Skip to contentSkip to conversation

target/s390x: vfmin/vfmax fixes

PATCH Superseded Ilya Leoshkevich 3 participants 12 messages Jul 12, 2022 – Jul 13, 2022

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover target/s390x: vfmin/vfmax fixes

Hi,

Uli has found an issue with finding maximum of different kinds of 0s; I wrote a test and found another one with finding maximum of different kinds of NaNs.

Patches 1 and 2 fix those issues, patch 3 adds a vfmin/vfmax test.

Best regards,
Ilya

Ilya Leoshkevich (3):
target/s390x: fix handling of zeroes in vfmin/vfmax target/s390x: fix NaN propagation rules tests/tcg/s390x: test signed vfmin/vfmax

 fpu/softfloat-specialize.c.inc    |   3 +-
 target/s390x/tcg/vec_fpu_helper.c |   4 +-
 tests/tcg/s390x/Makefile.target   |   7 +
 tests/tcg/s390x/vfminmax.c        | 426 ++++++++++++++++++++++++++++++
 4 files changed, 437 insertions(+), 3 deletions(-)
 create mode 100644 tests/tcg/s390x/vfminmax.c
Signature · 1 line
2.35.3
#1 lore raw

PATCH 1/3 target/s390x: fix handling of zeroes in vfmin/vfmax

vfmin_res() / vfmax_res() are trying to check whether a and b are both zeroes, but in reality they check that they are the same kind of zero. This causes incorrect results when comparing positive and negative zeroes.

Fixes da4807527f3b ("s390x/tcg: Implement VECTOR FP (MAXIMUM|MINIMUM)")
---
 target/s390x/tcg/vec_fpu_helper.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
target/s390x/tcg/vec_fpu_helper.c +2 -2
index 2a618a1093..75cf605b9f 100644
@@ -824,7 +824,7 @@ static S390MinMaxRes vfmin_res(uint16_t dcmask_a, uint16_t dcmask_b,
         default:
             g_assert_not_reached();
         }
-    } else if (unlikely(dcmask_a & dcmask_b & DCMASK_ZERO)) {
+    } else if (unlikely((dcmask_a & DCMASK_ZERO) && (dcmask_b & DCMASK_ZERO))) {
         switch (type) {
         case S390_MINMAX_TYPE_JAVA:
             return neg_a ? S390_MINMAX_RES_A : S390_MINMAX_RES_B;
@@ -874,7 +874,7 @@ static S390MinMaxRes vfmax_res(uint16_t dcmask_a, uint16_t dcmask_b,
         default:
             g_assert_not_reached();
         }
-    } else if (unlikely(dcmask_a & dcmask_b & DCMASK_ZERO)) {
+    } else if (unlikely((dcmask_a & DCMASK_ZERO) && (dcmask_b & DCMASK_ZERO))) {
         const bool neg_a = dcmask_a & DCMASK_NEGATIVE;
 
         switch (type) {
Signature · 1 line
2.35.3
#4 lore raw

PATCH 2/3 target/s390x: fix NaN propagation rules

In reply to Ilya Leoshkevich

s390x has the same NaN propagation rules as ARM, and not as x86.

---
 fpu/softfloat-specialize.c.inc | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fpu/softfloat-specialize.c.inc +2 -1
index 943e3301d2..a43ff5e02e 100644
@@ -390,7 +390,8 @@ bool float32_is_signaling_nan(float32 a_, float_status *status)
 static int pickNaN(FloatClass a_cls, FloatClass b_cls,
                    bool aIsLargerSignificand, float_status *status)
 {
-#if defined(TARGET_ARM) || defined(TARGET_MIPS) || defined(TARGET_HPPA)
+#if defined(TARGET_ARM) || defined(TARGET_MIPS) || defined(TARGET_HPPA) || \
+    defined(TARGET_S390X)
     /* ARM mandated NaN propagation rules (see FPProcessNaNs()), take
      * the first of:
      *  1. A if it is signaling
Signature · 1 line
2.35.3
#7 lore raw

PATCH 3/3 tests/tcg/s390x: test signed vfmin/vfmax

In reply to Ilya Leoshkevich

Add a test to prevent regressions. Try all floating point value sizes and all combinations of floating point value classes. Verify the results against PoP tables, which are represented as close to the original as possible - this produces a lot of checkpatch complaints, but it seems to be justified in this case.

