This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [mips patch RFA] add -mfix-sb1 to work around SB-1 CPUerrata.
- From: cgd at broadcom dot com
- To: gcc-patches at gcc dot gnu dot org
- Cc: echristo at redhat dot com
- Date: 22 Aug 2003 15:22:10 -0700
- Subject: Re: [mips patch RFA] add -mfix-sb1 to work around SB-1 CPUerrata.
- References: <yov5vfsqhhru.fsf@ldt-sj3-010.sj.broadcom.com><mailpost.1061498100.19652@news-sj1-1>
OK, after some discussion with echristo & rsandifo off-line, I've
reformulated this as the following.
Less duplication of instruction patterns seems like a good thing, to
me.
> I've checked this on mipsisa64-elf against the simulator, C only (w/
> sources as of a day or two ago). No regressions.
Ditto.
chris
--
2003-08-22 Chris Demetriou <cgd@broadcom.com>
* config/mips/mips.h (MASK_FIX_SB1, TARGET_FIX_SB1): New defines.
(TARGET_SWITCHES): Add -mfix-sb1 and -mno-fix-sb1.
* config/mips/mips.md (divdf3, divsf3, sqrtdf2, sqrtsf2): Work
around SB-1 errata if TARGET_FIX_SB1 is set.
(recip.d insn, recip.s insn, rsqrt.d insn, rsqrt.s insn): Likewise.
* doc/invoke.texi: Document MIPS -mfix-sb1 and -mno-fix-sb1.
Index: config/mips/mips.h
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/mips/mips.h,v
retrieving revision 1.284
diff -u -p -r1.284 mips.h
--- config/mips/mips.h 12 Aug 2003 06:44:54 -0000 1.284
+++ config/mips/mips.h 22 Aug 2003 22:15:49 -0000
@@ -170,6 +170,7 @@ extern const struct mips_cpu_info *mips_
#define MASK_UNINIT_CONST_IN_RODATA \
0x00800000 /* Store uninitialized
consts in rodata */
+#define MASK_FIX_SB1 0x01000000 /* Work around SB-1 errata. */
/* Debug switches, not documented */
#define MASK_DEBUG 0 /* unused */
@@ -255,6 +256,7 @@ extern const struct mips_cpu_info *mips_
#define TARGET_BRANCHLIKELY (target_flags & MASK_BRANCHLIKELY)
+#define TARGET_FIX_SB1 (target_flags & MASK_FIX_SB1)
/* True if we should use NewABI-style relocation operators for
symbolic addresses. This is never true for mips16 code,
@@ -580,6 +582,10 @@ extern const struct mips_cpu_info *mips_
N_("Work around early 4300 hardware bug")}, \
{"no-fix4300", -MASK_4300_MUL_FIX, \
N_("Don't work around early 4300 hardware bug")}, \
+ {"fix-sb1", MASK_FIX_SB1, \
+ N_("Work around errata for early SB-1 revision 2 cores")}, \
+ {"no-fix-sb1", -MASK_FIX_SB1, \
+ N_("Don't work around errata for early SB-1 revision 2 cores")}, \
{"check-zero-division",-MASK_NO_CHECK_ZERO_DIV, \
N_("Trap on integer divide by zero")}, \
{"no-check-zero-division", MASK_NO_CHECK_ZERO_DIV, \
Index: config/mips/mips.md
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/mips/mips.md,v
retrieving revision 1.191
diff -u -p -r1.191 mips.md
--- config/mips/mips.md 19 Aug 2003 18:09:06 -0000 1.191
+++ config/mips/mips.md 22 Aug 2003 22:15:49 -0000
@@ -2365,41 +2365,104 @@
;; ....................
;;
+;; This pattern works around the early SB-1 rev2 core "F1" erratum:
+;;
+;; If an mfc1 or dmfc1 happens to access the floating point register
+;; file at the same time a long latency operation (div, sqrt, recip,
+;; sqrt) iterates an intermediate result back through the floating
+;; point register file bypass, then instead returning the correct
+;; register value the mfc1 or dmfc1 operation returns the intermediate
+;; result of the long latency operation.
+;;
+;; The workaround is to insert an unconditional 'mov' from/to the
+;; long latency op destination register.
(define_insn "divdf3"
[(set (match_operand:DF 0 "register_operand" "=f")
(div:DF (match_operand:DF 1 "register_operand" "f")
(match_operand:DF 2 "register_operand" "f")))]
"TARGET_HARD_FLOAT && TARGET_DOUBLE_FLOAT"
- "div.d\t%0,%1,%2"
+{
+ if (TARGET_FIX_SB1)
+ return "div.d\t%0,%1,%2\;mov.d\t%0,%0";
+ else
+ return "div.d\t%0,%1,%2";
+}
[(set_attr "type" "fdiv")
- (set_attr "mode" "DF")])
+ (set_attr "mode" "DF")
+ (set (attr "length")
+ (if_then_else (ne (symbol_ref "TARGET_FIX_SB1") (const_int 0))
+ (const_int 8)
+ (const_int 4)))])
+
+;; This pattern works around the early SB-1 rev2 core "F1" erratum (see
+;; "divdf3" comment for details).
