From 519786ec38bb5123547071021e4bd7678d7673ad Mon Sep 17 00:00:00 2001

From: Andreas Krebbel <krebbel@linux.ibm.com>

Date: Mon, 23 Mar 2020 09:08:27 +0100

Subject: [PATCH 3/4] Optimize s390x instructions

---

 src/asm/jump_s390x_sysv_elf_gas.S  | 84 ++++++++++--------------------

 src/asm/make_s390x_sysv_elf_gas.S  | 27 ++++++----

 src/asm/ontop_s390x_sysv_elf_gas.S | 81 +++++++++-------------------

 3 files changed, 70 insertions(+), 122 deletions(-)

diff --git a/src/asm/jump_s390x_sysv_elf_gas.S b/src/asm/jump_s390x_sysv_elf_gas.S

index c011d53..b2163cc 100644

--- a/libs/context/src/asm/jump_s390x_sysv_elf_gas.S

+++ b/libs/context/src/asm/jump_s390x_sysv_elf_gas.S

@@ -36,47 +36,34 @@

 .global jump_fcontext

 .type   jump_fcontext, @function

+#define GR_OFFSET	0

+#define LR_OFFSET	64

+#define SP_OFFSET	72

+#define FP_OFFSET	80

+#define PC_OFFSET	112

+#define L_CTX		120

+#define L_STACKFRAME	120

+

 jump_fcontext:

     # Reserved the space for stack to store the data of current context

     # before we jump to the new context.

-    lay 15,-120(15)

+    aghi %r15,-L_STACKFRAME

     # save the registers to the stack

-    stg 6,  0(15)       # save R6     

-    stg 7,  8(15)       # save R7     

-    stg 8,  16(15)      # save R8

-    stg 9,  24(15)      # save R9

-    stg 10, 32(15)      # save R10

-    stg 11, 40(15)      # save R11

-    stg 12, 48(15)      # save R12

-    stg 13, 56(15)      # save R13

-    stg 14, 64(15)      # save R14

-    stg 15, 72(15)      # save R15

+    stmg %r6, %r15, GR_OFFSET(%r15)

     # save the floating point registers

-    # Load the FPR into R0 then save it to the stack

-    # Load F1 into R0

-    lgdr 0,1 

-    stg  0,80(15)       # save F1

-

-    # Load F3 into R0

-    lgdr 0,3

-    stg  0,88(15)       # save F3

-

-    # Load F5 into R0

-    lgdr 0,5

-    stg  0,96(15)       # save F5

-

-    # Load F7 into R0

-    lgdr 0,7

-    stg  0,104(15)      # save F7

+    std  %f0,FP_OFFSET(%r15)

+    std  %f3,FP_OFFSET+8(%r15)

+    std  %f5,FP_OFFSET+16(%r15)

+    std  %f7,FP_OFFSET+24(%r15)

     # Save LR as PC

-    stg 14,112(15)

+    stg  %r14,PC_OFFSET(%r15)

     # Store the SP pointing to the old context-data into R0

-    lgr 0,15

+    lgr	 %r0,%r15

     # Get the SP pointing to the new context-data

     # Note: Since the return type of the jump_fcontext is struct whose

@@ -88,46 +75,31 @@ jump_fcontext:

     # R2 --> Address of the return transfer_t struct

     # R3 --> Context we want to switch to

     # R4 --> Data

-    lgr 15,3

+    lgr	%r15,%r3

     # Load the registers with the data present in context-data of the

     # context we are going to switch to

-    lg 6,  0(15)       # restore R6     

-    lg 7,  8(15)       # restore R7     

-    lg 8,  16(15)      # restore R8

-    lg 9,  24(15)      # restore R9

-    lg 10, 32(15)      # restore R10

-    lg 11, 40(15)      # restore R11

-    lg 12, 48(15)      # restore R12

-    lg 13, 56(15)      # restore R13

-    lg 14, 64(15)      # restore R14

+    lmg	%r6, %r14, GR_OFFSET(%r15)

     # Restore Floating point registers

-    lg   1,80(15)

-    ldgr 1,1            # restore F1

-

-    lg   1,88(15)

-    ldgr 1,3            # restore F3

-

-    lg   1,96(15)

-    ldgr 1,5            # restore F5

-

-    lg   1,104(15)

