target/s390x: Fix shift instructions
v3 cover target/s390x: Fix shift instructions
v2: https://lists.nongnu.org/archive/html/qemu-devel/2022-01/msg02488.html
v2 -> v3: Unify CC_OP_SLA_32 and CC_OP_SLA_64.
Add underscores to test macro parameters.
Shift CC in test asm.
Add a second SLAG test.
Add tags to commit messages.
v1: https://lists.nongnu.org/archive/html/qemu-devel/2022-01/msg02035.html
v1 -> v2: Fix cc_calc_sla_32().
Fix cc_calc_sla_64().
Fix SLDA sign bit index.
Inline help_l2_shift().
Fix wout_r1_D32().
Add all shift instructions to the test.
Split the series.
Ilya Leoshkevich (5):
target/s390x: Fix SLDA sign bit index
target/s390x: Fix SRDA CC calculation
target/s390x: Fix cc_calc_sla_64() missing overflows target/s390x: Fix shifting 32-bit values for more than 31 bits tests/tcg/s390x: Test shift instructions
target/s390x/cpu-dump.c | 3 +- target/s390x/s390x-internal.h | 3 +- target/s390x/tcg/cc_helper.c | 39 +---- target/s390x/tcg/insn-data.def | 36 ++--- target/s390x/tcg/translate.c | 49 +++--- tests/tcg/s390x/Makefile.target | 1 + tests/tcg/s390x/shift.c | 270 ++++++++++++++++++++++++++++++++ 7 files changed, 318 insertions(+), 83 deletions(-) create mode 100644 tests/tcg/s390x/shift.c
Signature · 1 line
2.31.1
PATCH v3 1/5 target/s390x: Fix SLDA sign bit index
SLDA operates on 64-bit values, so its sign bit index should be 63, not 31.
--- target/s390x/tcg/insn-data.def | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
target/s390x/tcg/insn-data.def +1 -1
@@ -800,7 +800,7 @@ C(0xebde, SRLK, RSY_a, DO, r3_32u, sh32, new, r1_32, srl, 0) C(0xeb0c, SRLG, RSY_a, Z, r3_o, sh64, r1, 0, srl, 0) /* SHIFT LEFT DOUBLE */ - D(0x8f00, SLDA, RS_a, Z, r1_D32, sh64, new, r1_D32, sla, 0, 31) + D(0x8f00, SLDA, RS_a, Z, r1_D32, sh64, new, r1_D32, sla, 0, 63) /* SHIFT LEFT DOUBLE LOGICAL */ C(0x8d00, SLDL, RS_a, Z, r1_D32, sh64, new, r1_D32, sll, 0) /* SHIFT RIGHT DOUBLE */
Signature · 1 line
2.31.1
PATCH v3 2/5 target/s390x: Fix SRDA CC calculation
In reply to Ilya LeoshkevichSRDA uses r1_D32 for binding the first operand and s64 for setting CC. cout_s64() relies on o->out being the shift result, however, wout_r1_D32() clobbers it.
Fix by using a temporary.
--- target/s390x/tcg/translate.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
target/s390x/tcg/translate.c +3 -2
@@ -5420,9 +5420,10 @@ static void wout_r1_P32(DisasContext *s, DisasOps *o)static void wout_r1_D32 (DisasContext *s,DisasOps *o) {int r1 =get_field (s, r1); +TCGv_i64 t =tcg_temp_new_i64 ();store_reg32_i64 (r1 +1 , o->out); -tcg_gen_shri_i64 (o->out, o->out,32 ); -store_reg32_i64 (r1, o->out); +tcg_gen_shri_i64 (t, o->out,32 ); +store_reg32_i64 (r1, t); }#define SPEC_wout_r1_D32 SPEC_r1_even
Signature · 1 line
2.31.1
Re: [PATCH v3 2/5] target/s390x: Fix SRDA CC calculation
Quoted 18 lines · Ilya Leoshkevich
SRDA uses
r1_D32for binding the first operand and s64 for setting CC.cout_s64()relies ono->outbeing the shift result, however,wout_r1_D32()clobbers it.Fix by using a temporary.
