Fix extlh instruction on Alpha
PATCH 1/1 Fix extlh instruction on Alpha
(re-sending)
The extlh instruction on Alpha currently doesn't work properly. It's a combination of a cut/paste bug (16 where it should be 32) as well as a "shift by 64" bug.
Below is a patch that fixes the problem. The previous e-mails on this problem have test cases that exhibit the bug.
This patch uses tcg_temp_local_new() at the suggestion of Filip Navara.
Vince
target-alpha/translate.c +12 -2
@@ -526,14 +526,24 @@ static always_inline void gen_ext_h(void (*tcg_gen_ext_i64)(TCGv t0, TCGv t1),else tcg_gen_mov_i64 (cpu_ir[rc], cpu_ir[ra]); }else { +int l1;TCGv tmp1, tmp2; - tmp1 =tcg_temp_new (); + tmp1 =tcg_temp_local_new (); + l1 =gen_new_label (); +tcg_gen_andi_i64 (tmp1, cpu_ir[rb],7 );tcg_gen_shli_i64 (tmp1, tmp1,3 ); + +tcg_gen_mov_i64 (cpu_ir[rc], cpu_ir[ra]); +tcg_gen_brcondi_i64 (TCG_COND_EQ , tmp1,0 , l1); + tmp2 =tcg_const_i64 (64 );tcg_gen_sub_i64 (tmp1, tmp2, tmp1);tcg_temp_free (tmp2);tcg_gen_shl_i64 (cpu_ir[rc], cpu_ir[ra], tmp1); + +gen_set_label (l1); +tcg_temp_free (tmp1); }if (tcg_gen_ext_i64)
@@ -1320,7 +1330,7 @@ static always_inline int translate_one (DisasContext *ctx, uint32_t insn)break ;case 0x6A :/* EXTLH */ -gen_ext_h (&tcg_gen_ext16u_i64, ra, rb, rc, islit, lit); +gen_ext_h (&tcg_gen_ext32u_i64, ra, rb, rc, islit, lit);break ;case 0x72 :/* MSKQH */
Quoted 29 lines · Vince Weaver
(re-sending)
The extlh instruction on Alpha currently doesn't work properly. It's a combination of a cut/paste bug (16 where it should be 32) as well as a "shift by 64" bug.
Below is a patch that fixes the problem. The previous e-mails on this problem have test cases that exhibit the bug.
This patch uses
tcg_temp_local_new()at the suggestion of Filip Navara.Vince
Signed-off-by Vince Weaver <[email protected]>target-alpha/translate.c +3
-1@@ -526,14 +526,24 @@ static always_inline void gen_ext_h(void (*tcg_gen_ext_i64)(TCGv t0, TCGv t1),else tcg_gen_mov_i64 (cpu_ir[rc], cpu_ir[ra]); }else { +int l1;TCGv tmp1, tmp2; - tmp1 =tcg_temp_new (); + tmp1 =tcg_temp_local_new (); + l1 =gen_new_label ();
target-alpha/translate.c static always_inline void gen_ext_h(void (*tcg_gen_ext_i64)(TCGv t0, TCGv t1), · line 542
Quoted 10 lines
target-alpha/translate.c +5
-0+tcg_gen_andi_i64 (tmp1, cpu_ir[rb],7 );tcg_gen_shli_i64 (tmp1, tmp1,3 ); + +tcg_gen_mov_i64 (cpu_ir[rc], cpu_ir[ra]); +tcg_gen_brcondi_i64 (TCG_COND_EQ , tmp1,0 , l1); + tmp2 =tcg_const_i64 (64 );tcg_gen_sub_i64 (tmp1, tmp2, tmp1);tcg_temp_free (tmp2);
Given that a test costs a lot (partly due to the fact temp local variable must be used), I do wonder if doing a AND here wouldn't be better:
tcg_gen_andi_i64(tmp1, tmp1, 0x3f);
Quoted 19 lines
target-alpha/translate.c +4
-1tcg_gen_shl_i64 (cpu_ir[rc], cpu_ir[ra], tmp1); + +gen_set_label (l1); +tcg_temp_free (tmp1); }if (tcg_gen_ext_i64)@@ -1320,7 +1330,7 @@ static always_inline int translate_one (DisasContext *ctx, uint32_t insn)break ;case 0x6A :/* EXTLH */ -gen_ext_h (&tcg_gen_ext16u_i64, ra, rb, rc, islit, lit); +gen_ext_h (&tcg_gen_ext32u_i64, ra, rb, rc, islit, lit);break ;case 0x72 :/* MSKQH */
Signature · 2 lines
Aurelien Jarno GPG: 1024D/F1BCDB73 [email protected] http://www.aurel32.net
Quoted 23 lines · Aurelien Jarno
Quoted 6 lines · Vince Weaver
} else {target-alpha/translate.c +3
-1+int l1;TCGv tmp1, tmp2; - tmp1 =tcg_temp_new (); + tmp1 =tcg_temp_local_new (); + l1 =gen_new_label ();target-alpha/translate.c static always_inline void gen_ext_h(void (*tcg_gen_ext_i64)(TCGv t0, TCGv t1), · line 542
Quoted 10 lines
target-alpha/translate.c +5
-0+tcg_gen_andi_i64 (tmp1, cpu_ir[rb],7 );tcg_gen_shli_i64 (tmp1, tmp1,3 ); + +tcg_gen_mov_i64 (cpu_ir[rc], cpu_ir[ra]); +tcg_gen_brcondi_i64 (TCG_COND_EQ , tmp1,0 , l1); + tmp2 =tcg_const_i64 (64 );tcg_gen_sub_i64 (tmp1, tmp2, tmp1);tcg_temp_free (tmp2);Given that a test costs a lot (partly due to the fact temp local variable must be used), I do wonder if doing a AND here wouldn't be better:
Quoted 1 line
tcg_gen_andi_i64(tmp1, tmp1, 0x3f);
I'm not sure I follow.