-    ldgr 1,7            # restore F7

+    ld	 %f1,FP_OFFSET(%r15)

+    ld	 %f3,FP_OFFSET+8(%r15)

+    ld	 %f5,FP_OFFSET+16(%r15)

+    ld	 %f7,FP_OFFSET+24(%r15)

     # Load PC

-    lg  1,112(15)

+    lg   %r1,PC_OFFSET(%r15)

-    # Adjust the stack

-    lay 15, 120(15)

+    # Adjust the stack 

+    aghi %r15,120

     # R2 --> Address where the return transfer_t is stored

     # R0 --> FCTX

     # R4 --> DATA

     # Store the elements to return transfer_t

-    stg 15, 0(2)

-    stg 4, 8(2)

+    stg %r15, 0(%r2)

+    stg %r4, 8(%r2)

     # Note: The address in R2 points to the place where the return

     # transfer_t is stored. Since context_function take transfer_t

@@ -135,7 +107,7 @@ jump_fcontext:

     # first parameter value.

     #jump to context

-    br 1

+    br  %r1

 .size   jump_fcontext,.-jump_fcontext

 # Mark that we don't need executable stack.

diff --git a/src/asm/make_s390x_sysv_elf_gas.S b/src/asm/make_s390x_sysv_elf_gas.S

index f566533..d02856c 100644

--- a/libs/context/src/asm/make_s390x_sysv_elf_gas.S

+++ b/libs/context/src/asm/make_s390x_sysv_elf_gas.S

@@ -36,6 +36,14 @@

 .global make_fcontext

 .type 	 make_fcontext, @function

+#define GR_OFFSET	0

+#define LR_OFFSET	64

+#define SP_OFFSET	72

+#define FP_OFFSET	80

+#define PC_OFFSET	112

+#define L_CTX		120

+#define L_STACKFRAME	120

+

 make_fcontext:

 		# make_fcontext takes in 3 arguments

@@ -56,40 +64,39 @@ make_fcontext:

 		# address is zero or not. If not AND it with `-8`. 

 		# Here we AND the lower 16 bits of the memory address present in the 

-		# R2 with the bits 1111 1111 1111 1000 which when converted into

-		# decimal is 65528

-		nill    2,65528

+		# R2 with the bits 1111 1111 1111 1000 

+		nill    %r2,0xfff0

 		# Reserve space for context-data on context-stack.

 		# This is done by shifting the SP/address by 112 bytes.

-		lay 2,-120(2)

+		aghi	%r2,-L_CTX

 		# third arg of make_fcontext() == address of the context-function

 		# Store the address as a PC to jump in, whenever we call the 

 		# make_fcontext.

-		stg 4,112(2)

+		stg 	%r4,PC_OFFSET(%r2)

 		# Save the address of finish as return-address for context-function

 		# This will be entered after context-function return

 		# The address of finish will be saved in Link register, this register

 		# specifies where we need to jump after the function executes

 		# completely.

-		larl 1,finish

-		stg  1,64(2)

+		larl 	%r1,finish

+		stg  	%r1,LR_OFFSET(%r2)

 		# Return pointer to context data

 		# R14 acts as the link register

 		# R2 holds the address of the context stack. When we return from the

 		# make_fcontext, R2 is passed back.

-		br 14 

+		br 	%r14 

 	finish:

 		# In finish tasks, you load the exit code and exit the make_fcontext

 		# This is called when the context-function is entirely executed

-		lghi 2,0

-		brasl 14,_exit

+		lghi 	%r2,0

+		brasl 	%r14,_exit@PLT

 .size   make_fcontext,.-make_fcontext

 # Mark that we don't need executable stack.

diff --git a/src/asm/ontop_s390x_sysv_elf_gas.S b/src/asm/ontop_s390x_sysv_elf_gas.S

index 7ab2cf5..4488654 100644

--- a/libs/context/src/asm/ontop_s390x_sysv_elf_gas.S

+++ b/libs/context/src/asm/ontop_s390x_sysv_elf_gas.S

@@ -36,47 +36,32 @@

 .global ontop_fcontext

 .type   ontop_fcontext, @function

+#define GR_OFFSET	0

+#define LR_OFFSET	64

+#define SP_OFFSET	72

+#define FP_OFFSET	80

+#define PC_OFFSET	112

+#define L_CTX		120

+

 ontop_fcontext:

