qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH] target-mips: apply workaround for TCG optimizat


From: Aurelien Jarno
Subject: Re: [Qemu-devel] [PATCH] target-mips: apply workaround for TCG optimizations for MFC1
Date: Wed, 15 Jul 2015 12:14:47 +0200
User-agent: Mutt/1.5.23 (2014-03-12)

On 2015-07-15 11:02, Richard Henderson wrote:
> On 07/15/2015 09:06 AM, Aurelien Jarno wrote:
> >On 2015-07-15 09:31, Paolo Bonzini wrote:
> >>Ok, I see your point.  If you put it like this :) the fault definitely
> >>lies in the backends.  What I'm proposing would be in a new
> >>tcg_reg_alloc_trunc function, and it would require implementing a
> >>non-noop trunc.
> >
> >Why not reusing the existing trunc_shr_i64_i32 op? AFAIU, it has been
> >designed exactly for that.
> 
> Indeed.
> 
> >Actually I think we should implement the following ops as optional but
> >*real* TCG ops:
> >- trunc_shr_i64_i32
> >- extu_i32_i64
> >- ext_i32_i64
> 
> While we could perhaps gain something from the last two, reliably using the
> first is probably the most important.
>
> I've been unable to reproduce the binary in question.  I'm curious if
> something as simple as this helps.  Alternately, the two hunks might just
> cancel each other out and result in no change.

No, besides that tcg_assert doesn't exist, it doesn't work, as it is
later optimized out. What I meant is to implement trunc_shr_i64_i32 in
the x86 backend, either as a 32-bit zero extension when count is 0, or a
64-bit shift zero extended otherwise.

I think for 2.4 we should go with the ADDR32 way, using the following
kind of patch. It works when %gs is in use, but has to be extended when
it's not in use. The idea is that qemu_ld/st should behave like other
32-bit ops, that is ignore the high part of the register.

diff --git a/tcg/i386/tcg-target.c b/tcg/i386/tcg-target.c
index ff4d9cf..e139c44 100644
--- a/tcg/i386/tcg-target.c
+++ b/tcg/i386/tcg-target.c
@@ -1572,6 +1572,9 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg 
*args, bool is64)
         if (GUEST_BASE && guest_base_flags) {
             seg = guest_base_flags;
             offset = 0;
+            if (TARGET_LONG_BITS == 32) {
+                seg = P_ADDR32;
+            }
         } else if (TCG_TARGET_REG_BITS == 64 && offset != GUEST_BASE) {
             tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_L1, GUEST_BASE);
             tgen_arithr(s, ARITH_ADD + P_REXW, TCG_REG_L1, base);
@@ -1705,6 +1708,9 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg 
*args, bool is64)
         if (GUEST_BASE && guest_base_flags) {
             seg = guest_base_flags;
             offset = 0;
+            if (TARGET_LONG_BITS == 32) {
+                seg = P_ADDR32;
+            }
         } else if (TCG_TARGET_REG_BITS == 64 && offset != GUEST_BASE) {
             tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_L1, GUEST_BASE);
             tgen_arithr(s, ARITH_ADD + P_REXW, TCG_REG_L1, base);


-- 
Aurelien Jarno                          GPG: 4096R/1DDD8C9B
address@hidden                 http://www.aurel32.net



reply via email to

[Prev in Thread] Current Thread [Next in Thread]