Bring throw helpers to PUSH_COOP_PINVOKE_FRAME plan

src/coreclr/pal/inc/unixasmmacrosamd64.inc

@@ -428,6 +428,39 @@ C_FUNC(\Name\()_End):
         POP_CALLEE_SAVED_REGISTERS
 .endm
 
+// Pushes a full TransitionBlock on the stack including argument registers and
+// floating point argument registers. Used for exception throw helpers where we
+// need to capture the complete register state.
+//
+// Stack layout (from high to low address after prologue):
+//   Return address
+//   CalleeSavedRegisters (rbp, rbx, r15, r14, r13, r12 - 48 bytes)
+//   ArgumentRegisters (r9, r8, rcx, rdx, rsi, rdi - 48 bytes)  <- TransitionBlock
+//   Padding (16 bytes for 16-byte alignment)
+//   FloatArgumentRegisters (xmm0-xmm7, 128 bytes)
+//   sp points here
+//
+// Stack alignment: After call (8) + callee-saved (48) + arg regs (48) + alloc (144) + call (8) = 256 bytes
+// 256 % 16 = 0, so stack is properly aligned for the call to IL_Throw_Impl.
+//
+// On exit, \target contains the TransitionBlock pointer (after float args area).
+.macro PUSH_COOP_PINVOKE_FRAME_WITH_FLOATS target
+        set_cfa_register rsp, 8
+
+        PUSH_CALLEE_SAVED_REGISTERS
+        PUSH_ARGUMENT_REGISTERS
+
+        // Allocate space for alignment (16 bytes) + float argument registers (128 bytes) = 144 bytes
+        alloc_stack     144
+        // Save float argument registers at offset 16 (after alignment padding)
+        // This ensures floats are at TransitionBlock - 128 (matching GetOffsetOfFloatArgumentRegisters)
+        SAVE_FLOAT_ARGUMENT_REGISTERS 16
+
+        END_PROLOGUE
+
+        lea     \target, [rsp + 144]
+.endm
+
 .macro INLINE_GETTHREAD
         // Inlined version of call C_FUNC(RhpGetThread)
         INLINE_GET_TLS_VAR t_CurrentThreadInfo

src/coreclr/pal/inc/unixasmmacrosarm.inc

@@ -288,6 +288,42 @@ C_FUNC(\Name):
         add     \target, sp, 4
 .endm
 
+// Pushes a full TransitionBlock on the stack including float argument registers.
+// On exit, \target contains the TransitionBlock pointer.
+//
+// Stack layout (from sp going up):
+//   sp+0:   d8-d15 (64 bytes) - FP callee-saved
+//   sp+64:  padding (4 bytes) - to make d0-d7 8-byte aligned at TransitionBlock-68
+//   sp+68:  d0-d7 (64 bytes) - float argument registers (at TransitionBlock - 68)
+//   sp+132: padding (4 bytes) - to keep total allocation 8-byte aligned  
+//   sp+136: TransitionBlock starts here (CalleeSavedRegisters + ArgumentRegisters pushed above)
+//
+// GetNegSpaceSize() for ARM32 = 64 (FloatArgumentRegisters) + 4 (padding) = 68
+// GetOffsetOfFloatArgumentRegisters() = -68
+//
+// Total stack alloc: 64 + 4 + 64 + 4 = 136 bytes
+// Stack: Arguments(16) + callee-saved(36) + alloc(136) = 188 bytes
+// 188 % 4 = 0, properly aligned for ARM32
+.macro PUSH_COOP_PINVOKE_FRAME_WITH_FLOATS target
+        // Push argument registers (r0-r3) - these will be at highest address in TransitionBlock
+        PUSH_ARGUMENT_REGISTERS
+        PUSH_CALLEE_SAVED_REGISTERS
+        PROLOG_STACK_SAVE_OFFSET r7, #12
+        // let r7 point the saved r7 in the stack (clang FP style)
+        // Allocate space for: d8-d15 (64) + padding (4) + d0-d7 (64) + padding (4) = 136 bytes
+        alloc_stack 136
+        // Save floating point argument registers (d0-d7) at sp+68 (TransitionBlock - 68)
+        add     r12, sp, #68
+        vstm    r12, {d0-d7}
+        // Save FP callee-saved registers (d8-d15) at sp+0
+        add     r12, sp, #0
+        vstm    r12, {d8-d15}
+        CHECK_STACK_ALIGNMENT
+        END_PROLOGUE
+        // TransitionBlock is at sp + 136
+        add     \target, sp, #136
+.endm
+
 .macro POP_COOP_PINVOKE_FRAME
         free_stack 4
         POP_CALLEE_SAVED_REGISTERS

