s390x/tcg: Vector instruction fixes
v2 cover s390x/tcg: Vector instruction fixes
Ivan reported that a simple
$ go get -v -d github.com/FactomProject/factom
Will result in errors when vector instructions are in use. Turns out golang makes excessive use of vector instructions, e.g., for crypto, hashes but also basic math.
I tracked the involved vector instructions and started writing more tests for them (will upstream them once they are in a better shape). Turns out there are quite some issues remaining. golang uses instructions not yet used by the kernel or by glibc.
With these patches, "go get" works again. It wouldn't surprise me if there are more BUGs in the vector instructions. Will continue writing more tests.
v1 -> v2:
- "s390x/tcg: Fix VECTOR SUBTRACT COMPUTE BORROW INDICATION" -- Fix by inverting the result
- "s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW INDICATION"
-- Reuse gen_ac2_i64()
- "s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW COMPUTE BORROW INDICATION" -- Reuse gen_accc2_i64()
- Ran against more tests
David Hildenbrand (6):
s390x/tcg: Fix VECTOR MULTIPLY LOGICAL ODD
s390x/tcg: Fix VECTOR MULTIPLY AND ADD *
s390x/tcg: Fix VECTOR SHIFT RIGHT ARITHMETIC BY BYTE s390x/tcg: Fix VECTOR SUBTRACT COMPUTE BORROW INDICATION s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW INDICATION s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW COMPUTE BORROW INDICATION
target/s390x/translate_vx.inc.c | 30 ++++++++++++------------------ target/s390x/vec_int_helper.c | 18 ++++++++++-------- 2 files changed, 22 insertions(+), 26 deletions(-)
Signature · 1 line
2.21.0
PATCH v2 3/6 s390x/tcg: Fix VECTOR SHIFT RIGHT ARITHMETIC BY BYTE
We forgot to propagate the highest bit accross the high doubleword in two cases (shift >=64).
--- target/s390x/vec_int_helper.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
target/s390x/vec_int_helper.c +4 -2
@@ -70,15 +70,17 @@ static void s390_vec_sar(S390Vector *d, const S390Vector *a, uint64_t count) d->doubleword[0 ] = a->doubleword[0 ]; d->doubleword[1 ] = a->doubleword[1 ]; }else if (count ==64 ) { + tmp = (int64_t )a->doubleword[0 ] >>63 ; d->doubleword[1 ] = a->doubleword[0 ]; - d->doubleword[0 ] =0 ; + d->doubleword[0 ] = tmp; }else if (count <64 ) { tmp = a->doubleword[1 ] >> count; d->doubleword[1 ] =deposit64 (tmp,64 - count, count, a->doubleword[0 ]); d->doubleword[0 ] = (int64_t )a->doubleword[0 ] >> count; }else { + tmp = (int64_t )a->doubleword[0 ] >>63 ; d->doubleword[1 ] = (int64_t )a->doubleword[0 ] >> (count -64 ); - d->doubleword[0 ] =0 ; + d->doubleword[0 ] = tmp; } }
Signature · 1 line
2.21.0
PATCH v2 4/6 s390x/tcg: Fix VECTOR SUBTRACT COMPUTE BORROW INDICATION
In reply to David HildenbrandLooks like my idea of what a "borrow" is was wrong. The PoP says:
"If the resulting subtraction results in a carry out of bit zero, a value of one is placed in the corresponding element of the first operand; otherwise, a value of zero is placed in the corresponding element"
As clarified by Richard, all we have to do is invert the result.
