[codegen] Use ABI caller saved regs for Pop/PushCallerSaved

The set of registers used for Pop/PushCallerSaved is a superset of the
ABI caller-saved registers. In the past it may have been the case that
these extra registers had to be saved, but at this point
Pop/PushCallerSaved is only used for fast C calls from JS, so we can
rely on the C-compiled functions saving callee-saved registers
correctly, and only save ABI-required registers ourselves.

Change-Id: I2a172bdbb381a1485654e54e3561d695b6672ed0
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3735130
Commit-Queue: Jakob Kummerow <jkummerow@chromium.org>
Reviewed-by: Jakob Kummerow <jkummerow@chromium.org>
Auto-Submit: Leszek Swirski <leszeks@chromium.org>
Cr-Commit-Position: refs/heads/main@{#81467}

src/codegen/ia32/macro-assembler-ia32.cc

@@ -299,21 +299,11 @@ void TurboAssembler::LoadAddress(Register destination,
   mov(destination, Immediate(source));
 }
 
-static constexpr Register saved_regs[] = {eax, ecx, edx};
-
-static constexpr int kNumberOfSavedRegs = sizeof(saved_regs) / sizeof(Register);
-
 int TurboAssembler::RequiredStackSizeForCallerSaved(SaveFPRegsMode fp_mode,
-                                                    Register exclusion1,
-                                                    Register exclusion2,
-                                                    Register exclusion3) const {
+                                                    Register exclusion) const {
   int bytes = 0;
-  for (int i = 0; i < kNumberOfSavedRegs; i++) {
-    Register reg = saved_regs[i];
-    if (reg != exclusion1 && reg != exclusion2 && reg != exclusion3) {
-      bytes += kSystemPointerSize;
-    }
-  }
+  RegList saved_regs = kCallerSaved - exclusion;
+  bytes += kSystemPointerSize * saved_regs.Count();
 
   if (fp_mode == SaveFPRegsMode::kSave) {
     // Count all XMM registers except XMM0.
@@ -323,19 +313,17 @@ int TurboAssembler::RequiredStackSizeForCallerSaved(SaveFPRegsMode fp_mode,
   return bytes;
 }
 
-int TurboAssembler::PushCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
-                                    Register exclusion2, Register exclusion3) {
+int TurboAssembler::PushCallerSaved(SaveFPRegsMode fp_mode,
+                                    Register exclusion) {
   ASM_CODE_COMMENT(this);
   // We don't allow a GC in a write barrier slow path so there is no need to
   // store the registers in any particular way, but we do have to store and
   // restore them.
   int bytes = 0;
-  for (int i = 0; i < kNumberOfSavedRegs; i++) {
-    Register reg = saved_regs[i];
-    if (reg != exclusion1 && reg != exclusion2 && reg != exclusion3) {
-      push(reg);
-      bytes += kSystemPointerSize;
-    }
+  RegList saved_regs = kCallerSaved - exclusion;
+  for (Register reg : saved_regs) {
+    push(reg);
+    bytes += kSystemPointerSize;
   }
 
   if (fp_mode == SaveFPRegsMode::kSave) {
@@ -356,8 +344,7 @@ int TurboAssembler::PushCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
   return bytes;
 }
 
-int TurboAssembler::PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
-                                   Register exclusion2, Register exclusion3) {
+int TurboAssembler::PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion) {
   ASM_CODE_COMMENT(this);
   int bytes = 0;
   if (fp_mode == SaveFPRegsMode::kSave) {
@@ -375,12 +362,10 @@ int TurboAssembler::PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
     bytes += delta;
   }
 
-  for (int i = kNumberOfSavedRegs - 1; i >= 0; i--) {
-    Register reg = saved_regs[i];
-    if (reg != exclusion1 && reg != exclusion2 && reg != exclusion3) {
-      pop(reg);
-      bytes += kSystemPointerSize;
-    }
+  RegList saved_regs = kCallerSaved - exclusion;
+  for (Register reg : base::Reversed(saved_regs)) {
+    pop(reg);
+    bytes += kSystemPointerSize;
   }
 
   return bytes;

src/codegen/ia32/macro-assembler-ia32.h

@@ -375,25 +375,19 @@ class V8_EXPORT_PRIVATE TurboAssembler
   // Calculate how much stack space (in bytes) are required to store caller
   // registers excluding those specified in the arguments.
   int RequiredStackSizeForCallerSaved(SaveFPRegsMode fp_mode,
-                                      Register exclusion1 = no_reg,
-                                      Register exclusion2 = no_reg,
-                                      Register exclusion3 = no_reg) const;
+                                      Register exclusion = no_reg) const;
 
   // PushCallerSaved and PopCallerSaved do not arrange the registers in any
   // particular order so they are not useful for calls that can cause a GC.
-  // The caller can exclude up to 3 registers that do not need to be saved and
+  // The caller can exclude a register that does not need to be saved and
   // restored.
 