The code is attempting the following:
tmp1=rb&0x7;
tmp1=temp1<<3;
if (tmp1!=0) {
tmp1=64-tmp1;
rc=ra<<tmp1;
}
else {
rc=ra;
}The problem with the original code is that in the case of tmp1 being 0, the shift left by 64 would result in 0, instead of the identity.
I tried to avoid the jump but couldn't. Am I missing something?
Quoted 4 lines
target-alpha/translate.c static always_inline void gen_ext_h(void (*tcg_gen_ext_i64)(TCGv t0, TCGv t1), · line 546
Quoted 4 lines
tcg_gen_shl_i64(cpu_ir[rc], cpu_ir[ra], tmp1);target-alpha/translate.c +3
-0+ +gen_set_label (l1); +
Vince
Quoted 42 lines · Vince Weaver
Quoted 23 lines · Aurelien Jarno
Quoted 6 lines · Vince Weaver
} else {target-alpha/translate.c +3
-1+int l1;TCGv tmp1, tmp2; - tmp1 =tcg_temp_new (); + tmp1 =tcg_temp_local_new (); + l1 =gen_new_label ();target-alpha/translate.c static always_inline void gen_ext_h(void (*tcg_gen_ext_i64)(TCGv t0, TCGv t1), · line 542
Quoted 10 lines
target-alpha/translate.c +5
-0+tcg_gen_andi_i64 (tmp1, cpu_ir[rb],7 );tcg_gen_shli_i64 (tmp1, tmp1,3 ); + +tcg_gen_mov_i64 (cpu_ir[rc], cpu_ir[ra]); +tcg_gen_brcondi_i64 (TCG_COND_EQ , tmp1,0 , l1); + tmp2 =tcg_const_i64 (64 );tcg_gen_sub_i64 (tmp1, tmp2, tmp1);tcg_temp_free (tmp2);Given that a test costs a lot (partly due to the fact temp local variable must be used), I do wonder if doing a AND here wouldn't be better:
Quoted 1 line
tcg_gen_andi_i64(tmp1, tmp1, 0x3f);I'm not sure I follow.
The code is attempting the following:
tmp1=rb&0x7; tmp1=temp1<<3; if (tmp1!=0) { tmp1=64-tmp1; rc=ra<<tmp1; } else { rc=ra; }
Quoted 5 lines
The problem with the original code is that in the case of tmp1 being 0, the shift left by 64 would result in 0, instead of the identity.
I tried to avoid the jump but couldn't. Am I missing something?
I mean the following code:
tmp1=rb&0x7;
tmp1=temp1<<3;
tmp1=64-tmp1;
tmp1=tmp1 & 0x3f;
rc=ra<<tmp1;In case tmp1 = 0, it becomes 64, and then 0 again after the and, so rc=ra<<0.
Signature · 2 lines
Aurelien Jarno GPG: 1024D/F1BCDB73 [email protected] http://www.aurel32.net
Quoted 2 lines · Aurelien Jarno
In case tmp1 = 0, it becomes 64, and then 0 again after the and, so rc=ra<<0.
Ah, I see. I completely missed that optimization.
How does this updated patch look? I removed one of the TCGv variables too. Does that help performance? What would be nice is a tcg subtract-from instruction, which I know some architectures have. Maybe tcg does have it and I should look harder.
target-alpha/translate.c +7 -5
@@ -524,14 +524,16 @@ static inline void gen_ext_h(void(*tcg_gen_ext_i64)(TCGv t0, TCGv t1),else tcg_gen_mov_i64 (cpu_ir[rc], cpu_ir[ra]); }else { -TCGv tmp1, tmp2; +TCGv tmp1; tmp1 =tcg_temp_new (); +tcg_gen_andi_i64 (tmp1, cpu_ir[rb],7 );tcg_gen_shli_i64 (tmp1, tmp1,3 ); - tmp2 =tcg_const_i64 (64 ); -tcg_gen_sub_i64 (tmp1, tmp2, tmp1); -tcg_temp_free (tmp2); +tcg_gen_andi_i64 (tmp1, tmp1,0x3f ); +tcg_gen_neg_i64 (tmp1, tmp1); +tcg_gen_addi_i64 (tmp1, tmp1,64 );tcg_gen_shl_i64 (cpu_ir[rc], cpu_ir[ra], tmp1); +tcg_temp_free (tmp1); }if (tcg_gen_ext_i64)
@@ -1316,7 +1318,7 @@ static inline int translate_one(DisasContext *ctx, uint32_t insn)break ;case 0x6A :/* EXTLH */ -gen_ext_h (&tcg_gen_ext16u_i64, ra, rb, rc, islit, lit); +gen_ext_h (&tcg_gen_ext32u_i64, ra, rb, rc, islit, lit);break ;case 0x72 :/* MSKQH */
Quoted 4 lines · Vince Weaver
What would be nice is a tcg
subtract-from instruction, which I know some architectures have. Maybe tcg does have it and I should look harder.
