Misc goodies in the JIT in preparation for more aggressive code motion.
- Set up resource masks correctly for Thumb push/pop when LR/PC are involved.
- Preserve LR around simulated heap references under self-verification mode.
- Compact a few simple flags in ArmLIR into bit fields.
- Minor performance tuning in TEMPLATE_MEM_OP_DECODE
Change-Id: Id73edac837c5bb37dfd21f372d6fa21c238cf42a
diff --git a/vm/compiler/codegen/arm/CodegenDriver.c b/vm/compiler/codegen/arm/CodegenDriver.c
index 1b7c6ed..74f648b 100644
--- a/vm/compiler/codegen/arm/CodegenDriver.c
+++ b/vm/compiler/codegen/arm/CodegenDriver.c
@@ -167,6 +167,9 @@
dvmCompilerClobberCallRegs(cUnit);
rlResult = dvmCompilerGetReturnWide(cUnit);
storeValueWide(cUnit, rlDest, rlResult);
+#if defined(WITH_SELF_VERIFICATION)
+ cUnit->usesLinkRegister = true;
+#endif
return false;
}
@@ -213,6 +216,31 @@
dvmCompilerInsertLIRBefore(currentLIR, (LIR *) insn);
}
+/*
+ * Example where r14 (LR) is preserved around a heap access under
+ * self-verification mode in Thumb2:
+ *
+ * D/dalvikvm( 1538): 0x59414c5e (0026): ldr r14, [rpc, #220] <-hoisted
+ * D/dalvikvm( 1538): 0x59414c62 (002a): mla r4, r0, r8, r4
+ * D/dalvikvm( 1538): 0x59414c66 (002e): adds r3, r4, r3
+ * D/dalvikvm( 1538): 0x59414c6a (0032): push <r5, r14> ---+
+ * D/dalvikvm( 1538): 0x59414c6c (0034): blx_1 0x5940f494 |
+ * D/dalvikvm( 1538): 0x59414c6e (0036): blx_2 see above <-MEM_OP_DECODE
+ * D/dalvikvm( 1538): 0x59414c70 (0038): ldr r10, [r9, #0] |
+ * D/dalvikvm( 1538): 0x59414c74 (003c): pop <r5, r14> ---+
+ * D/dalvikvm( 1538): 0x59414c78 (0040): mov r11, r10
+ * D/dalvikvm( 1538): 0x59414c7a (0042): asr r12, r11, #31
+ * D/dalvikvm( 1538): 0x59414c7e (0046): movs r0, r2
+ * D/dalvikvm( 1538): 0x59414c80 (0048): movs r1, r3
+ * D/dalvikvm( 1538): 0x59414c82 (004a): str r2, [r5, #16]
+ * D/dalvikvm( 1538): 0x59414c84 (004c): mov r2, r11
+ * D/dalvikvm( 1538): 0x59414c86 (004e): str r3, [r5, #20]
+ * D/dalvikvm( 1538): 0x59414c88 (0050): mov r3, r12
+ * D/dalvikvm( 1538): 0x59414c8a (0052): str r11, [r5, #24]
+ * D/dalvikvm( 1538): 0x59414c8e (0056): str r12, [r5, #28]
+ * D/dalvikvm( 1538): 0x59414c92 (005a): blx r14 <-use of LR
+ *
+ */
static void selfVerificationBranchInsertPass(CompilationUnit *cUnit)
{
ArmLIR *thisLIR;
@@ -221,7 +249,19 @@
for (thisLIR = (ArmLIR *) cUnit->firstLIRInsn;
thisLIR != (ArmLIR *) cUnit->lastLIRInsn;
thisLIR = NEXT_LIR(thisLIR)) {
- if (thisLIR->branchInsertSV) {
+ if (!thisLIR->flags.isNop && thisLIR->flags.insertWrapper) {
+ /*
+ * Push r5(FP) and r14(LR) onto stack. We need to make sure that
+ * SP is 8-byte aligned, and we use r5 as a temp to restore LR
+ * for Thumb-only target since LR cannot be directly accessed in
+ * Thumb mode. Another reason to choose r5 here is it is the Dalvik
+ * frame pointer and cannot be the target of the emulated heap
+ * load.
+ */
+ if (cUnit->usesLinkRegister) {
+ genSelfVerificationPreBranch(cUnit, thisLIR);
+ }
+
/* Branch to mem op decode template */
selfVerificationBranchInsert((LIR *) thisLIR, kThumbBlx1,
(int) gDvmJit.codeCache + templateEntryOffsets[opcode],
@@ -229,6 +269,11 @@
selfVerificationBranchInsert((LIR *) thisLIR, kThumbBlx2,
(int) gDvmJit.codeCache + templateEntryOffsets[opcode],
(int) gDvmJit.codeCache + templateEntryOffsets[opcode]);
+
+ /* Restore LR */
+ if (cUnit->usesLinkRegister) {
+ genSelfVerificationPostBranch(cUnit, thisLIR);
+ }
}
}
}
@@ -708,6 +753,9 @@
else
rlResult = dvmCompilerGetReturnWideAlt(cUnit);
storeValueWide(cUnit, rlDest, rlResult);
+#if defined(WITH_SELF_VERIFICATION)
+ cUnit->usesLinkRegister = true;
+#endif
}
return false;
}