src/coreclr/pal/inc/unixasmmacrosarm64.inc

@@ -419,7 +419,54 @@ C_FUNC(\Name\()_End):
         EPILOG_RESTORE_REG_PAIR   x25, x26, 64
         EPILOG_RESTORE_REG_PAIR   x27, x28, 80
         EPILOG_RESTORE_REG_PAIR_INDEXED fp, lr, 176
-.endm 
+.endm
+
+// Pushes a full TransitionBlock on the stack including argument registers and
+// floating point argument registers. Used for exception throw helpers where we
+// need to capture the complete register state including FP callee-saved registers.
+//
+// Stack layout (from low to high address):
+//   sp+0:     FP callee-saved registers (d8-d15, 64 bytes)
+//   sp+64:    FloatArgumentRegisters (q0-q7, 128 bytes)
+//   sp+192:   TransitionBlock start (176 bytes)
+//             - CalleeSavedRegisters (fp, lr, x19-x28 - 96 bytes)
+//             - padding (8 bytes)
+//             - x8 (8 bytes)
+//             - ArgumentRegisters (x0-x7, 64 bytes)
+//
+// On exit, \target contains the TransitionBlock pointer (sp+192).
+.macro PUSH_COOP_PINVOKE_FRAME_WITH_FLOATS target
+        PROLOG_SAVE_REG_PAIR_INDEXED   fp, lr, -176
+
+        // Spill callee saved registers
+        PROLOG_SAVE_REG_PAIR   x19, x20, 16
+        PROLOG_SAVE_REG_PAIR   x21, x22, 32
+        PROLOG_SAVE_REG_PAIR   x23, x24, 48
+        PROLOG_SAVE_REG_PAIR   x25, x26, 64
+        PROLOG_SAVE_REG_PAIR   x27, x28, 80
+
+        // Allocate space for FloatArgumentRegisters (128) + FP callee-saved (64) = 192 bytes
+        PROLOG_STACK_ALLOC 192
+
+        // Save argument registers (x8, x0-x7) at offset 296 from sp (192 + 104)
+        SAVE_ARGUMENT_REGISTERS sp, 296
+
+        // Save floating point argument registers (q0-q7) at sp+64
+        SAVE_FLOAT_ARGUMENT_REGISTERS sp, 64
+
+        // Save FP callee-saved registers (d8-d15) at sp+0
+        str     d8,  [sp, #0]
+        str     d9,  [sp, #8]
+        str     d10, [sp, #16]
+        str     d11, [sp, #24]
+        str     d12, [sp, #32]
+        str     d13, [sp, #40]
+        str     d14, [sp, #48]
+        str     d15, [sp, #56]
+
+        // Set target to TransitionBlock pointer
+        add     \target, sp, #192
+.endm
 
 // ------------------------------------------------------------------
 // Macro to generate Redirection Stubs

src/coreclr/pal/inc/unixasmmacrosloongarch64.inc

@@ -404,6 +404,45 @@ C_FUNC(\Name\()_End):
         EPILOG_STACK_FREE  160
 .endm
 
+// Pushes a full TransitionBlock on the stack including argument registers and
+// floating point argument registers. Used for exception throw helpers where we
+// need to capture the complete register state.
+//
+// Stack layout (from low to high address):
+//   sp+0:     FloatArgumentRegisters (fa0-fa7, 64 bytes)
+//   sp+64:    TransitionBlock start
+//             - CalleeSavedRegisters (fp, ra, s0-s8 - 96 bytes)
+//             - ArgumentRegisters (a0-a7, 64 bytes)
+//
+// On exit, \target contains the TransitionBlock pointer (sp+128).
+.macro PUSH_COOP_PINVOKE_FRAME_WITH_FLOATS target
+        // Stack: FPCalleeSaved(64) + FloatArgs(64) + CalleeSaved(96) + Args(64) = 288 bytes
+        PROLOG_STACK_ALLOC 288
+        PROLOG_SAVE_REG_PAIR  22, 1, 128, 1
+
+        // Save callee-saved registers at offset 128 (after FP callee-saved and FloatArgumentRegisters)
+        SAVE_CALLEESAVED_REGISTERS $sp, 128
+
+        // Save argument registers (a0-a7) at offset 224
+        SAVE_ARGUMENT_REGISTERS $sp, 224
+
+        // Save floating-point argument registers (fa0-fa7) at offset 64
+        SAVE_FLOAT_ARGUMENT_REGISTERS $sp, 64
+
+        // Save FP callee-saved registers (f24-f31) at offset 0
+        fst.d   $f24, $sp, 0
+        fst.d   $f25, $sp, 8
+        fst.d   $f26, $sp, 16
+        fst.d   $f27, $sp, 24
+        fst.d   $f28, $sp, 32
+        fst.d   $f29, $sp, 40
+        fst.d   $f30, $sp, 48
+        fst.d   $f31, $sp, 56
+
+        // Set target to TransitionBlock pointer
+        addi.d  \target, $sp, 128
+.endm
+
 // ------------------------------------------------------------------
 // Macro to generate Redirection Stubs
 //

