thu, Aug 17, 2023 07:48 PM
To:&nbsp;"Lehua Ding"<lehua.ding@rivai.ai&gt;;"gcc-patches"<gcc-patches@gcc.gnu.org&gt;;"kito.cheng"<kito.cheng@gmail.com&gt;;
Cc:&nbsp;"rdapp.gcc"<rdapp.gcc@gmail.com&gt;;"juzhe.zhong"<juzhe.zhong@rivai.ai&gt;;"palmer"<palmer@rivosinc.com&gt;;"jeffreyalaw"<jeffreyalaw@gmail.com&gt;;
Subject:&nbsp;Re: [PATCH] RISC-V: Add the missed half floating-point mode patterns of local_pic_load/store when only use zfhmin



Hi Lehua,

thanks for fixing this.&nbsp; Looks like the same reason we have the
separation of zvfh and zvfhmin for vector loads/stores.

&gt; +;; Iterator for hardware-supported load/store floating-point modes.
&gt; +(define_mode_iterator ANYLSF [(SF "TARGET_HARD_FLOAT || TARGET_ZFINX")
&gt; + &nbsp;&nbsp;&nbsp;&nbsp;&nbsp; (DF "TARGET_DOUBLE_FLOAT || TARGET_ZDINX")
&gt; + &nbsp;&nbsp;&nbsp;&nbsp;&nbsp; (HF "TARGET_ZFHMIN || TARGET_ZHINX")])
&gt; +

I first thought we needed TARGET_ZFH here as well but it appears that
TARGET_ZFH implies TARGET_ZFHMIN via riscv_implied_info.&nbsp; We're lacking
that on the vector side and this should be addressed separately.

You likely want TARGET_ZHINXMIN instead of ZHINX though?&nbsp; I mean the
hardware support is obviously always there but the patterns should
be available for the min extension already.&nbsp; Please double check as
I haven't worked with that extension before.
Our test coverage for the *inx extensions is honestly a bit sparse,
maybe you would also want to add a testcase for a similar scenario?

&gt; -;; We can support ANYF loads into X register if there is no double support
&gt; +;; We can support ANYLSF loads into X register if there is no double support
&gt;&nbsp; ;; or if the target is 64-bit&gt; -(define_insn "*local_pic_load<ANYF:mode&gt;"
&gt; -&nbsp; [(set (match_operand:ANYF 0 "register_operand" "=f,*r")
&gt; - (mem:ANYF (match_operand 1 "absolute_symbolic_operand" "")))
&gt; +(define_insn "*local_pic_load<ANYLSF:mode&gt;"
&gt; +&nbsp; [(set (match_operand:ANYLSF 0 "register_operand" "=f,*r")
&gt; + (mem:ANYLSF (match_operand 1 "absolute_symbolic_operand" "")))
&gt;&nbsp;&nbsp;&nbsp;&nbsp; (clobber (match_scratch:P 2 "=r,X"))]
&gt;&nbsp;&nbsp;&nbsp; "TARGET_HARD_FLOAT &amp;&amp; USE_LOAD_ADDRESS_MACRO (operands[1])
&gt;&nbsp;&nbsp;&nbsp;&nbsp; &amp;&amp; (!TARGET_DOUBLE_FLOAT || TARGET_64BIT)"
&gt;&nbsp;&nbsp;&nbsp; "@
&gt; -&nbsp;&nbsp; <ANYF:load&gt;\t%0,%1,%2
&gt; +&nbsp;&nbsp; <ANYLSF:load&gt;\t%0,%1,%2
&gt;&nbsp;&nbsp;&nbsp;&nbsp; <softload&gt;\t%0,%1"
&gt;&nbsp;&nbsp;&nbsp; [(set (attr "length") (const_int 8))])

Unrelated to your patch - but from a quick glimpse here I didn't see
why we require TARGET_HARD_FLOAT for the softload alternatives.&nbsp; Aren't
zdinx, zfinx, zhinx a bit of a SOFT_FLOAT thing?&nbsp; Well probably just
semantics... 

Apart from that LGTM.

Regards
Robin
diff mbox series

Patch

diff --git a/gcc/config/riscv/iterators.md b/gcc/config/riscv/iterators.md
index d374a10810c..39c2dd629a2 100644
--- a/gcc/config/riscv/iterators.md
+++ b/gcc/config/riscv/iterators.md
@@ -67,6 +67,11 @@ 
 			    (DF "TARGET_DOUBLE_FLOAT || TARGET_ZDINX")
 			    (HF "TARGET_ZFH || TARGET_ZHINX")])
 
+;; Iterator for hardware-supported load/store floating-point modes.
+(define_mode_iterator ANYLSF [(SF "TARGET_HARD_FLOAT || TARGET_ZFINX")
+			      (DF "TARGET_DOUBLE_FLOAT || TARGET_ZDINX")
+			      (HF "TARGET_ZFHMIN || TARGET_ZHINX")])
+
 ;; Iterator for floating-point modes that can be loaded into X registers.
 (define_mode_iterator SOFTF [SF (DF "TARGET_64BIT") (HF "TARGET_ZFHMIN")])
 
diff --git a/gcc/config/riscv/pic.md b/gcc/config/riscv/pic.md
index 9507850455a..da636e31619 100644
--- a/gcc/config/riscv/pic.md
+++ b/gcc/config/riscv/pic.md
@@ -43,17 +43,17 @@ 
   "<SUBX:load>u\t%0,%1"
   [(set (attr "length") (const_int 8))])
 
-;; We can support ANYF loads into X register if there is no double support
+;; We can support ANYLSF loads into X register if there is no double support
 ;; or if the target is 64-bit.
 
