[PATCH] MIPS: Add `-mfix-r5900' option for the R5900 short loop erratum

[Fredrik Noring <noring at nocrew dot org>]

The short loop bug under certain conditions causes loops to
execute only once or twice, due to a hardware bug in the R5900 chip.

`-march=r5900' already enables the R5900 short loop workaround.
However, the R5900 ISA and most other MIPS ISAs are mutually
exclusive since R5900-specific instructions are generated as well.

The `-mfix-r5900' option can be used in combination with e.g.
`-mips2' or `-mips3' to generate generic MIPS binaries that also
work with the R5900 target.  The workaround is implemented by GAS
rather than by GCC.

The following small `shortloop.c' file has been used as a test
with GCC 8.2.0:

void shortloop(void)
{
    __asm__ __volatile__ (
	"	li $3, 300\n"
	"loop:\n"
	"	addi $3, -1\n"
	"	addi $4, -1\n"
	"	bne $3, $0, loop\n"
	"	li $4, 3\n"
	::);
}

The following six combinations have been tested:

% mipsr5900el-unknown-linux-gnu-gcc -O1 -c shortloop.c
% mipsr5900el-unknown-linux-gnu-gcc -O1 -c shortloop.c -mfix-r5900
% mipsr5900el-unknown-linux-gnu-gcc -O1 -c shortloop.c -mno-fix-r5900

% mipsr4000el-unknown-linux-gnu-gcc -O1 -c shortloop.c
% mipsr4000el-unknown-linux-gnu-gcc -O1 -c shortloop.c -mfix-r5900
% mipsr4000el-unknown-linux-gnu-gcc -O1 -c shortloop.c -mno-fix-r5900

The R5900 short loop erratum is corrected in exactly three cases:

1. for the target `mipsr5900el' by default;

2. for the target `mipsr5900el' with `-mfix-r5900';

3. for any other MIPS target (e.g. `mipsr4000el') with `-mfix-r5900'.

In all other cases the correction is not made.

	* gcc/config/mips/mips.c (mips_reorg_process_insns)
	  (mips_option_override): Default to working around R5900
	  errata only if the processor was selected explicitly.
	* gcc/config/mips/mips.h: Declare `mfix-r5900' and
	  `mno-fix-r5900'.
	* gcc/config/mips/mips.opt: Define MASK_FIX_R5900.
	* gcc/doc/invoke.texi: Document the R5900, `mfix-r5900' and
	  `mno-fix-r5900'.
---
 gcc/config/mips/mips.c   | 14 ++++++++++----
 gcc/config/mips/mips.h   |  1 +
 gcc/config/mips/mips.opt |  4 ++++
 gcc/doc/invoke.texi      | 14 +++++++++++++-
 4 files changed, 28 insertions(+), 5 deletions(-)

diff --git a/gcc/config/mips/mips.c b/gcc/config/mips/mips.c
index ea2fae1d6db..5763ce21427 100644
--- a/gcc/config/mips/mips.c
+++ b/gcc/config/mips/mips.c
@@ -18881,13 +18881,13 @@ mips_reorg_process_insns (void)
   if (crtl->profile)
     cfun->machine->all_noreorder_p = false;
 
-  /* Code compiled with -mfix-vr4120, -mfix-rm7000 or -mfix-24k can't be
-     all noreorder because we rely on the assembler to work around some
-     errata.  The R5900 too has several bugs.  */
+  /* Code compiled with -mfix-vr4120, -mfix-r5900, -mfix-rm7000 or
+     -mfix-24k can't be all noreorder because we rely on the assembler
+     to work around some errata.  The R5900 target has several bugs.  */
   if (TARGET_FIX_VR4120
       || TARGET_FIX_RM7000
       || TARGET_FIX_24K
-      || TARGET_MIPS5900)
+      || TARGET_FIX_R5900)
     cfun->machine->all_noreorder_p = false;
 
   /* The same is true for -mfix-vr4130 if we might generate MFLO or
@@ -20244,6 +20244,12 @@ mips_option_override (void)
       && strcmp (mips_arch_info->name, "r4400") == 0)
     target_flags |= MASK_FIX_R4400;
 
