[RuntimeDyldChecker][AArch32] Add a PC offset to next_PC for ARM targets (#91746)

In ARM mode, the Program Counter (PC) points to the current instruction's
address + 8 instead of + 4. An offset is added to RuntimeDyldChecker to
use `next_pc` expression in JITLink tests with both Thumb and Arm.

Author: eymay

llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldChecker.cpp

@@ -369,7 +369,13 @@ class RuntimeDyldCheckerExprEval {
     uint64_t SymbolAddr = PCtx.IsInsideLoad
                               ? Checker.getSymbolLocalAddr(Symbol)
                               : Checker.getSymbolRemoteAddr(Symbol);
-    uint64_t NextPC = SymbolAddr + InstSize;
+
+    // ARM PC offset is 8 instead of 4, because it accounts for an additional
+    // prefetch instruction that increments PC even though it is implicit.
+    auto TT = Checker.getTripleForSymbol(Checker.getTargetFlag(Symbol));
+    uint64_t PCOffset = TT.getArch() == Triple::ArchType::arm ? 4 : 0;
+
+    uint64_t NextPC = SymbolAddr + InstSize + PCOffset;
 
     return std::make_pair(EvalResult(NextPC), RemainingExpr);
   }

llvm/test/ExecutionEngine/JITLink/AArch32/ELF_relocations_arm.s

@@ -23,23 +23,29 @@
 	.syntax unified
 
 # CHECK-TYPE: {{[0-9a-f]+}} R_ARM_CALL call_target_arm
-# CHECK-INSTR: 	00000000 <call_site>:
+# CHECK-INSTR: 	00000000 <call_site_a2a>:
 # CHECK-INSTR: 	       0: ebfffffe     bl
-# CHECK-INSTR: 	       4: ebfffffe     bl
-# CHECK-INSTR: 	0000000c <call_target_arm>
-# CHECK-INSTR: 	00000010 <call_target_thumb>
-# ARM branch offset is 8, because it accounts for an additional prefetch
-# instruction that increments PC even though it is implicit
-# jitlink-check: decode_operand(call_site + 0, 0) = call_target_arm   - (call_site +  8)
-# jitlink-check: decode_operand(call_site + 4, 0) = call_target_thumb - (call_site + 12)
-	.globl	call_site
-	.type	call_site,%function
+# CHECK-INSTR: 	00000008 <call_site_a2t>:
+# CHECK-INSTR: 	       8: ebfffffe     bl
+# CHECK-INSTR: 	00000010 <call_target_arm>
+# CHECK-INSTR: 	00000014 <call_target_thumb>
+# jitlink-check: decode_operand(call_site_a2a, 0) = call_target_arm   - next_pc(call_site_a2a)
+# jitlink-check: decode_operand(call_site_a2t, 0) = call_target_thumb - next_pc(call_site_a2t)
+	.globl	call_site_a2a
+	.type	call_site_a2a, %function
 	.p2align	2
-call_site:
+call_site_a2a:
 	bl	call_target_arm
+	bx	lr
+	.size	call_site_a2a, .-call_site_a2a
+
+	.globl	call_site_a2t
+	.type	call_site_a2t,%function
+	.p2align	2
+call_site_a2t:
 	bl	call_target_thumb
 	bx	lr
-	.size	call_site, .-call_site
+	.size	call_site_a2t, .-call_site_a2t
 
 	.globl	call_target_arm
 	.type	call_target_arm,%function
@@ -59,10 +65,10 @@ call_target_thumb:
 	.code 32
 
