|
 |
8e0b80c |
commit 71002d8a5111d02ce8049c55017a8d948c820e35
|
|
|
8e0b80c |
Author: Andreas Arnez <arnez@linux.ibm.com>
|
|
|
8e0b80c |
Date: Thu Oct 25 13:47:12 2018 +0200
|
|
|
8e0b80c |
|
|
|
8e0b80c |
Bug 400490 s390x: Fix register allocation for VRs vs FPRs
|
|
|
8e0b80c |
|
|
|
8e0b80c |
On s390x, if vector registers are available, they are fed to the register
|
|
|
8e0b80c |
allocator as if they were separate from the floating-point registers. But
|
|
|
8e0b80c |
in fact the FPRs are embedded in the VRs. So for instance, if both f3 and
|
|
|
8e0b80c |
v3 are allocated and used at the same time, corruption will result.
|
|
|
8e0b80c |
|
|
|
8e0b80c |
This is fixed by offering only the non-overlapping VRs, v16 to v31, to the
|
|
|
8e0b80c |
register allocator instead.
|
|
|
8e0b80c |
|
|
|
8e0b80c |
diff --git a/VEX/priv/host_s390_defs.c b/VEX/priv/host_s390_defs.c
|
|
|
8e0b80c |
index 6c22ac8..98ac938 100644
|
|
|
8e0b80c |
--- a/VEX/priv/host_s390_defs.c
|
|
|
8e0b80c |
+++ b/VEX/priv/host_s390_defs.c
|
|
|
8e0b80c |
@@ -59,7 +59,6 @@ static UInt s390_tchain_load64_len(void);
|
|
|
8e0b80c |
|
|
|
8e0b80c |
/* A mapping from register number to register index */
|
|
|
8e0b80c |
static Int gpr_index[16]; // GPR regno -> register index
|
|
|
8e0b80c |
-static Int fpr_index[16]; // FPR regno -> register index
|
|
|
8e0b80c |
static Int vr_index[32]; // VR regno -> register index
|
|
|
8e0b80c |
|
|
|
8e0b80c |
HReg
|
|
|
8e0b80c |
@@ -73,7 +72,7 @@ s390_hreg_gpr(UInt regno)
|
|
|
8e0b80c |
HReg
|
|
|
8e0b80c |
s390_hreg_fpr(UInt regno)
|
|
|
8e0b80c |
{
|
|
|
8e0b80c |
- Int ix = fpr_index[regno];
|
|
|
8e0b80c |
+ Int ix = vr_index[regno];
|
|
|
8e0b80c |
vassert(ix >= 0);
|
|
|
8e0b80c |
return mkHReg(/*virtual*/False, HRcFlt64, regno, ix);
|
|
|
8e0b80c |
}
|
|
|
8e0b80c |
@@ -463,11 +462,9 @@ getRRegUniverse_S390(void)
|
|
|
8e0b80c |
|
|
|
8e0b80c |
RRegUniverse__init(ru);
|
|
|
8e0b80c |
|
|
|
8e0b80c |
- /* Assign invalid values to the gpr/fpr/vr_index */
|
|
|
8e0b80c |
+ /* Assign invalid values to the gpr/vr_index */
|
|
|
8e0b80c |
for (UInt i = 0; i < sizeof gpr_index / sizeof gpr_index[0]; ++i)
|
|
|
8e0b80c |
gpr_index[i] = -1;
|
|
|
8e0b80c |
- for (UInt i = 0; i < sizeof fpr_index / sizeof fpr_index[0]; ++i)
|
|
|
8e0b80c |
- fpr_index[i] = -1;
|
|
|
8e0b80c |
for (UInt i = 0; i < sizeof vr_index / sizeof vr_index[0]; ++i)
|
|
|
8e0b80c |
vr_index[i] = -1;
|
|
|
8e0b80c |
|
|
|
8e0b80c |
@@ -494,17 +491,17 @@ getRRegUniverse_S390(void)
|
|
|
8e0b80c |
|
|
|
8e0b80c |
ru->allocable_start[HRcFlt64] = ru->size;
|
|
|
8e0b80c |
for (UInt regno = 8; regno <= 15; ++regno) {
|
|
|
8e0b80c |
- fpr_index[regno] = ru->size;
|
|
|
8e0b80c |
+ vr_index[regno] = ru->size;
|
|
|
8e0b80c |
ru->regs[ru->size++] = s390_hreg_fpr(regno);
|
|
|
8e0b80c |
}
|
|
|
8e0b80c |
for (UInt regno = 0; regno <= 7; ++regno) {
|
|
|
8e0b80c |
- fpr_index[regno] = ru->size;
|
|
|
8e0b80c |
+ vr_index[regno] = ru->size;
|
|
|
8e0b80c |
ru->regs[ru->size++] = s390_hreg_fpr(regno);
|
|
|
8e0b80c |
}
|
|
|
8e0b80c |
ru->allocable_end[HRcFlt64] = ru->size - 1;
|
|
|
8e0b80c |
|
|
|
8e0b80c |
ru->allocable_start[HRcVec128] = ru->size;
|
|
|
8e0b80c |
- for (UInt regno = 0; regno <= 31; ++regno) {
|
|
|
8e0b80c |
+ for (UInt regno = 16; regno <= 31; ++regno) {
|
|
|
8e0b80c |
vr_index[regno] = ru->size;
|
|
|
8e0b80c |
ru->regs[ru->size++] = s390_hreg_vr(regno);
|
|
|
8e0b80c |
}
|
|
|
8e0b80c |
@@ -527,12 +524,12 @@ getRRegUniverse_S390(void)
|
|
|
8e0b80c |
/* Sanity checking */
|
|
|
8e0b80c |
for (UInt i = 0; i < sizeof gpr_index / sizeof gpr_index[0]; ++i)
|
|
|
8e0b80c |
vassert(gpr_index[i] >= 0);
|
|
|
8e0b80c |
- for (UInt i = 0; i < sizeof fpr_index / sizeof fpr_index[0]; ++i)
|
|
|
8e0b80c |
- vassert(fpr_index[i] >= 0);
|
|
|
8e0b80c |
for (UInt i = 0; i < sizeof vr_index / sizeof vr_index[0]; ++i)
|
|
|
8e0b80c |
vassert(vr_index[i] >= 0);
|
|
|
8e0b80c |
|
|
|
8e0b80c |
initialised = True;
|
|
|
8e0b80c |
+
|
|
|
8e0b80c |
+ RRegUniverse__check_is_sane(ru);
|
|
|
8e0b80c |
return ru;
|
|
|
8e0b80c |
}
|
|
|
8e0b80c |
|