+;;
+;; This pattern works around the early SB-1 rev2 core "F2" erratum:
+;;
+;; In certain cases, div.s and div.ps may have a rounding error
+;; and/or wrong inexact flag.
+;;
+;; Therefore, we only allow div.s if not working around SB-1 rev2
+;; errata, or if working around those errata and a slight loss of
+;; precision is OK (i.e., flag_unsafe_math_optimizations is set).
(define_insn "divsf3"
[(set (match_operand:SF 0 "register_operand" "=f")
(div:SF (match_operand:SF 1 "register_operand" "f")
(match_operand:SF 2 "register_operand" "f")))]
- "TARGET_HARD_FLOAT"
- "div.s\t%0,%1,%2"
+ "TARGET_HARD_FLOAT && (!TARGET_FIX_SB1 || flag_unsafe_math_optimizations)"
+{
+ if (TARGET_FIX_SB1)
+ return "div.s\t%0,%1,%2\;mov.s\t%0,%0";
+ else
+ return "div.s\t%0,%1,%2";
+}
[(set_attr "type" "fdiv")
- (set_attr "mode" "SF")])
+ (set_attr "mode" "SF")
+ (set (attr "length")
+ (if_then_else (ne (symbol_ref "TARGET_FIX_SB1") (const_int 0))
+ (const_int 8)
+ (const_int 4)))])
+;; This pattern works around the early SB-1 rev2 core "F1" erratum (see
+;; "divdf3" comment for details).
(define_insn ""
[(set (match_operand:DF 0 "register_operand" "=f")
(div:DF (match_operand:DF 1 "const_float_1_operand" "")
(match_operand:DF 2 "register_operand" "f")))]
"ISA_HAS_FP4 && TARGET_HARD_FLOAT && TARGET_DOUBLE_FLOAT && flag_unsafe_math_optimizations"
- "recip.d\t%0,%2"
+{
+ if (TARGET_FIX_SB1)
+ return "recip.d\t%0,%2\;mov.d\t%0,%0";
+ else
+ return "recip.d\t%0,%2";
+}
[(set_attr "type" "fdiv")
- (set_attr "mode" "DF")])
+ (set_attr "mode" "DF")
+ (set (attr "length")
+ (if_then_else (ne (symbol_ref "TARGET_FIX_SB1") (const_int 0))
+ (const_int 8)
+ (const_int 4)))])
+;; This pattern works around the early SB-1 rev2 core "F1" erratum (see
+;; "divdf3" comment for details).
(define_insn ""
[(set (match_operand:SF 0 "register_operand" "=f")
(div:SF (match_operand:SF 1 "const_float_1_operand" "")
(match_operand:SF 2 "register_operand" "f")))]
"ISA_HAS_FP4 && TARGET_HARD_FLOAT && flag_unsafe_math_optimizations"
- "recip.s\t%0,%2"
+{
+ if (TARGET_FIX_SB1)
+ return "recip.s\t%0,%2\;mov.s\t%0,%0";
+ else
+ return "recip.s\t%0,%2";
+}
[(set_attr "type" "fdiv")
- (set_attr "mode" "SF")])
+ (set_attr "mode" "SF")
+ (set (attr "length")
+ (if_then_else (ne (symbol_ref "TARGET_FIX_SB1") (const_int 0))
+ (const_int 8)
+ (const_int 4)))])
(define_insn "divmodsi4"
[(set (match_operand:SI 0 "register_operand" "=l")
@@ -2456,39 +2519,83 @@
;;
;; ...................
+;; This pattern works around the early SB-1 rev2 core "F1" erratum (see
+;; "divdf3" comment for details).
(define_insn "sqrtdf2"
[(set (match_operand:DF 0 "register_operand" "=f")
(sqrt:DF (match_operand:DF 1 "register_operand" "f")))]
"TARGET_HARD_FLOAT && HAVE_SQRT_P() && TARGET_DOUBLE_FLOAT"
- "sqrt.d\t%0,%1"
+{
+ if (TARGET_FIX_SB1)
+ return "sqrt.d\t%0,%1\;mov.d\t%0,%0";
+ else
+ return "sqrt.d\t%0,%1";
+}
[(set_attr "type" "fsqrt")
- (set_attr "mode" "DF")])
+ (set_attr "mode" "DF")
+ (set (attr "length")
+ (if_then_else (ne (symbol_ref "TARGET_FIX_SB1") (const_int 0))
+ (const_int 8)
+ (const_int 4)))])
+;; This pattern works around the early SB-1 rev2 core "F1" erratum (see
+;; "divdf3" comment for details).