--- target/s390x/translate_vx.inc.c | 7 ++++--- target/s390x/vec_int_helper.c | 2 +- 2 files changed, 5 insertions(+), 4 deletions(-)
target/s390x/translate_vx.inc.c +4 -3
@@ -2132,12 +2132,12 @@ static DisasJumpType op_vs(DisasContext *s, DisasOps *o)static void gen_scbi_i32 (TCGv_i32 d,TCGv_i32 a,TCGv_i32 b) { -tcg_gen_setcond_i32 (TCG_COND_LTU , d, a, b); +tcg_gen_setcond_i32 (TCG_COND_GEU , d, a, b); }static void gen_scbi_i64 (TCGv_i64 d,TCGv_i64 a,TCGv_i64 b) { -tcg_gen_setcond_i64 (TCG_COND_LTU , d, a, b); +tcg_gen_setcond_i64 (TCG_COND_GEU , d, a, b); } static void gen_scbi2_i64(TCGv_i64 dl, TCGv_i64 dh, TCGv_i64 al,
@@ -2151,7 +2151,8 @@ static void gen_scbi2_i64(TCGv_i64 dl, TCGv_i64 dh, TCGv_i64 al,tcg_gen_andi_i64 (th, th,1 );tcg_gen_sub2_i64 (tl, th, ah, zero, th, zero);tcg_gen_sub2_i64 (tl, th, tl, th, bh, zero); -tcg_gen_andi_i64 (dl, th,1 ); +/* "invert" the result: -1 -> 0; 0 -> 1 */ +tcg_gen_addi_i64 (dl, th,1 );tcg_gen_mov_i64 (dh, zero);tcg_temp_free_i64 (th);
target/s390x/vec_int_helper.c +1 -1
@@ -593,7 +593,7 @@ void HELPER(gvec_vscbi##BITS)(void *v1, const void *v2, const void *v3, \ const uint##BITS##_t a = s390_vec_read_element##BITS(v2, i); \ const uint##BITS##_t b = s390_vec_read_element##BITS(v3, i); \ \ - s390_vec_write_element##BITS(v1, i, a < b); \ + s390_vec_write_element##BITS(v1, i, a >= b); \ } \ }DEF_VSCBI (8 )
Signature · 1 line
2.21.0
PATCH v2 1/6 s390x/tcg: Fix VECTOR MULTIPLY LOGICAL ODD
In reply to David HildenbrandWe have to read from odd offsets.
--- target/s390x/vec_int_helper.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
target/s390x/vec_int_helper.c +1 -1
@@ -488,7 +488,7 @@ void HELPER(gvec_vmlo##BITS)(void *v1, const void *v2, const void *v3, \ { \int i, j; \ \ -for (i =0 , j =0 ; i < (128 /TBITS ); i++, j +=2 ) { \ +for (i =0 , j =1 ; i < (128 /TBITS ); i++, j +=2 ) { \ const uint##TBITS##_t a = s390_vec_read_element##BITS(v2, j); \ const uint##TBITS##_t b = s390_vec_read_element##BITS(v3, j); \ \
Signature · 1 line
2.21.0
PATCH v2 2/6 s390x/tcg: Fix VECTOR MULTIPLY AND ADD *
In reply to David HildenbrandWe missed that we always read a "double-wide even-odd element pair of the fourth operand". Fix it in all four variants.
--- target/s390x/vec_int_helper.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
target/s390x/vec_int_helper.c +4 -4
@@ -336,7 +336,7 @@ void HELPER(gvec_vmae##BITS)(void *v1, const void *v2, const void *v3, \for (i =0 , j =0 ; i < (128 /TBITS ); i++, j +=2 ) { \ int##TBITS##_t a = (int##BITS##_t)s390_vec_read_element##BITS(v2, j); \ int##TBITS##_t b = (int##BITS##_t)s390_vec_read_element##BITS(v3, j); \ - int##TBITS##_t c = (int##BITS##_t)s390_vec_read_element##BITS(v4, j); \ + int##TBITS##_t c = s390_vec_read_element##TBITS(v4, i); \ \ s390_vec_write_element##TBITS(v1, i, a * b + c); \ } \
@@ -354,7 +354,7 @@ void HELPER(gvec_vmale##BITS)(void *v1, const void *v2, const void *v3, \for (i =0 , j =0 ; i < (128 /TBITS ); i++, j +=2 ) { \ uint##TBITS##_t a = s390_vec_read_element##BITS(v2, j); \ uint##TBITS##_t b = s390_vec_read_element##BITS(v3, j); \ - uint##TBITS##_t c = s390_vec_read_element##BITS(v4, j); \ + uint##TBITS##_t c = s390_vec_read_element##TBITS(v4, i); \ \ s390_vec_write_element##TBITS(v1, i, a * b + c); \ } \
@@ -372,7 +372,7 @@ void HELPER(gvec_vmao##BITS)(void *v1, const void *v2, const void *v3, \for (i =0 , j =1 ; i < (128 /TBITS ); i++, j +=2 ) { \ int##TBITS##_t a = (int##BITS##_t)s390_vec_read_element##BITS(v2, j); \ int##TBITS##_t b = (int##BITS##_t)s390_vec_read_element##BITS(v3, j); \ - int##TBITS##_t c = (int##BITS##_t)s390_vec_read_element##BITS(v4, j); \ + int##TBITS##_t c = s390_vec_read_element##TBITS(v4, i); \ \ s390_vec_write_element##TBITS(v1, i, a * b + c); \ } \
@@ -390,7 +390,7 @@ void HELPER(gvec_vmalo##BITS)(void *v1, const void *v2, const void *v3, \for (i =0 , j =1 ; i < (128 /TBITS ); i++, j +=2 ) { \ uint##TBITS##_t a = s390_vec_read_element##BITS(v2, j); \ uint##TBITS##_t b = s390_vec_read_element##BITS(v3, j); \ - uint##TBITS##_t c = s390_vec_read_element##BITS(v4, j); \ + uint##TBITS##_t c = s390_vec_read_element##TBITS(v4, i); \ \ s390_vec_write_element##TBITS(v1, i, a * b + c); \ } \
Signature · 1 line
2.21.0
PATCH v2 5/6 s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW INDICATION
In reply to David HildenbrandTesting this, there seems to be something messed up. We are dealing with unsigned numbers. "Each operand is treated as an unsigned binary integer." Let's just implement as written in the PoP:
"A subtraction is performed by adding the contents of
the second operand with the bitwise complement of the third operand along with a borrow indication from the rightmost bit position of the fourth operand and the result is placed in the first operand."
