[PATCH 0/2] MIPS: Use Octeon BBIT instructions in TLB handlers.

[PATCH 0/2] MIPS: Use Octeon BBIT instructions in TLB handlers.

[David Daney]

Octeon has instructions that conditionally branch based on the value
of any single bit in any register.  We use these to reduce the number
of instructions in the generated TLB handlers.

This set applies on top of the recent KScratch patch set.

David Daney (2):
  MIPS: Declare uasm bbit0 and bbit1 functions.
  MIPS: Use BBIT instructions in TLB handlers

 arch/mips/include/asm/uasm.h |    2 +
 arch/mips/mm/tlbex.c         |  119 +++++++++++++++++++++++++++++++----------
 2 files changed, 92 insertions(+), 29 deletions(-)

-- 
1.7.2.3

[PATCH 1/2] MIPS: Declare uasm bbit0 and bbit1 functions.

[David Daney]

these are already defined, but declaring them allow them to be used
outside of uasm.c.

Signed-off-by: David Daney <ddaney@caviumnetworks.com>
---
 arch/mips/include/asm/uasm.h |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)

diff --git a/arch/mips/include/asm/uasm.h b/arch/mips/include/asm/uasm.h
index 99dae68..d361df3 100644
--- a/arch/mips/include/asm/uasm.h
+++ b/arch/mips/include/asm/uasm.h
@@ -117,6 +117,8 @@ Ip_u2u1u3(_xori);
 Ip_u2u1msbu3(_dins);
 Ip_u2u1msbu3(_dinsm);
 Ip_u1(_syscall);
+Ip_u1u2s3(_bbit0);
+Ip_u1u2s3(_bbit1);
 
 /* Handle labels. */
 struct uasm_label {
-- 
1.7.2.3

[PATCH 2/2] MIPS: Use BBIT instructions in TLB handlers

[David Daney]

If the CPU supports BBIT0 and BBIT1, use them in TLB handlers as they
are more efficient than an AND followed by an branch and then
restoring the clobbered register.

Signed-off-by: David Daney <ddaney@caviumnetworks.com>
---
 arch/mips/mm/tlbex.c |  119 +++++++++++++++++++++++++++++++++++++------------
 1 files changed, 90 insertions(+), 29 deletions(-)

diff --git a/arch/mips/mm/tlbex.c b/arch/mips/mm/tlbex.c
index cec0e1b..601f4c2 100644
--- a/arch/mips/mm/tlbex.c
+++ b/arch/mips/mm/tlbex.c
@@ -65,6 +65,18 @@ static inline int __maybe_unused r10000_llsc_war(void)
 	return R10000_LLSC_WAR;
 }
 
+static int use_bbit_insns(void)
+{
+	switch (current_cpu_type()) {
+	case CPU_CAVIUM_OCTEON:
+	case CPU_CAVIUM_OCTEON_PLUS:
+	case CPU_CAVIUM_OCTEON2:
+		return 1;
+	default:
+		return 0;
+	}
+}
+
 /*
  * Found by experiment: At least some revisions of the 4kc throw under
  * some circumstances a machine check exception, triggered by invalid
@@ -507,8 +519,12 @@ build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
 		unsigned int pmd, int lid)
 {
 	UASM_i_LW(p, tmp, 0, pmd);
-	uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
-	uasm_il_bnez(p, r, tmp, lid);
+	if (use_bbit_insns()) {
+		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
+	} else {
+		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
+		uasm_il_bnez(p, r, tmp, lid);
+	}
 }
 
 static __cpuinit void build_huge_update_entries(u32 **p,
@@ -1183,14 +1199,20 @@ build_pte_present(u32 **p, struct uasm_reloc **r,
 		  unsigned int pte, unsigned int ptr, enum label_id lid)
 {
 	if (kernel_uses_smartmips_rixi) {
-		uasm_i_andi(p, pte, pte, _PAGE_PRESENT);
-		uasm_il_beqz(p, r, pte, lid);
+		if (use_bbit_insns()) {
+			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
+			uasm_i_nop(p);
+		} else {
+			uasm_i_andi(p, pte, pte, _PAGE_PRESENT);
+			uasm_il_beqz(p, r, pte, lid);
+			iPTE_LW(p, pte, ptr);
+		}
 	} else {
 		uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
 		uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
 		uasm_il_bnez(p, r, pte, lid);
+		iPTE_LW(p, pte, ptr);
 	}
-	iPTE_LW(p, pte, ptr);
 }
 
