ARM: Do not set FPSCR when converting to clamped uint8

src/arm/macro-assembler-arm.cc

Index: src/arm/macro-assembler-arm.cc
diff --git a/src/arm/macro-assembler-arm.cc b/src/arm/macro-assembler-arm.cc
index 1633adc1331e203edbacf85a6419a81b02e13653..f8599c6b5637d7faa99bd699aea40811c87171b4 100644
--- a/src/arm/macro-assembler-arm.cc
+++ b/src/arm/macro-assembler-arm.cc
@@ -796,6 +796,10 @@ void MacroAssembler::VFPEnsureFPSCRState(Register scratch) {
   // If needed, restore wanted bits of FPSCR.
   Label fpscr_done;
   vmrs(scratch);
+  if (emit_debug_code()) {
+    tst(scratch, Operand(kVFPRoundingModeMask));
+    Assert(eq, kDefaultRoundingModeNotSet);
+  }
   tst(scratch, Operand(kVFPDefaultNaNModeControlBit));
   b(ne, &fpscr_done);
   orr(scratch, scratch, Operand(kVFPDefaultNaNModeControlBit));
@@ -3800,36 +3804,24 @@ void MacroAssembler::ClampUint8(Register output_reg, Register input_reg) {
 void MacroAssembler::ClampDoubleToUint8(Register result_reg,
                                         DwVfpRegister input_reg,
                                         LowDwVfpRegister double_scratch) {
-  Label above_zero;
   Label done;
-  Label in_bounds;
-
-  VFPCompareAndSetFlags(input_reg, 0.0);
-  b(gt, &above_zero);
 
-  // Double value is less than zero, NaN or Inf, return 0.
-  mov(result_reg, Operand::Zero());
-  b(al, &done);
-
-  // Double value is >= 255, return 255.
-  bind(&above_zero);
+  // Handle inputs >= 255 (including +infinity).
   Vmov(double_scratch, 255.0, result_reg);
-  VFPCompareAndSetFlags(input_reg, double_scratch);
-  b(le, &in_bounds);
   mov(result_reg, Operand(255));
-  b(al, &done);
-
-  // In 0-255 range, round and truncate.
-  bind(&in_bounds);
-  // Save FPSCR.
-  vmrs(ip);
-  // Set rounding mode to round to the nearest integer by clearing bits[23:22].
-  bic(result_reg, ip, Operand(kVFPRoundingModeMask));
-  vmsr(result_reg);
-  vcvt_s32_f64(double_scratch.low(), input_reg, kFPSCRRounding);
+  VFPCompareAndSetFlags(input_reg, double_scratch);
+  b(ge, &done);
+
+  if (emit_debug_code()) {

[jbramley, 2014/04/09 14:18:26]

I'd be tempted to omit this, since everything else depends on it too and we
(now) check it in VFPEnsureFPSCRState, but I won't object if you insist.

[oetuaho-nv, 2014/04/09 14:24:32]

Yes, I thought about leaving this out too, but this is the only place in the
code where the vcvt variant with kFPSCRRounding set seems to be called, so it
doesn't seem too out-of-place to double-check. But if there's a strong opposite
opinion from someone we can surely throw this out.

[rmcilroy, 2014/04/09 15:53:30]

This may be the only place vcvt uses the fpscr's rounding mode, but all the
other FP arithmetic operations depend on fpscr's rounding mode being correct, so
I don't think it's worth adding this check here (but adding it to
VFPEnsureFPSCRState seems like a good move).

+    vmrs(result_reg);
+    tst(result_reg, Operand(kVFPRoundingModeMask));
+    Assert(eq, kDefaultRoundingModeNotSet);
+  }
+  // vcvt_u32_f64 agrees with what ClampDoubleToUint8 should do as long as
+  // value in input_reg < 255.

[rmcilroy, 2014/04/09 15:53:30]

nit - "For inputs < 255 (including negative) vcvt_u32_f64 with round-to-nearest
rounding mode will provide the correct."

+  vcvt_u32_f64(double_scratch.low(), input_reg, kFPSCRRounding);
   vmov(result_reg, double_scratch.low());
-  // Restore FPSCR.
-  vmsr(ip);
+
   bind(&done);
 }