   // Push caller saved registers on the stack, and return the number of bytes
   // stack pointer is adjusted.
-  int PushCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1 = no_reg,
-                      Register exclusion2 = no_reg,
-                      Register exclusion3 = no_reg);
+  int PushCallerSaved(SaveFPRegsMode fp_mode, Register exclusion = no_reg);
   // Restore caller saved registers from the stack, and return the number of
   // bytes stack pointer is adjusted.
-  int PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1 = no_reg,
-                     Register exclusion2 = no_reg,
-                     Register exclusion3 = no_reg);
+  int PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion = no_reg);
 
   // Compute the start of the generated instruction stream from the current PC.
   // This is an alternative to embedding the {CodeObject} handle as a reference.

src/codegen/ia32/reglist-ia32.h

@@ -22,6 +22,9 @@ constexpr RegList kJSCallerSaved = {
     ebx,   // used as caller-saved register in JavaScript code
     edi};  // callee function
 
+// Caller-saved registers according to the x86 ABI
+constexpr RegList kCallerSaved = {eax, ecx, edx};
+
 constexpr int kNumJSCallerSaved = 5;
 
 }  // namespace internal

src/codegen/x64/macro-assembler-x64.cc

@@ -776,22 +776,11 @@ void MacroAssembler::JumpToExternalReference(const ExternalReference& ext,
   Jump(code, RelocInfo::CODE_TARGET);
 }
 
-static constexpr Register saved_regs[] = {rax, rcx, rdx, rbx, rbp, rsi,
-                                          rdi, r8,  r9,  r10, r11};
-
-static constexpr int kNumberOfSavedRegs = sizeof(saved_regs) / sizeof(Register);
-
 int TurboAssembler::RequiredStackSizeForCallerSaved(SaveFPRegsMode fp_mode,
-                                                    Register exclusion1,
-                                                    Register exclusion2,
-                                                    Register exclusion3) const {
+                                                    Register exclusion) const {
   int bytes = 0;
-  for (int i = 0; i < kNumberOfSavedRegs; i++) {
-    Register reg = saved_regs[i];
-    if (reg != exclusion1 && reg != exclusion2 && reg != exclusion3) {
-      bytes += kSystemPointerSize;
-    }
-  }
+  RegList saved_regs = kCallerSaved - exclusion;
+  bytes += kSystemPointerSize * saved_regs.Count();
 
   // R12 to r15 are callee save on all platforms.
   if (fp_mode == SaveFPRegsMode::kSave) {
@@ -801,19 +790,17 @@ int TurboAssembler::RequiredStackSizeForCallerSaved(SaveFPRegsMode fp_mode,
   return bytes;
 }
 
-int TurboAssembler::PushCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
-                                    Register exclusion2, Register exclusion3) {
+int TurboAssembler::PushCallerSaved(SaveFPRegsMode fp_mode,
+                                    Register exclusion) {
   ASM_CODE_COMMENT(this);
   // We don't allow a GC in a write barrier slow path so there is no need to
   // store the registers in any particular way, but we do have to store and
   // restore them.
   int bytes = 0;
-  for (int i = 0; i < kNumberOfSavedRegs; i++) {
-    Register reg = saved_regs[i];
-    if (reg != exclusion1 && reg != exclusion2 && reg != exclusion3) {
-      pushq(reg);
-      bytes += kSystemPointerSize;
-    }
+  RegList saved_regs = kCallerSaved - exclusion;
+  for (Register reg : saved_regs) {
+    pushq(reg);
+    bytes += kSystemPointerSize;
   }
 
   // R12 to r15 are callee save on all platforms.
@@ -834,8 +821,7 @@ int TurboAssembler::PushCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
   return bytes;
 }
 
-int TurboAssembler::PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
-                                   Register exclusion2, Register exclusion3) {
+int TurboAssembler::PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion) {
   ASM_CODE_COMMENT(this);
   int bytes = 0;
   if (fp_mode == SaveFPRegsMode::kSave) {
@@ -852,12 +838,10 @@ int TurboAssembler::PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
     bytes += delta;
   }
 
-  for (int i = kNumberOfSavedRegs - 1; i >= 0; i--) {
-    Register reg = saved_regs[i];
-    if (reg != exclusion1 && reg != exclusion2 && reg != exclusion3) {
-      popq(reg);
-      bytes += kSystemPointerSize;
-    }
+  RegList saved_regs = kCallerSaved - exclusion;
+  for (Register reg : base::Reversed(saved_regs)) {
+    popq(reg);
+    bytes += kSystemPointerSize;
   }
 
   return bytes;

src/codegen/x64/macro-assembler-x64.h

@@ -530,25 +530,19 @@ class V8_EXPORT_PRIVATE TurboAssembler
   // Calculate how much stack space (in bytes) are required to store caller
   // registers excluding those specified in the arguments.
   int RequiredStackSizeForCallerSaved(SaveFPRegsMode fp_mode,
-                                      Register exclusion1 = no_reg,
-                                      Register exclusion2 = no_reg,
-                                      Register exclusion3 = no_reg) const;
+                                      Register exclusion = no_reg) const;
 
   // PushCallerSaved and PopCallerSaved do not arrange the registers in any
   // particular order so they are not useful for calls that can cause a GC.
-  // The caller can exclude up to 3 registers that do not need to be saved and
+  // The caller can exclude a register that does not need to be saved and
   // restored.
 
   // Push caller saved registers on the stack, and return the number of bytes
   // stack pointer is adjusted.
-  int PushCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1 = no_reg,
-                      Register exclusion2 = no_reg,
-                      Register exclusion3 = no_reg);
+  int PushCallerSaved(SaveFPRegsMode fp_mode, Register exclusion = no_reg);
   // Restore caller saved registers from the stack, and return the number of
   // bytes stack pointer is adjusted.
-  int PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1 = no_reg,
-                     Register exclusion2 = no_reg,
-                     Register exclusion3 = no_reg);
+  int PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion = no_reg);
 
   // Compute the start of the generated instruction stream from the current PC.
   // This is an alternative to embedding the {CodeObject} handle as a reference.