---
 tests/tcg/s390x/Makefile.target |   7 +
 tests/tcg/s390x/vfminmax.c      | 426 ++++++++++++++++++++++++++++++++
 2 files changed, 433 insertions(+)
 create mode 100644 tests/tcg/s390x/vfminmax.c
tests/tcg/s390x/Makefile.target +7 -0
index 3124172736..1a7a4a2f59 100644
@@ -17,6 +17,13 @@ TESTS+=trap
 TESTS+=signals-s390x
 TESTS+=branch-relative-long
 
+Z14_TESTS=vfminmax
+vfminmax: LDFLAGS+=-lm
+$(Z14_TESTS): CFLAGS+=-march=z14 -O2
+
+TESTS+=$(if $(shell $(CC) -march=z14 -S -o /dev/null -xc /dev/null \
+			 >/dev/null 2>&1 && echo OK),$(Z14_TESTS))
+
 VECTOR_TESTS=vxeh2_vs
 VECTOR_TESTS+=vxeh2_vcvt
 VECTOR_TESTS+=vxeh2_vlstr
tests/tcg/s390x/vfminmax.c new file +426 -0

426 lines not on this page · raw message

Signature · 1 line
2.35.3
Reviewed-by on v1 1/3
#2 lore raw

Re: [PATCH 1/3] target/s390x: fix handling of zeroes in vfmin/vfmax

In reply to Ilya Leoshkevich
Quoted 11 lines · Ilya Leoshkevich

vfmin_res() / vfmax_res() are trying to check whether a and b are both zeroes, but in reality they check that they are the same kind of zero. This causes incorrect results when comparing positive and negative zeroes.

Fixes da4807527f3b ("s390x/tcg: Implement VECTOR FP (MAXIMUM|MINIMUM)")
---
  target/s390x/tcg/vec_fpu_helper.c | 4 ++--
  1 file changed, 2 insertions(+), 2 deletions(-)

r~

Reviewed-by on v1 2/3
#5 lore raw

Re: [PATCH 2/3] target/s390x: fix NaN propagation rules

In reply to Ilya Leoshkevich
Quoted 6 lines · Ilya Leoshkevich

s390x has the same NaN propagation rules as ARM, and not as x86.

---
  fpu/softfloat-specialize.c.inc | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)

Yes, I think no one has those x86 rules, including x86. :-)

r~

Reviewed-by on v1 3/3
#8 lore raw

Re: [PATCH 3/3] tests/tcg/s390x: test signed vfmin/vfmax

In reply to Ilya Leoshkevich
Quoted 18 lines · Ilya Leoshkevich
tests/tcg/s390x/vfminmax.c new file +18 -0
+/*
+ * vfmin/vfmax code generation.
+ */
+extern const char vfminmax_template[];
+extern const int vfminmax_template_size;
+extern const int vfminmax_offset;
+asm(".globl vfminmax_template\n"
+    "vfminmax_template:\n"
+    "vl %v25,0(%r3)\n"
+    "vl %v26,0(%r4)\n"
+    "0: vfmax %v24,%v25,%v26,2,0,0\n"
+    "vst %v24,0(%r2)\n"
+    "br %r14\n"
+    "1: .align 4\n"
+    ".globl vfminmax_template_size\n"
+    "vfminmax_template_size: .long 1b - vfminmax_template\n"
+    ".globl vfminmax_offset\n"
+    "vfminmax_offset: .long 0b - vfminmax_template\n");

...

Quoted 5 lines
tests/tcg/s390x/vfminmax.c new file +5 -0
+
+#define VFMIN 0xEE
+#define VFMAX 0xEF
+
+static void vfminmax(unsigned char *buf, unsigned int op,

tests/tcg/s390x/vfminmax.c line 40

Quoted 10 lines
tests/tcg/s390x/vfminmax.c new file +10 -0
+                     unsigned int m4, unsigned int m5, unsigned int m6,
+                     void *v1, const void *v2, const void *v3)
+{
+    memcpy(buf, vfminmax_template, vfminmax_template_size);
+    buf[vfminmax_offset + 3] = (m6 << 4) | m5;
+    buf[vfminmax_offset + 4] &= 0x0F;
+    buf[vfminmax_offset + 4] |= (m4 << 4);
+    buf[vfminmax_offset + 5] = op;
+    ((void (*)(void *, const void *, const void *))buf)(v1, v2, v3);
+}

