XTF: Use fastcall by default for 32bit

This passes up to 3 parameters in registers, rather than on the stack.  Most
transformations are easy.

The exec_user() infrastructure took two parameters on the stack and used an
ad-hoc %eax/%ecx arrangement with %edx as a scratch register. Fastcall uses
%eax/%edx, so switch the scratch regsiter to %ecx instead.

memop-seg was already using %eax, and needs a bit of care now that addr isn't
on the stack.  However, it does remove the need for stack_adj which simplifies
the result.

Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>

diff --git a/arch/x86/entry_32.S b/arch/x86/entry_32.S
index b76d7ef3ac7e64d5d49635c470d52bb40b1e592b..31c652626bdb15b359df52db175b1ef0d595079d 100644 (file)
--- a/arch/x86/entry_32.S
+++ b/arch/x86/entry_32.S
@@ -84,9 +84,8 @@ handle_exception:
         mov %eax, %ds
         mov %eax, %es
 
-        push %esp               /* struct cpu_regs * */
+        mov %esp, %eax          /* struct cpu_regs * */
         call do_exception
-        add $4, %esp
 
         RESTORE_ALL
 
@@ -104,17 +103,10 @@ ENTRY(entry_ret_to_kernel)      /* int $X86_VEC_RET2KERN */
         ret
 ENDFUNC(entry_ret_to_kernel)
 
-ENTRY(exec_user_param)
-        /*
-         * 2*4(%esp) ulong p1
-         * 1*4(%esp) ulong (*fn)(ulong)
-         * 0*4(%esp) return address
-         */
+ENTRY(exec_user_param)          /* %eax = ulong (*fn)(ulong p1), %edx = ulong p1 */
         push %ebp
 
         /* Prepare to "call" exec_user_stub(). */
-        mov (1+1)*4(%esp), %eax /* Pass fn() in %eax */
-        mov (1+2)*4(%esp), %ecx /* Pass p1 in %ecx */
         push $1f                /* Fake return addr as if we'd called exec_user_stub(). */
         mov %esp, %ebp          /* Stash %esp for entry_ret_to_kernel(). */
 
@@ -125,10 +117,10 @@ ENTRY(exec_user_param)
         pushf                   /* EFLAGS */
 
         /* Apply and/or masks to eflags. */
-        mov exec_user_efl_and_mask, %edx
-        and %edx, (%esp)
-        mov exec_user_efl_or_mask, %edx
-        or %edx, (%esp)
+        mov exec_user_efl_and_mask, %ecx
+        and %ecx, (%esp)
+        mov exec_user_efl_or_mask, %ecx
+        or %ecx, (%esp)
 
         push exec_user_cs       /* CS */
         push $exec_user_stub    /* EIP */
@@ -142,13 +134,9 @@ ENTRY(exec_user_param)
 ENDFUNC(exec_user_param)
 
 .pushsection .text.user, "ax", @progbits
-ENTRY(exec_user_stub)
-        /*
-         * For SMEP/SMAP safety, no shared stack can be used, so all
-         * parameters are passed in registers.
-         */
-        push %ecx               /* Push p1 for fn()'s call frame. */
-        call *%eax              /* fn(p1) */
+ENTRY(exec_user_stub)           /* %eax = ulong (*fn)(ulong p1), %edx = ulong p1 */
+        xchg %eax, %edx         /* Swap p1 to be first parameter to fn(). */
+        call *%edx              /* fn(p1) */
 
         int $X86_VEC_RET2KERN   /* Return to kernel privilege. */
 ENDFUNC(exec_user_stub)
@@ -167,9 +155,8 @@ ENTRY(entry_EVTCHN)
         mov %eax, %ds
         mov %eax, %es
 
-        push %esp               /* struct cpu_regs * */
+        mov %esp, %eax          /* struct cpu_regs * */
         call do_evtchn
-        add $4, %esp
 
         RESTORE_ALL
 
@@ -195,9 +182,8 @@ ENTRY(entry_SYSCALL)
         mov %eax, %ds
         mov %eax, %es
 
-        push %esp               /* struct cpu_regs * */
+        mov %esp, %eax          /* struct cpu_regs * */
         call do_syscall
-        add $4, %esp
 
         RESTORE_ALL
 
@@ -222,9 +208,8 @@ ENTRY(entry_SYSENTER)
         mov %eax, %ds
         mov %eax, %es
 
-        push %esp               /* struct cpu_regs * */
+        mov %esp, %eax          /* struct cpu_regs * */
         call do_sysenter
-        add $4, %esp
 