src/coreclr/pal/inc/unixasmmacrosriscv64.inc

@@ -349,6 +349,51 @@ C_FUNC(\Name):
     EPILOG_STACK_FREE             192
 .endm
 
+// Pushes a full TransitionBlock on the stack including argument registers and
+// floating point argument registers. Used for exception throw helpers where we
+// need to capture the complete register state.
+//
+// Stack layout (from low to high address):
+//   sp+0:     FloatArgumentRegisters (fa0-fa7, 64 bytes)
+//   sp+64:    TransitionBlock start
+//             - CalleeSavedRegisters (fp, ra, s1-s11, tp, gp - 120 bytes)
+//             - padding (8 bytes)
+//             - ArgumentRegisters (a0-a7, 64 bytes)
+//
+// On exit, \target contains the TransitionBlock pointer (sp+160).
+.macro PUSH_COOP_PINVOKE_FRAME_WITH_FLOATS target
+    // Stack: FPCalleeSaved(96) + FloatArgs(64) + CalleeSaved(120) + pad(8) + Args(64) = 352 bytes
+    PROLOG_STACK_ALLOC 352
+    PROLOG_SAVE_REG_PAIR   fp, ra, 160, 1
+
+    // Save callee-saved registers at offset 160 (after FP callee-saved and FloatArgumentRegisters)
+    SAVE_CALLEESAVED_REGISTERS sp, 160
+
+    // Save argument registers (a0-a7) at offset 288
+    SAVE_ARGUMENT_REGISTERS sp, 288
+
+    // Save floating-point argument registers (fa0-fa7) at offset 96
+    SAVE_FLOAT_ARGUMENT_REGISTERS sp, 96
+
+    // Save FP callee-saved registers (fs0-fs11 = f8,f9,f18-f27) at offset 0
+    // RISC-V FP callee-saved: fs0=f8, fs1=f9, fs2-fs11=f18-f27
+    fsd   fs0,  0(sp)   // f8
+    fsd   fs1,  8(sp)   // f9
+    fsd   fs2,  16(sp)  // f18
+    fsd   fs3,  24(sp)  // f19
+    fsd   fs4,  32(sp)  // f20
+    fsd   fs5,  40(sp)  // f21
+    fsd   fs6,  48(sp)  // f22
+    fsd   fs7,  56(sp)  // f23
+    fsd   fs8,  64(sp)  // f24
+    fsd   fs9,  72(sp)  // f25
+    fsd   fs10, 80(sp)  // f26
+    fsd   fs11, 88(sp)  // f27
+
+    // Set target to TransitionBlock pointer
+    addi  \target, sp, 160
+.endm
+
 // ------------------------------------------------------------------
 // Macro to generate Redirection Stubs
 //

src/coreclr/vm/amd64/AsmHelpers.asm

@@ -11,6 +11,9 @@ extern  ProfileLeave:proc
 extern  ProfileTailcall:proc
 extern OnHijackWorker:proc
 extern JIT_RareDisableHelperWorker:proc
+extern IL_Throw_Impl:proc
+extern IL_ThrowExact_Impl:proc
+extern IL_Rethrow_Impl:proc
 ifdef FEATURE_INTERPRETER
 extern ExecuteInterpretedMethod:proc
 extern GetInterpThreadContextWithPossiblyMissingThreadOrCallStub:proc
@@ -1200,4 +1203,51 @@ NESTED_END CallJittedMethodRetI8, _TEXT
 
