optimize x86 feclearexcept: only use save/restore x87 fenv if needed

the x87 exception summary (ES) and stack fault (SF) flags may be
spuriously cleared by feclearexcept using the fnclex instruction,
but these flags are not observable through libc hence maintaining
their state is not critical.

diff --git a/src/fenv/i386/fenv.s b/src/fenv/i386/fenv.s
index 9bba40a5e3a01f2ffeb25a7804383b40a6be2b66..a8540add424a3dcb30bd6e971919af6d3efac76a 100644 (file)
--- a/src/fenv/i386/fenv.s
+++ b/src/fenv/i386/fenv.s
@@ -4,26 +4,41 @@
 .type feclearexcept,@function
 feclearexcept: 
        mov 4(%esp),%ecx
-       not %ecx
+       fnstsw %ax
                # consider sse fenv as well if the cpu has XMM capability
        call 1f
 1:     addl $__hwcap-1b,(%esp)
        pop %edx
        testl $0x02000000,(%edx)
+       jz 2f
+               # maintain exceptions in the sse mxcsr, clear x87 exceptions
+       test %eax,%ecx
        jz 1f
-       stmxcsr 4(%esp)
-       and %ecx,4(%esp)
-       ldmxcsr 4(%esp)
-1:     test $0x3f,%ecx
-       jnz 2f
-1:     fnclex
-       xor %eax,%eax
+       fnclex
+1:     push %edx
+       stmxcsr (%esp)
+       pop %edx
+       and $0x3f,%eax
+       or %eax,%edx
+       test %edx,%ecx
+       jz 1f
+       not %ecx
+       and %ecx,%edx
+       push %edx
+       ldmxcsr (%esp)
+       pop %edx
+1:     xor %eax,%eax
        ret
-2:     fnstsw %ax
-               # TODO: only load/store fenv if exceptions arent clear yet
-       and %ecx,%eax
+               # only do the expensive x87 fenv load/store when needed
+2:     test %eax,%ecx
        jz 1b
-       sub $32,%esp
+       not %ecx
+       and %ecx,%eax
+       test $0x3f,%eax
+       jz 1f
+       fnclex
+       jmp 1b
+1:     sub $32,%esp
        fnstenv (%esp)
        mov %al,4(%esp)
        fldenv (%esp)

diff --git a/src/fenv/x86_64/fenv.s b/src/fenv/x86_64/fenv.s
index c48dade394c2e001f14ccf35c3b5ed8b9e8bd6d4..dda6b61a0b0bf61d9fce5bfeb3d605616c4d6bf4 100644 (file)
--- a/src/fenv/x86_64/fenv.s
+++ b/src/fenv/x86_64/fenv.s
@@ -1,25 +1,21 @@
 .global feclearexcept
 .type feclearexcept,@function
 feclearexcept:
+               # maintain exceptions in the sse mxcsr, clear x87 exceptions
        mov %edi,%ecx
+       fnstsw %ax
+       test %eax,%ecx
+       jz 1f
+       fnclex
+1:     stmxcsr -8(%rsp)
+       and $0x3f,%eax
+       or %eax,-8(%rsp)
+       test %ecx,-8(%rsp)
+       jz 1f
        not %ecx
-       stmxcsr -8(%rsp)
        and %ecx,-8(%rsp)
        ldmxcsr -8(%rsp)
-       test $0x3f,%ecx
-       jnz 2f
-1:     fnclex
-       xor %eax,%eax
-       ret
-2:     fnstsw %ax
-       and %ecx,%eax
-       jz 1b
-       sub $32,%rsp
-       fnstenv (%rsp)
-       mov %al,4(%rsp)
-       fldenv (%rsp)
-       add $32,%rsp
-       xor %eax,%eax
+1:     xor %eax,%eax
        ret
 
 .global feraiseexcept