We can reuse gen_ac2_i64().
--- target/s390x/translate_vx.inc.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-)
target/s390x/translate_vx.inc.c +5 -5
@@ -2187,13 +2187,13 @@ static void gen_sbi2_i64(TCGv_i64 dl, TCGv_i64 dh, TCGv_i64 al, TCGv_i64 ah, TCGv_i64 bl, TCGv_i64 bh, TCGv_i64 cl, TCGv_i64 ch) { TCGv_i64 tl = tcg_temp_new_i64(); -TCGv_i64 zero =tcg_const_i64 (0 ); +TCGv_i64 th =tcg_temp_new_i64 (); -tcg_gen_andi_i64 (tl, cl,1 ); -tcg_gen_sub2_i64 (dl, dh, al, ah, bl, bh); -tcg_gen_sub2_i64 (dl, dh, dl, dh, tl, zero); +tcg_gen_not_i64 (tl, bl); +tcg_gen_not_i64 (th, bh); +gen_ac2_i64 (dl, dh, al, ah, tl, th, cl, ch);tcg_temp_free_i64 (tl); -tcg_temp_free_i64 (zero); +tcg_temp_free_i64 (th); }static DisasJumpType op_vsbi (DisasContext *s,DisasOps *o)
Signature · 1 line
2.21.0
PATCH v2 6/6 s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW COMPUTE BORROW INDICATION
In reply to David HildenbrandThe numbers are unsigned, the computation is wrong. "Each operand is treated as an unsigned binary integer".
Let's implement as given in the PoP:
"A subtraction is performed by adding the contents of the second operand
with the bitwise complement of the third operand along with a borrow indication from the rightmost bit of the fourth operand."
Reuse gen_accc2_i64().
--- target/s390x/translate_vx.inc.c | 13 +++---------- 1 file changed, 3 insertions(+), 10 deletions(-)
target/s390x/translate_vx.inc.c +3 -10
@@ -2214,20 +2214,13 @@ static void gen_sbcbi2_i64(TCGv_i64 dl, TCGv_i64 dh, TCGv_i64 al, TCGv_i64 ah, {TCGv_i64 th =tcg_temp_new_i64 ();TCGv_i64 tl =tcg_temp_new_i64 (); -TCGv_i64 zero =tcg_const_i64 (0 ); -tcg_gen_andi_i64 (tl, cl,1 ); -tcg_gen_sub2_i64 (tl, th, al, zero, tl, zero); -tcg_gen_sub2_i64 (tl, th, tl, th, bl, zero); -tcg_gen_andi_i64 (th, th,1 ); -tcg_gen_sub2_i64 (tl, th, ah, zero, th, zero); -tcg_gen_sub2_i64 (tl, th, tl, th, bh, zero); -tcg_gen_andi_i64 (dl, th,1 ); -tcg_gen_mov_i64 (dh, zero); +tcg_gen_not_i64 (tl, bl); +tcg_gen_not_i64 (th, bh); +gen_accc2_i64 (dl, dh, al, ah, tl, th, cl, ch);tcg_temp_free_i64 (tl);tcg_temp_free_i64 (th); -tcg_temp_free_i64 (zero); }static DisasJumpType op_vsbcbi (DisasContext *s,DisasOps *o)
Signature · 1 line
2.21.0
Re: [PATCH v2 4/6] s390x/tcg: Fix VECTOR SUBTRACT COMPUTE BORROW INDICATION
In reply to David HildenbrandQuoted 6 lines · David Hildenbrand
Looks like my idea of what a "borrow" is was wrong. The PoP says:
"If the resulting subtraction results in a carry out of bit zero, a value of one is placed in the corresponding element of the first operand; otherwise, a value of zero is placed in the corresponding element"
Quoted 8 lines
As clarified by Richard, all we have to do is invert the result.