 # CHECK-TYPE: {{[0-9a-f]+}} R_ARM_JUMP24 jump24_target
-# CHECK-INSTR: 	00000014 <jump24_site>:
-# CHECK-INSTR: 	      14: eafffffe     b
-# CHECK-INSTR: 	00000018 <jump24_target>
-# jitlink-check: decode_operand(jump24_site, 0) = jump24_target - (jump24_site + 8)
+# CHECK-INSTR: 	00000018 <jump24_site>:
+# CHECK-INSTR: 	      18: eafffffe     b
+# CHECK-INSTR: 	0000001c <jump24_target>
+# jitlink-check: decode_operand(jump24_site, 0) = jump24_target - next_pc(jump24_site)
 	.globl	jump24_site
 	.type	jump24_site,%function
 	.p2align	2

llvm/test/ExecutionEngine/JITLink/AArch32/ELF_relocations_data.s

@@ -50,7 +50,7 @@ target1_abs32:
 # The +12 accounts for the ARM branch offset (8) and the .LPC offset (4), which
 # is stored as initial addend inline.
 # FIXME: We shouldn't need to substract the 64-bit sign-extension manually.
-# jitlink-check: *{4}got_prel_offset = got_addr(out.o, target) - (got_prel + 12) - 0xffffffff00000000
+# jitlink-check: *{4}got_prel_offset = got_addr(out.o, target) - (next_pc(got_prel) + 4) - 0xffffffff00000000
 	.globl	got_prel
 	.type	got_prel,%function
 	.p2align	2

llvm/test/ExecutionEngine/JITLink/AArch32/ELF_relocations_thumbv6m.s

@@ -30,7 +30,7 @@
 # CHECK-INSTR: f7ff fffe     bl
 # We decode the operand with index 2, because bl generates two leading implicit
 # predicate operands that we have to skip in order to decode the call_target operand
-# jitlink-check: decode_operand(call_site, 2) = call_target_thumb - (call_site + 4)
+# jitlink-check: decode_operand(call_site, 2) = call_target_thumb - next_pc(call_site)
 	.globl	call_site
 	.type	call_site,%function
 	.p2align	1

llvm/test/ExecutionEngine/JITLink/AArch32/ELF_stubs_arm.s

@@ -27,7 +27,7 @@
 # hard-code it in the immediate field.
 
 # The external function ext will return to the caller directly.
-# jitlink-check: decode_operand(test_arm_jump, 0) = stub_addr(out.o, ext) - (test_arm_jump + 8)
+# jitlink-check: decode_operand(test_arm_jump, 0) = stub_addr(out.o, ext) - next_pc(test_arm_jump)
 	.globl	test_arm_jump
 	.type	test_arm_jump,%function
 	.p2align	2
@@ -38,7 +38,7 @@ test_arm_jump:
 # The branch-with-link sets the LR register so that the external function ext
 # returns to us. We have to save the register (push) and return to main manually
 # (pop). This adds the +4 offset for the bl instruction we decode:
-# jitlink-check: decode_operand(test_arm_call + 4, 0) = stub_addr(out.o, ext) - (test_arm_call + 8) - 4
+# jitlink-check: decode_operand(test_arm_call + 4, 0) = stub_addr(out.o, ext) - next_pc(test_arm_call) - 4
 	.globl  test_arm_call
 	.type	test_arm_call,%function
 	.p2align	2

llvm/test/ExecutionEngine/JITLink/AArch32/ELF_stubs_multi.s

@@ -13,7 +13,7 @@
 # ascending size (because the default memory manager lays out blocks by size).
 
 # Thumb relocation site emits thumb stub
-# jitlink-check: decode_operand(test_stub_thumb, 0) = stub_addr(out.o, ext, thumb) - (test_stub_thumb + 4)
+# jitlink-check: decode_operand(test_stub_thumb, 0) = stub_addr(out.o, ext, thumb) - next_pc(test_stub_thumb)
 	.globl  test_stub_thumb
 	.type	test_stub_thumb,%function
 	.p2align	1
@@ -24,7 +24,7 @@ test_stub_thumb:
 	.size	test_stub_thumb, .-test_stub_thumb
 
 # Arm relocation site emits arm stub
-# jitlink-check: decode_operand(test_stub_arm, 0) = stub_addr(out.o, ext, arm) - (test_stub_arm + 8)
+# jitlink-check: decode_operand(test_stub_arm, 0) = stub_addr(out.o, ext, arm) - next_pc(test_stub_arm)
 	.globl  test_stub_arm
 	.type	test_stub_arm,%function
 	.p2align	2