This works, of course. It could be simpler using EXECUTE, to store just the one instruction and not worry about an executable mapped page, but I guess it doesn't matter.

r~

Reviewed-by on v1 1/3
#3 lore raw

Re: [PATCH 1/3] target/s390x: fix handling of zeroes in vfmin/vfmax

In reply to Ilya Leoshkevich
Quoted 26 lines · Ilya Leoshkevich

vfmin_res() / vfmax_res() are trying to check whether a and b are both zeroes, but in reality they check that they are the same kind of zero. This causes incorrect results when comparing positive and negative zeroes.

Fixes da4807527f3b ("s390x/tcg: Implement VECTOR FP (MAXIMUM|MINIMUM)")
---
 target/s390x/tcg/vec_fpu_helper.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
target/s390x/tcg/vec_fpu_helper.c +1 -1
index 2a618a1093..75cf605b9f 100644
@@ -824,7 +824,7 @@ static S390MinMaxRes vfmin_res(uint16_t dcmask_a, uint16_t dcmask_b,
         default:
             g_assert_not_reached();
         }
-    } else if (unlikely(dcmask_a & dcmask_b & DCMASK_ZERO)) {
+    } else if (unlikely((dcmask_a & DCMASK_ZERO) && (dcmask_b & DCMASK_ZERO))) {
         switch (type) {
         case S390_MINMAX_TYPE_JAVA:
             return neg_a ? S390_MINMAX_RES_A : S390_MINMAX_RES_B;

target/s390x/tcg/vec_fpu_helper.c static S390MinMaxRes vfmax_res(uint16_t dcmask_a, uint16_t dcmask_b, · line 880

Quoted 8 lines
target/s390x/tcg/vec_fpu_helper.c +1 -1
@@ -874,7 +874,7 @@ static S390MinMaxRes vfmax_res(uint16_t dcmask_a, uint16_t dcmask_b,
         default:
             g_assert_not_reached();
         }
-    } else if (unlikely(dcmask_a & dcmask_b & DCMASK_ZERO)) {
+    } else if (unlikely((dcmask_a & DCMASK_ZERO) && (dcmask_b & DCMASK_ZERO))) {
         const bool neg_a = dcmask_a & DCMASK_NEGATIVE;
 
         switch (type) {

Thanks!

Signature · 3 lines
Thanks,

David / dhildenb
Reviewed-by on v1 2/3
#6 lore raw

Re: [PATCH 2/3] target/s390x: fix NaN propagation rules

In reply to Ilya Leoshkevich
Quoted 12 lines · Ilya Leoshkevich

s390x has the same NaN propagation rules as ARM, and not as x86.

---
 fpu/softfloat-specialize.c.inc | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fpu/softfloat-specialize.c.inc +0 -0
index 943e3301d2..a43ff5e02e 100644

fpu/softfloat-specialize.c.inc bool float32_is_signaling_nan(float32 a_, float_status *status) · line 397

Quoted 9 lines
fpu/softfloat-specialize.c.inc +2 -1
@@ -390,7 +390,8 @@ bool float32_is_signaling_nan(float32 a_, float_status *status)
 static int pickNaN(FloatClass a_cls, FloatClass b_cls,
                    bool aIsLargerSignificand, float_status *status)
 {
-#if defined(TARGET_ARM) || defined(TARGET_MIPS) || defined(TARGET_HPPA)
+#if defined(TARGET_ARM) || defined(TARGET_MIPS) || defined(TARGET_HPPA) || \
+    defined(TARGET_S390X)
     /* ARM mandated NaN propagation rules (see FPProcessNaNs()), take
      * the first of:
      *  1. A if it is signaling
Signature · 3 lines
Thanks,

