tor-browser

commit 56c5cc20609ed6ff5c4c3af310755770b42ee6fc
parent da494342cf7c537190eeea74b7a82ec37de89468
Author: André Bargull <andre.bargull@gmail.com>
Date:   Mon, 20 Oct 2025 08:20:22 +0000

Bug 1993951 - Part 1: Implement MacroAssembler::copySign{Float32,Double} for ARM. r=spidermonkey-reviewers,iain

GCC emits this instruction sequence for `std::copysign`. It needs fewer instructions
and registers, which helps to move `LCopySign` into the shared code generator.

Differential Revision: https://phabricator.services.mozilla.com/D268366

Diffstat:
M
js/src/jit/MacroAssembler.h
 | 
2
+-
M
js/src/jit/arm/CodeGenerator-arm.cpp
 | 
44
++------------------------------------------
M
js/src/jit/arm/Lowering-arm.cpp
 | 
3
---
M
js/src/jit/arm/MacroAssembler-arm.cpp
 | 
28
+++++++++++++++++++++++++++-

4 files changed, 30 insertions(+), 47 deletions(-)
diff --git a/js/src/jit/MacroAssembler.h b/js/src/jit/MacroAssembler.h
@@ -1328,7 +1328,7 @@ class MacroAssembler : public MacroAssemblerSpecific {
   void copySignDouble(FloatRegister lhs, FloatRegister rhs,
                       FloatRegister output) PER_SHARED_ARCH;
   void copySignFloat32(FloatRegister lhs, FloatRegister rhs,
-                       FloatRegister output) DEFINED_ON(x86_shared, arm64);
+                       FloatRegister output) DEFINED_ON(arm, arm64, x86_shared);
 
   // Returns a random double in range [0, 1) in |dest|. The |rng| register must
   // hold a pointer to a mozilla::non_crypto::XorShift128PlusRNG.
diff --git a/js/src/jit/arm/CodeGenerator-arm.cpp b/js/src/jit/arm/CodeGenerator-arm.cpp
@@ -2366,24 +2366,7 @@ void CodeGenerator::visitCopySignF(LCopySignF* ins) {
   FloatRegister rhs = ToFloatRegister(ins->rhs());
   FloatRegister output = ToFloatRegister(ins->output());
 
-  Register lhsi = ToRegister(ins->temp0());
-  Register rhsi = ToRegister(ins->temp1());
-
-  masm.ma_vxfer(lhs, lhsi);
-  masm.ma_vxfer(rhs, rhsi);
-
-  ScratchRegisterScope scratch(masm);
-
-  // Clear lhs's sign.
-  masm.ma_and(Imm32(INT32_MAX), lhsi, lhsi, scratch);
-
-  // Keep rhs's sign.
-  masm.ma_and(Imm32(INT32_MIN), rhsi, rhsi, scratch);
-
-  // Combine.
-  masm.ma_orr(lhsi, rhsi, rhsi);
-
-  masm.ma_vxfer(rhsi, output);
+  masm.copySignFloat32(lhs, rhs, output);
 }
 
 void CodeGenerator::visitCopySignD(LCopySignD* ins) {
@@ -2391,30 +2374,7 @@ void CodeGenerator::visitCopySignD(LCopySignD* ins) {
   FloatRegister rhs = ToFloatRegister(ins->rhs());
   FloatRegister output = ToFloatRegister(ins->output());
 
-  Register lhsi = ToRegister(ins->temp0());
-  Register rhsi = ToRegister(ins->temp1());
-
-  // Manipulate high words of double inputs.
-  masm.as_vxfer(lhsi, InvalidReg, lhs, Assembler::FloatToCore,
-                Assembler::Always, 1);
-  masm.as_vxfer(rhsi, InvalidReg, rhs, Assembler::FloatToCore,
-                Assembler::Always, 1);
-
-  ScratchRegisterScope scratch(masm);
-
-  // Clear lhs's sign.
-  masm.ma_and(Imm32(INT32_MAX), lhsi, lhsi, scratch);
-
-  // Keep rhs's sign.
-  masm.ma_and(Imm32(INT32_MIN), rhsi, rhsi, scratch);
-
-  // Combine.
-  masm.ma_orr(lhsi, rhsi, rhsi);
-
-  // Reconstruct the output.
-  masm.as_vxfer(lhsi, InvalidReg, lhs, Assembler::FloatToCore,
-                Assembler::Always, 0);
-  masm.ma_vxfer(lhsi, rhsi, output);
+  masm.copySignDouble(lhs, rhs, output);
 }
 
 void CodeGenerator::visitWrapInt64ToInt32(LWrapInt64ToInt32* lir) {
diff --git a/js/src/jit/arm/Lowering-arm.cpp b/js/src/jit/arm/Lowering-arm.cpp
@@ -1145,9 +1145,6 @@ void LIRGenerator::visitCopySign(MCopySign* ins) {
     lir = new (alloc()) LCopySignF();
   }
 
-  lir->setTemp(0, temp());
-  lir->setTemp(1, temp());
-
   lowerForFPU(lir, ins, lhs, rhs);
 }
 
diff --git a/js/src/jit/arm/MacroAssembler-arm.cpp b/js/src/jit/arm/MacroAssembler-arm.cpp
@@ -6188,7 +6188,33 @@ void MacroAssembler::nearbyIntFloat32(RoundingMode mode, FloatRegister src,
 
 void MacroAssembler::copySignDouble(FloatRegister lhs, FloatRegister rhs,
                                     FloatRegister output) {
-  MOZ_CRASH("not supported on this platform");
+  ScratchRegisterScope scratch(*this);
+
+  // Transfer top word of |rhs| into GPR.
+  as_vxfer(scratch, InvalidReg, rhs, Assembler::FloatToCore, Assembler::Always,
+           1);
+
+  // Move |lhs| with sign-bit cleared into |output|.
+  ma_vabs(lhs, output);
+
+  // Negate |output| if sign bit of |rhs| is set.
+  as_cmp(scratch, Imm8(0));
+  ma_vneg(output, output, Assembler::LessThan);
+}
+
+void MacroAssembler::copySignFloat32(FloatRegister lhs, FloatRegister rhs,
+                                     FloatRegister output) {
+  ScratchRegisterScope scratch(*this);
+
+  // Transfer |rhs| into GPR.
+  ma_vxfer(rhs, scratch);
+
+  // Move |lhs| with sign-bit cleared into |output|.
+  ma_vabs_f32(lhs, output);
+
+  // Negate |output| if sign bit of |rhs| is set.
+  as_cmp(scratch, Imm8(0));
+  ma_vneg_f32(output, output, Assembler::LessThan);
 }
 
 void MacroAssembler::shiftIndex32AndAdd(Register indexTemp32, int shift,