Fixes a79ba3398a0a ("target-s390: Convert SHIFT DOUBLE")Signed-off-by Ilya Leoshkevich <[email protected]>Reviewed-by David Hildenbrand <[email protected]>--- target/s390x/tcg/translate.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)target/s390x/tcg/translate.c +0
-0
target/s390x/tcg/translate.c static void wout_r1_P32(DisasContext *s, DisasOps *o) · line 5426
Quoted 9 lines
target/s390x/tcg/translate.c +3
-2@@ -5420,9 +5420,10 @@ static void wout_r1_P32(DisasContext *s, DisasOps *o)static void wout_r1_D32 (DisasContext *s,DisasOps *o) {int r1 =get_field (s, r1); +TCGv_i64 t =tcg_temp_new_i64 ();store_reg32_i64 (r1 +1 , o->out); -tcg_gen_shri_i64 (o->out, o->out,32 ); -store_reg32_i64 (r1, o->out); +tcg_gen_shri_i64 (t, o->out,32 ); +store_reg32_i64 (r1, t);
I think you're missing a tcg_temp_free_i64().
Signature · 3 lines
Thanks, David / dhildenb
PATCH v3 3/5 target/s390x: Fix cc_calc_sla_64() missing overflows
In reply to Ilya LeoshkevichAn overflow occurs for SLAG when at least one shifted bit is not equal to sign bit. Therefore, we need to check that shift + 1 bits are neither all 0s nor all 1s. The current code checks only shift bits, missing some overflows.
--- target/s390x/tcg/cc_helper.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
target/s390x/tcg/cc_helper.c +2 -1
@@ -297,7 +297,8 @@ static uint32_t cc_calc_sla_32(uint32_t src, int shift)static uint32_t cc_calc_sla_64 (uint64_t src,int shift) { -uint64_t mask = ((1ULL << shift) -1ULL ) << (64 - shift); +/* Do not use (1ULL << (shift + 1)): it triggers UB when shift is 63. */ +uint64_t mask = ((((1ULL << shift) -1 ) <<1 ) +1 ) << (64 - (shift +1 ));uint64_t sign =1ULL <<63 ;uint64_t match;int64_t r;
Signature · 1 line
2.31.1
PATCH v3 4/5 target/s390x: Fix shifting 32-bit values for more than 31 bits
In reply to Ilya LeoshkevichAccording to PoP, both 32- and 64-bit shifts use lowest 6 address bits. The current code special-cases 32-bit shifts to use only 5 bits, which is not correct. For example, shifting by 32 bits currently preserves the initial value, however, it's supposed zero it out instead.
Fix by merging sh32 and sh64 and adapting CC calculation to shift values greater than 31.
--- target/s390x/cpu-dump.c | 3 +-- target/s390x/s390x-internal.h | 3 +-- target/s390x/tcg/cc_helper.c | 36 +++------------------------- target/s390x/tcg/insn-data.def | 36 ++++++++++++++-------------- target/s390x/tcg/translate.c | 44 +++++++++++++++------------------- 5 files changed, 42 insertions(+), 80 deletions(-)
target/s390x/cpu-dump.c +1 -2
@@ -121,8 +121,7 @@ const char *cc_name(enum cc_op cc_op) [CC_OP_NZ_F64] = "CC_OP_NZ_F64", [CC_OP_NZ_F128] = "CC_OP_NZ_F128", [CC_OP_ICM] = "CC_OP_ICM", - [CC_OP_SLA_32] = "CC_OP_SLA_32", - [CC_OP_SLA_64] = "CC_OP_SLA_64", + [CC_OP_SLA] = "CC_OP_SLA", [CC_OP_FLOGR] = "CC_OP_FLOGR", [CC_OP_LCBB] = "CC_OP_LCBB", [CC_OP_VC] = "CC_OP_VC",
target/s390x/s390x-internal.h +1 -2