David / dhildenb
#9 lore raw

Re: [PATCH 3/3] tests/tcg/s390x: test signed vfmin/vfmax

In reply to Richard Henderson
Quoted 43 lines · Richard Henderson
Quoted 18 lines · Ilya Leoshkevich
tests/tcg/s390x/vfminmax.c new file +18 -0
+/*
+ * vfmin/vfmax code generation.
+ */
+extern const char vfminmax_template[];
+extern const int vfminmax_template_size;
+extern const int vfminmax_offset;
+asm(".globl vfminmax_template\n"
+    "vfminmax_template:\n"
+    "vl %v25,0(%r3)\n"
+    "vl %v26,0(%r4)\n"
+    "0: vfmax %v24,%v25,%v26,2,0,0\n"
+    "vst %v24,0(%r2)\n"
+    "br %r14\n"
+    "1: .align 4\n"
+    ".globl vfminmax_template_size\n"
+    "vfminmax_template_size: .long 1b - vfminmax_template\n"
+    ".globl vfminmax_offset\n"
+    "vfminmax_offset: .long 0b - vfminmax_template\n");

...

Quoted 9 lines
tests/tcg/s390x/vfminmax.c new file +6 -0
+
+#define VFMIN 0xEE
+#define VFMAX 0xEF
+
+static void vfminmax(unsigned char *buf, unsigned int op,
+                     unsigned int m4, unsigned int m5, unsigned

int m6,
+                     void *v1, const void *v2, const void *v3) +{

Quoted 8 lines

+    memcpy(buf, vfminmax_template, vfminmax_template_size); +    buf[vfminmax_offset + 3] = (m6 << 4) | m5;
+    buf[vfminmax_offset + 4] &= 0x0F;
+    buf[vfminmax_offset + 4] |= (m4 << 4);
+    buf[vfminmax_offset + 5] = op;
+    ((void (*)(void *, const void *, const void *))buf)(v1, v2, v3);
+}

This works, of course.  It could be simpler using EXECUTE, to store just the one
instruction and not worry about an executable mapped page, but I guess it doesn't matter.

Quoted 4 lines

r~

Thanks!

I thought about this too, but EX/EXRL operate only on the second byte, and I need to modify bytes 3-5 here.

#10 lore raw

Re: [PATCH 3/3] tests/tcg/s390x: test signed vfmin/vfmax

Quoted 7 lines · Ilya Leoshkevich
Quoted 4 lines

This works, of course.  It could be simpler using EXECUTE, to store just the one
instruction and not worry about an executable mapped page, but I guess it doesn't matter.

I thought about this too, but EX/EXRL operate only on the second byte, and I need to modify bytes 3-5 here.

I didn't mean modify the instruction via EX, but something like

  static char minmax[6] __attribute__((aligned(2)))
    = { xx, yy, zz, 0, 0, 0 };

minmax[3] = m6 ...
minmax[4] = ...
minmax[5] = op;

asm("vl %%v25,0(%1)\n"

      "vl %%v26,0(%2)\n"
      "ex 0,0(%3)\n"
      "vst %%v24,0(%0)"
      : : "a"(v1), "a"(v2), "a"(v3), "a"(minmax)
      : "memory", "v24", "v25", "v26);

r~

#11 lore raw

Re: [PATCH 3/3] tests/tcg/s390x: test signed vfmin/vfmax

Quoted 28 lines · Richard Henderson
Quoted 9 lines · Ilya Leoshkevich
Quoted 5 lines

This works, of course.  It could be simpler using EXECUTE, to store
just the one
instruction and not worry about an executable mapped page, but I guess it doesn't matter.

I thought about this too, but EX/EXRL operate only on the second byte,
and I need to modify bytes 3-5 here.

I didn't mean modify the instruction via EX, but something like

   static char minmax[6] __attribute__((aligned(2))) = { xx, yy, zz, 0, 0, 0 };

   minmax[3] = m6 ...
   minmax[4] = ...
   minmax[5] = op;

   asm("vl %%v25,0(%1)\n"
       "vl %%v26,0(%2)\n"
       "ex 0,0(%3)\n"
       "vst %%v24,0(%0)"
       : : "a"(v1), "a"(v2), "a"(v3), "a"(minmax)
       : "memory", "v24", "v25", "v26);

Quoted 1 line

r~

Nice trick!

This works in qemu, but not natively: EX target must be executable. I'd still like to try to find a way to establish an rwx section, and send a v2 with this improvement.

I guess we'll need to fix the access check discrepancy some day.