         RESTORE_ALL
 

diff --git a/arch/x86/hvm/head.S b/arch/x86/hvm/head.S
index f94dd0bc284407dec17f169e0ae91a1861062104..a6bb5f56829fd20df839a19ea25de83dfaa8a1d7 100644 (file)
--- a/arch/x86/hvm/head.S
+++ b/arch/x86/hvm/head.S
@@ -73,7 +73,7 @@ GLOBAL(_elf_start)              /* HVM common setup. */
 #ifdef __x86_64__
         lea .Lmain_err_msg(%rip), %rdi
 #else
-        push $.Lmain_err_msg
+        mov $.Lmain_err_msg, %eax
 #endif
         call panic
 ENDFUNC(_elf_start)

diff --git a/arch/x86/pv/head.S b/arch/x86/pv/head.S
index 0b2592f9691c139e74142addf2353e1ef0b146a5..36c9a05fc4287a630d4f890f6fa34b6ed1e6cf27 100644 (file)
--- a/arch/x86/pv/head.S
+++ b/arch/x86/pv/head.S
@@ -35,7 +35,7 @@ GLOBAL(_elf_start)
 #ifdef __x86_64__
         lea .Lmain_err_msg(%rip), %rdi
 #else
-        push $.Lmain_err_msg
+        mov $.Lmain_err_msg, %eax
 #endif
         call panic
 ENDFUNC(_elf_start)

diff --git a/build/common.mk b/build/common.mk
index 644961c03fa835cd48e8709a7cfd08480f3e1684..a6e270f567cde220d959bb47d5c91ba654fda98f 100644 (file)
--- a/build/common.mk
+++ b/build/common.mk
@@ -43,7 +43,7 @@ COMMON_CFLAGS += -Wno-unused-parameter -Winline
 COMMON_AFLAGS-x86_32 := -m32
 COMMON_AFLAGS-x86_64 := -m64
 
-COMMON_CFLAGS-x86_32 := -m32
+COMMON_CFLAGS-x86_32 := -m32 -mregparm=3
 COMMON_CFLAGS-x86_64 := -m64
 
 defcfg-pv    := $(ROOT)/config/default-pv.cfg.in

diff --git a/tests/memop-seg/asm.S b/tests/memop-seg/asm.S
index 08e87b2ea2ee27765076a9e6c44b981dfd421281..be9fb09d54418a12a400a72494318a67732563bc 100644 (file)
--- a/tests/memop-seg/asm.S
+++ b/tests/memop-seg/asm.S
@@ -12,9 +12,6 @@ ENTRY(stub_\seg\()_abs) /* exinfo_t stub_\seg_abs(unsigned long addr) */
          * Switch segment if necessary.  The old segment is preserved on the
          * stack for the duration of the test.
          */
-        .local stack_adj
-        stack_adj = 0
-
         .if \load_seg
             .ifeqs "\seg", "none"
                 push %ds
@@ -25,19 +22,22 @@ ENTRY(stub_\seg\()_abs) /* exinfo_t stub_\seg_abs(unsigned long addr) */
                 push $(GDTE_AVAIL1 << 3 | 3)
                 pop  %\seg
             .endif
-            stack_adj = 1
         .endif
 #endif
 
-        /* No exception if we don't fault.  Also reused by the 64bit case. */
-        xor %eax, %eax
-
         /* The bottom bit of 'addr' encodes FEP. */
 #ifdef __i386__
-        testb $1, (1 + stack_adj)*4(%esp)
+        testb $1, %al
 #else
         testb $1, %dil
 #endif
+
+        /*
+         * No exception if we don't fault.
+         * Reused by the 64bit case, and careful to not clobber flags.
+         */
+        mov $0, %eax
+
         jz 1f
 
         _ASM_XEN_FEP
@@ -82,9 +82,6 @@ ENTRY(stub_\seg\()_\reg) /* exinfo_t stub_\seg_\reg(unsigned long addr) */
          * Switch segment if necessary.  The old segment is preserved on the
          * stack for the duration of the test.
          */
-        .local stack_adj
-        stack_adj = 0
-
         .if \load_seg
             .ifeqs "\seg", "none"
                 push %ds
@@ -95,7 +92,6 @@ ENTRY(stub_\seg\()_\reg) /* exinfo_t stub_\seg_\reg(unsigned long addr) */
                 push $(GDTE_AVAIL1 << 3 | 3)
                 pop  %\seg
             .endif
-            stack_adj = 1
         .endif
 #endif
 
@@ -106,7 +102,7 @@ ENTRY(stub_\seg\()_\reg) /* exinfo_t stub_\seg_\reg(unsigned long addr) */
 
         /* Move 'addr' into \reg */
 #ifdef __i386__
-        mov (1 + stack_adj)*4(%esp), %\reg
+        mov %eax, %\reg
 #else
         mov %rdi, %\reg
 #endif

diff --git a/tests/nmi-taskswitch-priv/main.c b/tests/nmi-taskswitch-priv/main.c
index b2e43f9c234afcd472b41a9daa84f7b5596aa6dd..299117c5a1c8fce391bf46407b4a65a5d64b89f1 100644 (file)
--- a/tests/nmi-taskswitch-priv/main.c
+++ b/tests/nmi-taskswitch-priv/main.c
@@ -54,9 +54,8 @@ asm("exit_NMI_task:"
     "push %ebx;"
     "push %ebp;"
 
-    "push %esp;"
+    "mov %esp, %eax;"
     "call do_exception;"
-    "add $1*4, %esp;"
 
     "pop %ebp;"
     "pop %ebx;"