(define_insn "sqrtsf2"
[(set (match_operand:SF 0 "register_operand" "=f")
(sqrt:SF (match_operand:SF 1 "register_operand" "f")))]
"TARGET_HARD_FLOAT && HAVE_SQRT_P()"
- "sqrt.s\t%0,%1"
+{
+ if (TARGET_FIX_SB1)
+ return "sqrt.s\t%0,%1\;mov.s\t%0,%0";
+ else
+ return "sqrt.s\t%0,%1";
+}
[(set_attr "type" "fsqrt")
- (set_attr "mode" "SF")])
+ (set_attr "mode" "SF")
+ (set (attr "length")
+ (if_then_else (ne (symbol_ref "TARGET_FIX_SB1") (const_int 0))
+ (const_int 8)
+ (const_int 4)))])
+;; This pattern works around the early SB-1 rev2 core "F1" erratum (see
+;; "divdf3" comment for details).
(define_insn ""
[(set (match_operand:DF 0 "register_operand" "=f")
(div:DF (match_operand:DF 1 "const_float_1_operand" "")
(sqrt:DF (match_operand:DF 2 "register_operand" "f"))))]
"ISA_HAS_FP4 && TARGET_HARD_FLOAT && TARGET_DOUBLE_FLOAT && flag_unsafe_math_optimizations"
- "rsqrt.d\t%0,%2"
+{
+ if (TARGET_FIX_SB1)
+ return "rsqrt.d\t%0,%2\;mov.d\t%0,%0";
+ else
+ return "rsqrt.d\t%0,%2";
+}
[(set_attr "type" "frsqrt")
- (set_attr "mode" "DF")])
+ (set_attr "mode" "DF")
+ (set (attr "length")
+ (if_then_else (ne (symbol_ref "TARGET_FIX_SB1") (const_int 0))
+ (const_int 8)
+ (const_int 4)))])
+;; This pattern works around the early SB-1 rev2 core "F1" erratum (see
+;; "divdf3" comment for details).
(define_insn ""
[(set (match_operand:SF 0 "register_operand" "=f")
(div:SF (match_operand:SF 1 "const_float_1_operand" "")
(sqrt:SF (match_operand:SF 2 "register_operand" "f"))))]
"ISA_HAS_FP4 && TARGET_HARD_FLOAT && flag_unsafe_math_optimizations"
- "rsqrt.s\t%0,%2"
+{
+ if (TARGET_FIX_SB1)
+ return "rsqrt.s\t%0,%2\;mov.s\t%0,%0";
+ else
+ return "rsqrt.s\t%0,%2";
+}
[(set_attr "type" "frsqrt")
- (set_attr "mode" "SF")])
+ (set_attr "mode" "SF")
+ (set (attr "length")
+ (if_then_else (ne (symbol_ref "TARGET_FIX_SB1") (const_int 0))
+ (const_int 8)
+ (const_int 4)))])
;;
;; ....................
Index: doc/invoke.texi
===================================================================
RCS file: /cvs/gcc/gcc/gcc/doc/invoke.texi,v
retrieving revision 1.329
diff -u -p -r1.329 invoke.texi
--- doc/invoke.texi 19 Aug 2003 20:29:00 -0000 1.329
+++ doc/invoke.texi 22 Aug 2003 22:15:50 -0000
@@ -479,7 +479,8 @@ in the following sections.
-m4650 -msingle-float -mmad @gol
-EL -EB -G @var{num} -nocpp @gol
-mabi=32 -mabi=n32 -mabi=64 -mabi=eabi -mabi-fake-default @gol
--mfix7000 -mno-crt0 -mflush-func=@var{func} -mno-flush-func @gol
+-mfix7000 -mfix-sb1 -mno-fix-sb1 @gol
+-mno-crt0 -mflush-func=@var{func} -mno-flush-func @gol
-mbranch-likely -mno-branch-likely}
@emph{i386 and x86-64 Options}
@@ -8167,6 +8168,13 @@ assembler files (with a @samp{.s} suffix
Pass an option to gas which will cause nops to be inserted if
the read of the destination register of an mfhi or mflo instruction
occurs in the following two instructions.
+
+@item -mfix-sb1
+@itemx -mno-fix-sb1
+@opindex mfix-sb1
+Work around certain SB-1 CPU core errata.
+(This flag currently works around the SB-1 revision 2
+``F1'' and ``F2'' floating point errata.)
@item -no-crt0
@opindex no-crt0