@@ -193,8 +193,7 @@ enum cc_op {CC_OP_NZ_F128 ,/* FP dst != 0 (128bit) */ CC_OP_ICM ,/* insert characters under mask */ -CC_OP_SLA_32 ,/* Calculate shift left signed (32bit) */ -CC_OP_SLA_64 ,/* Calculate shift left signed (64bit) */ +CC_OP_SLA ,/* Calculate shift left signed */ CC_OP_FLOGR ,/* find leftmost one */ CC_OP_LCBB ,/* load count to block boundary */ CC_OP_VC, /* vector compare result */
target/s390x/tcg/cc_helper.c +3 -33
@@ -268,34 +268,7 @@ static uint32_t cc_calc_icm(uint64_t mask, uint64_t val) } } -static uint32_t cc_calc_sla_32 (uint32_t src,int shift) -{ -uint32_t mask = ((1U << shift) -1U ) << (32 - shift); -uint32_t sign =1U <<31 ; -uint32_t match; -int32_t r; - -/* Check if the sign bit stays the same. */ -if (src & sign) { - match = mask; - }else { - match =0 ; - } -if ((src & mask) != match) { -/* Overflow. */ -return 3 ; - } - - r = ((src << shift) & ~sign) | (src & sign); -if (r ==0 ) { -return 0 ; - }else if (r <0 ) { -return 1 ; - } -return 2 ; -} - -static uint32_t cc_calc_sla_64 (uint64_t src,int shift) +static uint32_t cc_calc_sla (uint64_t src,int shift) {/* Do not use (1ULL << (shift + 1)): it triggers UB when shift is 63. */ uint64_t mask = ((((1ULL << shift) -1 ) <<1 ) +1 ) << (64 - (shift +1 ));
@@ -460,11 +433,8 @@ static uint32_t do_calc_cc(CPUS390XState *env, uint32_t cc_op,case CC_OP_ICM : r =cc_calc_icm (src, dst);break ; -case CC_OP_SLA_32 : - r =cc_calc_sla_32 (src, dst); -break ; -case CC_OP_SLA_64 : - r =cc_calc_sla_64 (src, dst); +case CC_OP_SLA : + r =cc_calc_sla (src, dst);break ;case CC_OP_FLOGR : r =cc_calc_flogr (dst);
target/s390x/tcg/insn-data.def +18 -18
@@ -747,8 +747,8 @@ C(0xb9e1, POPCNT, RRE, PC, 0, r2_o, r1, 0, popcnt, nz64) /* ROTATE LEFT SINGLE LOGICAL */ - C(0xeb1d, RLL, RSY_a, Z, r3_o, sh32, new, r1_32, rll32, 0) - C(0xeb1c, RLLG, RSY_a, Z, r3_o, sh64, r1, 0, rll64, 0) + C(0xeb1d, RLL, RSY_a, Z, r3_o, sh, new, r1_32, rll32, 0) + C(0xeb1c, RLLG, RSY_a, Z, r3_o, sh, r1, 0, rll64, 0) /* ROTATE THEN INSERT SELECTED BITS */ C(0xec55, RISBG, RIE_f, GIE, 0, r2, r1, 0, risbg, s64)
@@ -784,29 +784,29 @@ C(0x0400, SPM, RR_a, Z, r1, 0, 0, 0, spm, 0) /* SHIFT LEFT SINGLE */ - D(0x8b00, SLA, RS_a, Z, r1, sh32, new, r1_32, sla, 0, 31) - D(0xebdd, SLAK, RSY_a, DO, r3, sh32, new, r1_32, sla, 0, 31) - D(0xeb0b, SLAG, RSY_a, Z, r3, sh64, r1, 0, sla, 0, 63) + D(0x8b00, SLA, RS_a, Z, r1, sh, new, r1_32, sla, 0, 31) + D(0xebdd, SLAK, RSY_a, DO, r3, sh, new, r1_32, sla, 0, 31) + D(0xeb0b, SLAG, RSY_a, Z, r3, sh, r1, 0, sla, 0, 63) /* SHIFT LEFT SINGLE LOGICAL */ - C(0x8900, SLL, RS_a, Z, r1_o, sh32, new, r1_32, sll, 0) - C(0xebdf, SLLK, RSY_a, DO, r3_o, sh32, new, r1_32, sll, 0) - C(0xeb0d, SLLG, RSY_a, Z, r3_o, sh64, r1, 0, sll, 0) + C(0x8900, SLL, RS_a, Z, r1_o, sh, new, r1_32, sll, 0) + C(0xebdf, SLLK, RSY_a, DO, r3_o, sh, new, r1_32, sll, 0) + C(0xeb0d, SLLG, RSY_a, Z, r3_o, sh, r1, 0, sll, 0) /* SHIFT RIGHT SINGLE */ - C(0x8a00, SRA, RS_a, Z, r1_32s, sh32, new, r1_32, sra, s32) - C(0xebdc, SRAK, RSY_a, DO, r3_32s, sh32, new, r1_32, sra, s32) - C(0xeb0a, SRAG, RSY_a, Z, r3_o, sh64, r1, 0, sra, s64) + C(0x8a00, SRA, RS_a, Z, r1_32s, sh, new, r1_32, sra, s32) + C(0xebdc, SRAK, RSY_a, DO, r3_32s, sh, new, r1_32, sra, s32) + C(0xeb0a, SRAG, RSY_a, Z, r3_o, sh, r1, 0, sra, s64) /* SHIFT RIGHT SINGLE LOGICAL */ - C(0x8800, SRL, RS_a, Z, r1_32u, sh32, new, r1_32, srl, 0) - C(0xebde, SRLK, RSY_a, DO, r3_32u, sh32, new, r1_32, srl, 0) - C(0xeb0c, SRLG, RSY_a, Z, r3_o, sh64, r1, 0, srl, 0) + C(0x8800, SRL, RS_a, Z, r1_32u, sh, new, r1_32, srl, 0) + C(0xebde, SRLK, RSY_a, DO, r3_32u, sh, new, r1_32, srl, 0) + C(0xeb0c, SRLG, RSY_a, Z, r3_o, sh, r1, 0, srl, 0) /* SHIFT LEFT DOUBLE */ - D(0x8f00, SLDA, RS_a, Z, r1_D32, sh64, new, r1_D32, sla, 0, 63) + D(0x8f00, SLDA, RS_a, Z, r1_D32, sh, new, r1_D32, sla, 0, 63) /* SHIFT LEFT DOUBLE LOGICAL */ - C(0x8d00, SLDL, RS_a, Z, r1_D32, sh64, new, r1_D32, sll, 0) + C(0x8d00, SLDL, RS_a, Z, r1_D32, sh, new, r1_D32, sll, 0) /* SHIFT RIGHT DOUBLE */ - C(0x8e00, SRDA, RS_a, Z, r1_D32, sh64, new, r1_D32, sra, s64) + C(0x8e00, SRDA, RS_a, Z, r1_D32, sh, new, r1_D32, sra, s64) /* SHIFT RIGHT DOUBLE LOGICAL */ - C(0x8c00, SRDL, RS_a, Z, r1_D32, sh64, new, r1_D32, srl, 0) + C(0x8c00, SRDL, RS_a, Z, r1_D32, sh, new, r1_D32, srl, 0) /* SQUARE ROOT */ F(0xb314, SQEBR, RRE, Z, 0, e2, new, e1, sqeb, 0, IF_BFP)
target/s390x/tcg/translate.c +19 -25
72 lines not on this page · raw message
Signature · 1 line
2.31.1
Re: [PATCH v3 4/5] target/s390x: Fix shifting 32-bit values for more than 31 bits
Quoted 8 lines
static DisasJumpType
help_goto_direct(DisasContext *s, uint64_t dest){ if (dest == s->pc_tmp) {target/s390x/tcg/translate.c +0
-0@@ -4113,9 +4099,15 @@ static DisasJumpType op_soc(DisasContext *s, DisasOps *o)static DisasJumpType op_sla (DisasContext *s,DisasOps *o) {
target/s390x/tcg/translate.c static DisasJumpType op_soc(DisasContext *s, DisasOps *o) · line 4109
Quoted 10 lines
target/s390x/tcg/translate.c +7
-2+TCGv_i64 t;uint64_t sign =1ull << s->insn->data; -enum cc_op cco = s->insn->data ==31 ?CC_OP_SLA_32 :CC_OP_SLA_64 ; -gen_op_update2_cc_i64 (s, cco, o->in1, o->in2); +if (s->insn->data ==31 ) { + t =tcg_temp_new_i64 (); +tcg_gen_shli_i64 (t, o->in1,32 ); + }else { + t = o->in1; + }
Are you missing to free the temp in case you allocated one?
Signature · 3 lines
Thanks, David / dhildenb
PATCH v3 5/5 tests/tcg/s390x: Test shift instructions
In reply to Ilya LeoshkevichAdd a test for each shift instruction in order to to prevent regressions.
--- tests/tcg/s390x/Makefile.target | 1 + tests/tcg/s390x/shift.c | 270 ++++++++++++++++++++++++++++++++ 2 files changed, 271 insertions(+) create mode 100644 tests/tcg/s390x/shift.c
tests/tcg/s390x/Makefile.target +1 -0
@@ -9,6 +9,7 @@ TESTS+=exrl-trtr TESTS+=pack TESTS+=mvo TESTS+=mvc +TESTS+=shift TESTS+=trap TESTS+=signals-s390x
tests/tcg/s390x/shift.c new file +270 -0
270 lines not on this page · raw message
Signature · 1 line
2.31.1