 /* Make PTE valid, store result in PTR. */
@@ -1211,10 +1233,17 @@ static void __cpuinit
 build_pte_writable(u32 **p, struct uasm_reloc **r,
 		   unsigned int pte, unsigned int ptr, enum label_id lid)
 {
-	uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
-	uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
-	uasm_il_bnez(p, r, pte, lid);
-	iPTE_LW(p, pte, ptr);
+	if (use_bbit_insns()) {
+		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
+		uasm_i_nop(p);
+		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
+		uasm_i_nop(p);
+	} else {
+		uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
+		uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
+		uasm_il_bnez(p, r, pte, lid);
+		iPTE_LW(p, pte, ptr);
+	}
 }
 
 /* Make PTE writable, update software status bits as well, then store
@@ -1238,9 +1267,14 @@ static void __cpuinit
 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
 		     unsigned int pte, unsigned int ptr, enum label_id lid)
 {
-	uasm_i_andi(p, pte, pte, _PAGE_WRITE);
-	uasm_il_beqz(p, r, pte, lid);
-	iPTE_LW(p, pte, ptr);
+	if (use_bbit_insns()) {
+		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
+		uasm_i_nop(p);
+	} else {
+		uasm_i_andi(p, pte, pte, _PAGE_WRITE);
+		uasm_il_beqz(p, r, pte, lid);
+		iPTE_LW(p, pte, ptr);
+	}
 }
 
 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
@@ -1485,14 +1519,23 @@ static void __cpuinit build_r4000_tlb_load_handler(void)
 		 * If the page is not _PAGE_VALID, RI or XI could not
 		 * have triggered it.  Skip the expensive test..
 		 */
-		uasm_i_andi(&p, K0, K0, _PAGE_VALID);
-		uasm_il_beqz(&p, &r, K0, label_tlbl_goaround1);
+		if (use_bbit_insns()) {
+			uasm_il_bbit0(&p, &r, K0, ilog2(_PAGE_VALID),
+				      label_tlbl_goaround1);
+		} else {
+			uasm_i_andi(&p, K0, K0, _PAGE_VALID);
+			uasm_il_beqz(&p, &r, K0, label_tlbl_goaround1);
+		}
 		uasm_i_nop(&p);
 
 		uasm_i_tlbr(&p);
 		/* Examine  entrylo 0 or 1 based on ptr. */
-		uasm_i_andi(&p, K0, K1, sizeof(pte_t));
-		uasm_i_beqz(&p, K0, 8);
+		if (use_bbit_insns()) {
+			uasm_i_bbit0(&p, K1, ilog2(sizeof(pte_t)), 8);
+		} else {
+			uasm_i_andi(&p, K0, K1, sizeof(pte_t));
+			uasm_i_beqz(&p, K0, 8);
+		}
 