+  /* Default to working around R5900 errata only if the processor
+     was selected explicitly.  */
+  if ((target_flags_explicit & MASK_FIX_R5900) == 0
+      && strcmp (mips_arch_info->name, "r5900") == 0)
+    target_flags |= MASK_FIX_R5900;
+
   /* Default to working around R10000 errata only if the processor
      was selected explicitly.  */
   if ((target_flags_explicit & MASK_FIX_R10000) == 0
diff --git a/gcc/config/mips/mips.h b/gcc/config/mips/mips.h
index 32a88edc910..7dd19fc6f2d 100644
--- a/gcc/config/mips/mips.h
+++ b/gcc/config/mips/mips.h
@@ -1363,6 +1363,7 @@ struct mips_cpu_info {
 %{mmsa} %{mno-msa} \
 %{msmartmips} %{mno-smartmips} \
 %{mmt} %{mno-mt} \
+%{mfix-r5900} %{mno-fix-r5900} \
 %{mfix-rm7000} %{mno-fix-rm7000} \
 %{mfix-vr4120} %{mfix-vr4130} \
 %{mfix-24k} \
diff --git a/gcc/config/mips/mips.opt b/gcc/config/mips/mips.opt
index 5a9f255fe20..427ac4913fc 100644
--- a/gcc/config/mips/mips.opt
+++ b/gcc/config/mips/mips.opt
@@ -165,6 +165,10 @@ mfix-r4400
 Target Report Mask(FIX_R4400)
 Work around certain R4400 errata.
 
+mfix-r5900
+Target Report Mask(FIX_R5900)
+Work around the R5900 short loop erratum.
+
 mfix-rm7000
 Target Report Var(TARGET_FIX_RM7000)
 Work around certain RM7000 errata.
diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
index e290128f535..c9846d96304 100644
--- a/gcc/doc/invoke.texi
+++ b/gcc/doc/invoke.texi
@@ -939,6 +939,7 @@ Objective-C and Objective-C++ Dialects}.
 -mmad  -mno-mad  -mimadd  -mno-imadd  -mfused-madd  -mno-fused-madd  -nocpp @gol
 -mfix-24k  -mno-fix-24k @gol
 -mfix-r4000  -mno-fix-r4000  -mfix-r4400  -mno-fix-r4400 @gol
+-mfix-r5900  -mno-fix-r5900 @gol
 -mfix-r10000  -mno-fix-r10000  -mfix-rm7000  -mno-fix-rm7000 @gol
 -mfix-vr4120  -mno-fix-vr4120 @gol
 -mfix-vr4130  -mno-fix-vr4130  -mfix-sb1  -mno-fix-sb1 @gol
@@ -20804,7 +20805,8 @@ The processor names are:
 @samp{orion},
 @samp{p5600}, @samp{p6600},
 @samp{r2000}, @samp{r3000}, @samp{r3900}, @samp{r4000}, @samp{r4400},
-@samp{r4600}, @samp{r4650}, @samp{r4700}, @samp{r6000}, @samp{r8000},
+@samp{r4600}, @samp{r4650}, @samp{r4700}, @samp{r5900},
+@samp{r6000}, @samp{r8000},
 @samp{rm7000}, @samp{rm9000},
 @samp{r10000}, @samp{r12000}, @samp{r14000}, @samp{r16000},
 @samp{sb1},
@@ -21578,6 +21580,16 @@ branch-likely instructions.  @option{-mfix-r10000} is the default when
 @option{-march=r10000} is used; @option{-mno-fix-r10000} is the default
 otherwise.
 
+@item -mfix-r5900
+@itemx -mno-fix-r5900
+@opindex mfix-r5900
+Do not attempt to schedule the preceding instruction into the delay slot
+of a branch instruction placed at the end of a short loop of six
+instructions or fewer and always schedule a @code{nop} instruction there
+instead.  The short loop bug under certain conditions causes loops to
+execute only once or twice, due to a hardware bug in the R5900 chip.  The
+workaround is implemented by the assembler rather than by GCC@.
+
 @item -mfix-rm7000
 @itemx -mno-fix-rm7000
 @opindex mfix-rm7000
-- 
2.18.1