 endif ; FEATURE_INTERPRETER
 
+;==========================================================================
+; Capture a transition block with register values and call the IL_Throw_Impl
+; implementation written in C.
+;
+; Input state:
+;   RCX = Pointer to exception object
+;==========================================================================
+NESTED_ENTRY IL_Throw, _TEXT
+        PUSH_COOP_PINVOKE_FRAME_WITH_FLOATS rdx
+
+        ; RCX already contains exception object
+        ; RDX contains pointer to TransitionBlock
+        call    IL_Throw_Impl
+        ; Should never return
+        int     3
+NESTED_END IL_Throw, _TEXT
+
+;==========================================================================
+; Capture a transition block with register values and call the IL_ThrowExact_Impl
+; implementation written in C.
+;
+; Input state:
+;   RCX = Pointer to exception object
+;==========================================================================
+NESTED_ENTRY IL_ThrowExact, _TEXT
+        PUSH_COOP_PINVOKE_FRAME_WITH_FLOATS rdx
+
+        ; RCX already contains exception object
+        ; RDX contains pointer to TransitionBlock
+        call    IL_ThrowExact_Impl
+        ; Should never return
+        int     3
+NESTED_END IL_ThrowExact, _TEXT
+
+;==========================================================================
+; Capture a transition block with register values and call the IL_Rethrow_Impl
+; implementation written in C.
+;==========================================================================
+NESTED_ENTRY IL_Rethrow, _TEXT
+        PUSH_COOP_PINVOKE_FRAME_WITH_FLOATS rcx
+
+        ; RCX contains pointer to TransitionBlock
+        call    IL_Rethrow_Impl
+        ; Should never return
+        int     3
+NESTED_END IL_Rethrow, _TEXT
+
         end

src/coreclr/vm/amd64/AsmMacros.inc

@@ -485,5 +485,56 @@ POP_COOP_PINVOKE_FRAME macro
 
         endm
 
+; Pushes a full TransitionBlock on the stack including argument registers and
+; floating point argument registers. Used for exception throw helpers where we
+; need to capture the complete register state including FP callee-saved registers.
+;
+; Stack layout (from high to low address after prologue):
+;   Return address (m_ReturnAddress)
+;   CalleeSavedRegisters (r15, r14, r13, r12, rbp, rbx, rsi, rdi - 64 bytes) <- TransitionBlock starts here
+;   Outgoing argument homes (32 bytes)
+;   FloatArgumentRegisters (xmm0-xmm3, 64 bytes)
+;   FP Callee-saved registers (xmm6-xmm15, 160 bytes)
+;   Padding (8 bytes for 16-byte alignment)
+;   sp points here
+;
+; Stack alignment: After call to IL_Throw (8 bytes) + callee-saved (64 bytes) + alloc (264 bytes) = 336 bytes
+; The XMM save area at offset 8 is 16-byte aligned (336 - 8 = 328, 328 % 16 = 8, but RSP+8 % 16 = 0).
+;
+; On exit, target contains the TransitionBlock pointer (CalleeSavedRegisters).
+PUSH_COOP_PINVOKE_FRAME_WITH_FLOATS macro target
+
+        PUSH_CALLEE_SAVED_REGISTERS
+
+        ; Allocate space for: outgoing args (32) + float args (64) + FP callee-saved (160) + padding (8) = 264 bytes
+        alloc_stack     264
+
+        ; Save argument registers to shadow space area
+        ; Shadow space is at offset 232 (after float args and FP callee-saved and padding) from sp
+        SAVE_ARGUMENT_REGISTERS 232
+
+        ; Save float argument registers at offset 168 (after FP callee-saved and padding)
+        SAVE_FLOAT_ARGUMENT_REGISTERS 168
+
+        ; Save FP callee-saved registers (xmm6-xmm15) at offset 8 (after padding)
+        ; Offset 8 ensures 16-byte alignment for movaps
+        save_xmm128_postrsp xmm6,  8h
+        save_xmm128_postrsp xmm7,  18h
+        save_xmm128_postrsp xmm8,  28h
+        save_xmm128_postrsp xmm9,  38h
+        save_xmm128_postrsp xmm10, 48h
+        save_xmm128_postrsp xmm11, 58h
+        save_xmm128_postrsp xmm12, 68h
+        save_xmm128_postrsp xmm13, 78h
+        save_xmm128_postrsp xmm14, 88h
+        save_xmm128_postrsp xmm15, 98h
+
+        END_PROLOGUE
+
+        ; TransitionBlock pointer points to CalleeSavedRegisters at rsp + 264
+        lea     target, [rsp + 264]
+
+        endm
+
 ;; GC type flags
 GC_ALLOC_FINALIZE               equ 1