-(define_insn "*local_pic_load<ANYF:mode>"
-  [(set (match_operand:ANYF 0 "register_operand" "=f,*r")
-	(mem:ANYF (match_operand 1 "absolute_symbolic_operand" "")))
+(define_insn "*local_pic_load<ANYLSF:mode>"
+  [(set (match_operand:ANYLSF 0 "register_operand" "=f,*r")
+	(mem:ANYLSF (match_operand 1 "absolute_symbolic_operand" "")))
    (clobber (match_scratch:P 2 "=r,X"))]
   "TARGET_HARD_FLOAT && USE_LOAD_ADDRESS_MACRO (operands[1])
    && (!TARGET_DOUBLE_FLOAT || TARGET_64BIT)"
   "@
-   <ANYF:load>\t%0,%1,%2
+   <ANYLSF:load>\t%0,%1,%2
    <softload>\t%0,%1"
   [(set (attr "length") (const_int 8))])
 
@@ -61,13 +61,13 @@ 
 ;; supported.  ld is not valid in that case.  Punt for now.  Maybe add a split
 ;; for this later.
 
-(define_insn "*local_pic_load_32d<ANYF:mode>"
-  [(set (match_operand:ANYF 0 "register_operand" "=f")
-	(mem:ANYF (match_operand 1 "absolute_symbolic_operand" "")))
+(define_insn "*local_pic_load_32d<ANYLSF:mode>"
+  [(set (match_operand:ANYLSF 0 "register_operand" "=f")
+	(mem:ANYLSF (match_operand 1 "absolute_symbolic_operand" "")))
    (clobber (match_scratch:P 2 "=r"))]
   "TARGET_HARD_FLOAT && USE_LOAD_ADDRESS_MACRO (operands[1])
    && (TARGET_DOUBLE_FLOAT && !TARGET_64BIT)"
-  "<ANYF:load>\t%0,%1,%2"
+  "<ANYLSF:load>\t%0,%1,%2"
   [(set (attr "length") (const_int 8))])
 
 (define_insn "*local_pic_load_sf<mode>"
@@ -88,14 +88,14 @@ 
   "<ANYI:store>\t%z1,%0,%2"
   [(set (attr "length") (const_int 8))])
 
-(define_insn "*local_pic_store<ANYF:mode>"
-  [(set (mem:ANYF (match_operand 0 "absolute_symbolic_operand" ""))
-	(match_operand:ANYF 1 "register_operand" "f,*r"))
+(define_insn "*local_pic_store<ANYLSF:mode>"
+  [(set (mem:ANYLSF (match_operand 0 "absolute_symbolic_operand" ""))
+	(match_operand:ANYLSF 1 "register_operand" "f,*r"))
    (clobber (match_scratch:P 2 "=r,&r"))]
   "TARGET_HARD_FLOAT && USE_LOAD_ADDRESS_MACRO (operands[0])
    && (!TARGET_DOUBLE_FLOAT || TARGET_64BIT)"
   "@
-   <ANYF:store>\t%1,%0,%2
+   <ANYLSF:store>\t%1,%0,%2
    <softstore>\t%1,%0,%2"
   [(set (attr "length") (const_int 8))])
 
@@ -103,13 +103,13 @@ 
 ;; supported.  sd is not valid in that case.  Punt for now.  Maybe add a split
 ;; for this later.
 
-(define_insn "*local_pic_store_32d<ANYF:mode>"
-  [(set (match_operand:ANYF 0 "register_operand" "=f")
-	(mem:ANYF (match_operand 1 "absolute_symbolic_operand" "")))
+(define_insn "*local_pic_store_32d<ANYLSF:mode>"
+  [(set (match_operand:ANYLSF 0 "register_operand" "=f")
+	(mem:ANYLSF (match_operand 1 "absolute_symbolic_operand" "")))
    (clobber (match_scratch:P 2 "=r"))]
   "TARGET_HARD_FLOAT && USE_LOAD_ADDRESS_MACRO (operands[1])
    && (TARGET_DOUBLE_FLOAT && !TARGET_64BIT)"
-  "<ANYF:store>\t%1,%0,%2"
+  "<ANYLSF:store>\t%1,%0,%2"
   [(set (attr "length") (const_int 8))])
 
 (define_insn "*local_pic_store_sf<SOFTF:mode>"
diff --git a/gcc/testsuite/gcc.target/riscv/_Float16-zfhmin-4.c b/gcc/testsuite/gcc.target/riscv/_Float16-zfhmin-4.c
new file mode 100644
index 00000000000..42a238a0380
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/_Float16-zfhmin-4.c
@@ -0,0 +1,11 @@ 
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gc_zfhmin -mabi=lp64d -O3 -mcmodel=medany" } */
+/* { dg-final { check-function-bodies "**" "" } } */
+
+/* Make sure zfhmin behaves the same way as zfh. */
+/*
+** foo: { target { no-opts "-flto" } }
+**   flh\tfa0,\.LC0,[a-z0-9]+
+**   ...
+*/
+_Float16 foo() { return 0.8974; }