 		UASM_i_MFC0(&p, K0, C0_ENTRYLO0); /* load it in the delay slot*/
 		UASM_i_MFC0(&p, K0, C0_ENTRYLO1); /* load it if ptr is odd */
@@ -1500,12 +1543,18 @@ static void __cpuinit build_r4000_tlb_load_handler(void)
 		 * If the entryLo (now in K0) is valid (bit 1), RI or
 		 * XI must have triggered it.
 		 */
-		uasm_i_andi(&p, K0, K0, 2);
-		uasm_il_bnez(&p, &r, K0, label_nopage_tlbl);
-
-		uasm_l_tlbl_goaround1(&l, p);
-		/* Reload the PTE value */
-		iPTE_LW(&p, K0, K1);
+		if (use_bbit_insns()) {
+			uasm_il_bbit1(&p, &r, K0, 1, label_nopage_tlbl);
+			/* Reload the PTE value */
+			iPTE_LW(&p, K0, K1);
+			uasm_l_tlbl_goaround1(&l, p);
+		} else {
+			uasm_i_andi(&p, K0, K0, 2);
+			uasm_il_bnez(&p, &r, K0, label_nopage_tlbl);
+			uasm_l_tlbl_goaround1(&l, p);
+			/* Reload the PTE value */
+			iPTE_LW(&p, K0, K1);
+		}
 	}
 	build_make_valid(&p, &r, K0, K1);
 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
@@ -1525,23 +1574,35 @@ static void __cpuinit build_r4000_tlb_load_handler(void)
 		 * If the page is not _PAGE_VALID, RI or XI could not
 		 * have triggered it.  Skip the expensive test..
 		 */
-		uasm_i_andi(&p, K0, K0, _PAGE_VALID);
-		uasm_il_beqz(&p, &r, K0, label_tlbl_goaround2);
+		if (use_bbit_insns()) {
+			uasm_il_bbit0(&p, &r, K0, ilog2(_PAGE_VALID),
+				      label_tlbl_goaround2);
+		} else {
+			uasm_i_andi(&p, K0, K0, _PAGE_VALID);
+			uasm_il_beqz(&p, &r, K0, label_tlbl_goaround2);
+		}
 		uasm_i_nop(&p);
 
 		uasm_i_tlbr(&p);
 		/* Examine  entrylo 0 or 1 based on ptr. */
-		uasm_i_andi(&p, K0, K1, sizeof(pte_t));
-		uasm_i_beqz(&p, K0, 8);
-
+		if (use_bbit_insns()) {
+			uasm_i_bbit0(&p, K1, ilog2(sizeof(pte_t)), 8);
+		} else {
+			uasm_i_andi(&p, K0, K1, sizeof(pte_t));
+			uasm_i_beqz(&p, K0, 8);
+		}
 		UASM_i_MFC0(&p, K0, C0_ENTRYLO0); /* load it in the delay slot*/
 		UASM_i_MFC0(&p, K0, C0_ENTRYLO1); /* load it if ptr is odd */
 		/*
 		 * If the entryLo (now in K0) is valid (bit 1), RI or
 		 * XI must have triggered it.
 		 */
-		uasm_i_andi(&p, K0, K0, 2);
-		uasm_il_beqz(&p, &r, K0, label_tlbl_goaround2);
+		if (use_bbit_insns()) {
+			uasm_il_bbit0(&p, &r, K0, 1, label_tlbl_goaround2);
+		} else {
+			uasm_i_andi(&p, K0, K0, 2);
+			uasm_il_beqz(&p, &r, K0, label_tlbl_goaround2);
+		}
 		/* Reload the PTE value */
 		iPTE_LW(&p, K0, K1);
 
-- 
1.7.2.3

Re: [PATCH 1/2] MIPS: Declare uasm bbit0 and bbit1 functions.

[Ralf Baechle]

On Mon, Dec 20, 2010 at 03:54:49PM -0800, David Daney wrote:

> these are already defined, but declaring them allow them to be used
> outside of uasm.c.
> 
> Signed-off-by: David Daney <ddaney@caviumnetworks.com>
> ---
>  arch/mips/include/asm/uasm.h |    2 ++

Queued for 2.6.38 - but a few of your patches to uasm.h only apply with
fuzz so I'm wondering if I'm missing a patch or what your patches were
created against.

Thanks,

  Ralf

Re: [PATCH 2/2] MIPS: Use BBIT instructions in TLB handlers

[Ralf Baechle]

On Mon, Dec 20, 2010 at 03:54:50PM -0800, David Daney wrote:

> If the CPU supports BBIT0 and BBIT1, use them in TLB handlers as they
> are more efficient than an AND followed by an branch and then
> restoring the clobbered register.

Queued for 2.6.38.  Thanks,

  Ralf