There is tcg_gen_subfi.
Laurent
target-alpha/translate.c static inline void gen_ext_h(void(*tcg_gen_ext_i64)(TCGv t0, TCGv t1), · line 534
Quoted 8 lines · Vince Weaver
tcg_gen_andi_i64(tmp1, cpu_ir[rb], 7); tcg_gen_shli_i64(tmp1, tmp1, 3);target-alpha/translate.c +3
-3- tmp2 =tcg_const_i64 (64 ); -tcg_gen_sub_i64 (tmp1, tmp2, tmp1); -tcg_temp_free (tmp2); +tcg_gen_andi_i64 (tmp1, tmp1,0x3f ); +tcg_gen_neg_i64 (tmp1, tmp1); +tcg_gen_addi_i64 (tmp1, tmp1,64 );
This wastes an operation. If you switch andi and neg you don't need to add 64.
Andreas.
Signature · 3 lines
Andreas Schwab, [email protected] GPG Key fingerprint = 58CA 54C7 6D53 942B 1756 01D3 44D5 214B 8276 4ED5 "And now for something completely different."
Quoted 9 lines · Vince Weaver
Quoted 2 lines · Aurelien Jarno
In case tmp1 = 0, it becomes 64, and then 0 again after the and, so rc=ra<<0.
Ah, I see. I completely missed that optimization.
How does this updated patch look? I removed one of the TCGv variables too. Does that help performance? What would be nice is a tcg
Yes it looks ok. Removing one normal TCGv variable doesn't really help. What helps is removing a tcg temp_local variable or a branch.
Quoted 2 lines
subtract-from instruction, which I know some architectures have. Maybe tcg does have it and I should look harder.
You can use tcg_gen_subfi_i64 (tcg result, immediate, tcg arg).
target-alpha/translate.c static inline void gen_ext_h(void(*tcg_gen_ext_i64)(TCGv t0, TCGv t1), · line 527
Quoted 9 lines
target-alpha/translate.c +0
-1@@ -524,14 +524,16 @@ static inline void gen_ext_h(void(*tcg_gen_ext_i64)(TCGv t0, TCGv t1),else tcg_gen_mov_i64 (cpu_ir[rc], cpu_ir[ra]); }else { -TCGv tmp1, tmp2;
ap> + TCGv tmp1;
Quoted 25 lines
target-alpha/translate.c +6
-4tmp1 =tcg_temp_new (); +tcg_gen_andi_i64 (tmp1, cpu_ir[rb],7 );tcg_gen_shli_i64 (tmp1, tmp1,3 ); - tmp2 =tcg_const_i64 (64 ); -tcg_gen_sub_i64 (tmp1, tmp2, tmp1); -tcg_temp_free (tmp2); +tcg_gen_andi_i64 (tmp1, tmp1,0x3f ); +tcg_gen_neg_i64 (tmp1, tmp1); +tcg_gen_addi_i64 (tmp1, tmp1,64 );tcg_gen_shl_i64 (cpu_ir[rc], cpu_ir[ra], tmp1); +tcg_temp_free (tmp1); }if (tcg_gen_ext_i64)@@ -1316,7 +1318,7 @@ static inline int translate_one(DisasContext *ctx, uint32_t insn)break ;case 0x6A :/* EXTLH */ -gen_ext_h (&tcg_gen_ext16u_i64, ra, rb, rc, islit, lit); +gen_ext_h (&tcg_gen_ext32u_i64, ra, rb, rc, islit, lit);break ;case 0x72 :/* MSKQH */
Signature · 2 lines
Aurelien Jarno GPG: 1024D/F1BCDB73 [email protected] http://www.aurel32.net
Quoted 13 lines · Vince Weaver
The code is attempting the following:
tmp1=rb&0x7; tmp1=temp1<<3; if (tmp1!=0) { tmp1=64-tmp1; rc=ra<<tmp1; } else { rc=ra; }
Quoted 4 lines
The problem with the original code is that in the case of tmp1 being 0, the shift left by 64 would result in 0, instead of the identity.
I tried to avoid the jump but couldn't. Am I missing something?
Instead of tmp1 = 64 - tmp1 use tmp1 = -tmp1 & 0x3f.
Andreas.
Signature · 3 lines
Andreas Schwab, [email protected] GPG Key fingerprint = 58CA 54C7 6D53 942B 1756 01D3 44D5 214B 8276 4ED5 "And now for something completely different."