Revert "Revert "ART: Implement X86 hard float (Quick/JNI/Baseline)""
This reverts commit 949c91fb91f40a4a80b2b492913cf8541008975e.
This time, don't clobber EBX before saving it.
Redo some of the macros to make register usage explicit.
Change-Id: I8db8662877cd006816e16a28f42444ab7c36bfef
diff --git a/runtime/arch/x86/context_x86.cc b/runtime/arch/x86/context_x86.cc
index 2a6ff14..4ea4684 100644
--- a/runtime/arch/x86/context_x86.cc
+++ b/runtime/arch/x86/context_x86.cc
@@ -30,6 +30,9 @@
for (size_t i = 0; i < kNumberOfCpuRegisters; i++) {
gprs_[i] = nullptr;
}
+ for (size_t i = 0; i < kNumberOfFloatRegisters; ++i) {
+ fprs_[i] = nullptr;
+ }
gprs_[ESP] = &esp_;
// Initialize registers with easy to spot debug values.
esp_ = X86Context::kBadGprBase + ESP;
@@ -40,7 +43,7 @@
mirror::ArtMethod* method = fr.GetMethod();
const QuickMethodFrameInfo frame_info = method->GetQuickFrameInfo();
size_t spill_count = POPCOUNT(frame_info.CoreSpillMask());
- DCHECK_EQ(frame_info.FpSpillMask(), 0u);
+ size_t fp_spill_count = POPCOUNT(frame_info.FpSpillMask());
if (spill_count > 0) {
// Lowest number spill is farthest away, walk registers and fill into context.
int j = 2; // Offset j to skip return address spill.
@@ -51,6 +54,24 @@
}
}
}
+ if (fp_spill_count > 0) {
+ // Lowest number spill is farthest away, walk registers and fill into context.
+ size_t j = 2; // Offset j to skip return address spill.
+ size_t fp_spill_size_in_words = fp_spill_count * 2;
+ for (size_t i = 0; i < kNumberOfFloatRegisters; ++i) {
+ if (((frame_info.FpSpillMask() >> i) & 1) != 0) {
+ // There are 2 pieces to each XMM register, to match VR size.
+ fprs_[2*i] = reinterpret_cast<uint32_t*>(
+ fr.CalleeSaveAddress(spill_count + fp_spill_size_in_words - j,
+ frame_info.FrameSizeInBytes()));
+ fprs_[2*i+1] = reinterpret_cast<uint32_t*>(
+ fr.CalleeSaveAddress(spill_count + fp_spill_size_in_words - j - 1,
+ frame_info.FrameSizeInBytes()));
+ // Two void* per XMM register.
+ j += 2;
+ }
+ }
+ }
}
void X86Context::SmashCallerSaves() {
@@ -59,6 +80,7 @@
gprs_[EDX] = const_cast<uintptr_t*>(&gZero);
gprs_[ECX] = nullptr;
gprs_[EBX] = nullptr;
+ memset(&fprs_[0], '\0', sizeof(fprs_));
}
void X86Context::SetGPR(uint32_t reg, uintptr_t value) {
@@ -68,14 +90,11 @@
*gprs_[reg] = value;
}
-uintptr_t X86Context::GetFPR(uint32_t reg ATTRIBUTE_UNUSED) {
- LOG(FATAL) << "Floating-point registers are all caller save in X86";
- UNREACHABLE();
-}
-
-void X86Context::SetFPR(uint32_t reg ATTRIBUTE_UNUSED, uintptr_t value ATTRIBUTE_UNUSED) {
- LOG(FATAL) << "Floating-point registers are all caller save in X86";
- UNREACHABLE();
+void X86Context::SetFPR(uint32_t reg, uintptr_t value) {
+ CHECK_LT(reg, static_cast<uint32_t>(kNumberOfFloatRegisters));
+ DCHECK(IsAccessibleFPR(reg));
+ CHECK_NE(fprs_[reg], reinterpret_cast<const uint32_t*>(&gZero));
+ *fprs_[reg] = value;
}
void X86Context::DoLongJump() {
@@ -86,17 +105,30 @@
for (size_t i = 0; i < kNumberOfCpuRegisters; ++i) {
gprs[kNumberOfCpuRegisters - i - 1] = gprs_[i] != nullptr ? *gprs_[i] : X86Context::kBadGprBase + i;
}
+ uint32_t fprs[kNumberOfFloatRegisters];
+ for (size_t i = 0; i < kNumberOfFloatRegisters; ++i) {
+ fprs[i] = fprs_[i] != nullptr ? *fprs_[i] : X86Context::kBadFprBase + i;
+ }
// We want to load the stack pointer one slot below so that the ret will pop eip.
uintptr_t esp = gprs[kNumberOfCpuRegisters - ESP - 1] - sizeof(intptr_t);
gprs[kNumberOfCpuRegisters] = esp;
*(reinterpret_cast<uintptr_t*>(esp)) = eip_;
__asm__ __volatile__(
+ "movl %1, %%ebx\n\t" // Address base of FPRs.
+ "movsd 0(%%ebx), %%xmm0\n\t" // Load up XMM0-XMM7.
+ "movsd 8(%%ebx), %%xmm1\n\t"
+ "movsd 16(%%ebx), %%xmm2\n\t"
+ "movsd 24(%%ebx), %%xmm3\n\t"
+ "movsd 32(%%ebx), %%xmm4\n\t"
+ "movsd 40(%%ebx), %%xmm5\n\t"
+ "movsd 48(%%ebx), %%xmm6\n\t"
+ "movsd 56(%%ebx), %%xmm7\n\t"
"movl %0, %%esp\n\t" // ESP points to gprs.
"popal\n\t" // Load all registers except ESP and EIP with values in gprs.
"popl %%esp\n\t" // Load stack pointer.
"ret\n\t" // From higher in the stack pop eip.
: // output.
- : "g"(&gprs[0]) // input.
+ : "g"(&gprs[0]), "g"(&fprs[0]) // input.
:); // clobber.
#else
UNIMPLEMENTED(FATAL);