     # Reserved the space for stack to store the data of current context

     # before we jump to the new context.

-    lay 15,-120(15)

+    aghi %r15,-L_CTX

     # save the registers to the stack

-    stg 6,  0(15)       # save R6     

-    stg 7,  8(15)       # save R7     

-    stg 8,  16(15)      # save R8

-    stg 9,  24(15)      # save R9

-    stg 10, 32(15)      # save R10

-    stg 11, 40(15)      # save R11

-    stg 12, 48(15)      # save R12

-    stg 13, 56(15)      # save R13

-    stg 14, 64(15)      # save R14

-    stg 15, 72(15)      # save R15

+    stmg %r6, %r15, GR_OFFSET(%r15)

     # save the floating point registers

-    # Load the FPR into R0 then save it to the stack

-    # Load F1 into R0

-    lgdr 0,1 

-    stg  0,80(15)       # save F1

-

-    # Load F3 into R0

-    lgdr 0,3

-    stg  0,88(15)       # save F3

-

-    # Load F5 into R0

-    lgdr 0,5

-    stg  0,96(15)       # save F5

-

-    # Load F7 into R0

-    lgdr 0,7

-    stg  0,104(15)      # save F7

-

+    std  %f0,FP_OFFSET(%r15)

+    std  %f3,FP_OFFSET+8(%r15)

+    std  %f5,FP_OFFSET+16(%r15)

+    std  %f7,FP_OFFSET+24(%r15)

     # Save LR as PC

-    stg 14,112(15)

+    stg  %r14,PC_OFFSET(%r15)

     # Store the SP pointing to the old context-data into R0

-    lgr 0,15

+    lgr  %r0,%r15

     # Get the SP pointing to the new context-data

     # Note: Since the return type of the jump_fcontext is struct whose

@@ -88,38 +73,22 @@ ontop_fcontext:

     # R2 --> Address of the return transfer_t struct

     # R3 --> Context we want to switch to

     # R4 --> Data

-    lgr 15,3

+    lgr  %r15,%r3

     # Load the registers with the data present in context-data of the

     # context we are going to switch to

-    lg 6,  0(15)       # restore R6     

-    lg 7,  8(15)       # restore R7     

-    lg 8,  16(15)      # restore R8

-    lg 9,  24(15)      # restore R9

-    lg 10, 32(15)      # restore R10

-    lg 11, 40(15)      # restore R11

-    lg 12, 48(15)      # restore R12

-    lg 13, 56(15)      # restore R13

-    lg 14, 64(15)      # restore R14

-    lg 15, 72(15)      # restore R15

+    lmg  %r6,%r15,GR_OFFSET(%r15)

     # Restore Floating point registers

-    lg   1,80(15)

-    ldgr 1,1            # restore F1

-

-    lg   1,88(15)

-    ldgr 1,3            # restore F3

-

-    lg   1,96(15)

-    ldgr 1,5            # restore F5

-

-    lg   1,104(15)

-    ldgr 1,7            # restore F7

+    ld	 %f1,FP_OFFSET(%r15)

+    ld	 %f3,FP_OFFSET+8(%r15)

+    ld	 %f5,FP_OFFSET+16(%r15)

+    ld	 %f7,FP_OFFSET+24(%r15)

     # Skip PC

     # Adjust the stack

-    lay 15, 120(15)

+    aghi %r15,L_CTX

     # R2 --> Address where the return transfer_t is stored

     # R0 --> FCTX

@@ -127,8 +96,8 @@ ontop_fcontext:

     # R5 --> Context function

     # Store the elements to return transfer_t

-    stg 15, 0(2)

-    stg 4, 8(2)

+    stg  %r15, 0(%r2)

+    stg  %r4, 8(%r2)

     # Note: The address in R2 points to the place where the return

     # transfer_t is stored. Since context_function take transfer_t

@@ -136,7 +105,7 @@ ontop_fcontext:

     # first parameter value.

     #jump to context function

-    br 5

+    br 	%r5

 .size   ontop_fcontext,.-ontop_fcontext

 # Mark that we don't need executable stack.

-- 

2.18.1