Fixes 1ee2d7ba72f6 ("s390x/tcg: Implement VECTOR SUBTRACT COMPUTE BORROW INDICATION")Signed-off-by David Hildenbrand <[email protected]>--- target/s390x/translate_vx.inc.c | 7 ++++--- target/s390x/vec_int_helper.c | 2 +- 2 files changed, 5 insertions(+), 4 deletions(-)
r~
Re: [PATCH v2 5/6] s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW INDICATION
In reply to David HildenbrandQuoted 10 lines · David Hildenbrand
Testing this, there seems to be something messed up. We are dealing with unsigned numbers. "Each operand is treated as an unsigned binary integer." Let's just implement as written in the PoP:
"A subtraction is performed by adding the contents of
the second operand with the bitwise complement of the third operand along with a borrow indication from the rightmost bit position of the fourth operand and the result is placed in the first operand."
Quoted 7 lines
We can reuse
gen_ac2_i64().Fixes 48390a7c2716 ("s390x/tcg: Implement VECTOR SUBTRACT WITH BORROW INDICATION")Signed-off-by David Hildenbrand <[email protected]>--- target/s390x/translate_vx.inc.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-)
r~
Re: [PATCH v2 6/6] s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW COMPUTE BORROW INDICATION
In reply to David HildenbrandQuoted 8 lines · David Hildenbrand
The numbers are unsigned, the computation is wrong. "Each operand is treated as an unsigned binary integer".
Let's implement as given in the PoP:"A subtraction is performed by adding the contents of the second operand
with the bitwise complement of the third operand along with a borrow indication from the rightmost bit of the fourth operand."
Quoted 7 lines
Reuse
gen_accc2_i64().Fixes bc725e65152c ("s390x/tcg: Implement VECTOR SUBTRACT WITH BORROW COMPUTE BORROW INDICATION")Signed-off-by David Hildenbrand <[email protected]>--- target/s390x/translate_vx.inc.c | 13 +++---------- 1 file changed, 3 insertions(+), 10 deletions(-)
r~
Cornelia Huck
AppliedQuoted 34 lines · On Mon, 21 Oct 2019 10:57:09 +0200 David Hildenbrand <[email protected]> wrote:
Ivan reported that a simple
$ go get -v -d github.com/FactomProject/factomWill result in errors when vector instructions are in use. Turns out golang makes excessive use of vector instructions, e.g., for crypto, hashes but also basic math.
I tracked the involved vector instructions and started writing more tests for them (will upstream them once they are in a better shape). Turns out there are quite some issues remaining. golang uses instructions not yet used by the kernel or by glibc.
With these patches, "go get" works again. It wouldn't surprise me if there are more BUGs in the vector instructions. Will continue writing more tests.
Cc Ivan Warren <[email protected]>v1 -> v2:
- "s390x/tcg: Fix VECTOR SUBTRACT COMPUTE BORROW INDICATION" -- Fix by inverting the result
- "s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW INDICATION"
-- Reusegen_ac2_i64()
- "s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW COMPUTE BORROW INDICATION" -- Reusegen_accc2_i64()
- Ran against more testsDavid Hildenbrand (6):
s390x/tcg: Fix VECTOR MULTIPLY LOGICAL ODD
s390x/tcg: Fix VECTOR MULTIPLY AND ADD *
s390x/tcg: Fix VECTOR SHIFT RIGHT ARITHMETIC BY BYTE s390x/tcg: Fix VECTOR SUBTRACT COMPUTE BORROW INDICATION s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW INDICATION s390x/tcg: Fix VECTOR SUBTRACT WITH BORROW COMPUTE BORROW INDICATION
Quoted 4 lines
target/s390x/translate_vx.inc.c | 30 ++++++++++++------------------ target/s390x/vec_int_helper.c | 18 ++++++++++-------- 2 files changed, 22 insertions(+), 26 deletions(-)
Thanks, applied.