​MIPS: Improve Cvt_d_uw on mips32.

src/mips/macro-assembler-mips.cc

Index: src/mips/macro-assembler-mips.cc
diff --git a/src/mips/macro-assembler-mips.cc b/src/mips/macro-assembler-mips.cc
index 272f8147b94185aac05ed5dd75d18940e2fb7e03..5a239a78253bcff353834a0f528040c0fc37864f 100644
--- a/src/mips/macro-assembler-mips.cc
+++ b/src/mips/macro-assembler-mips.cc
@@ -1270,46 +1270,53 @@ void MacroAssembler::Ins(Register rt,
 void MacroAssembler::Cvt_d_uw(FPURegister fd,
                               FPURegister fs,
                               FPURegister scratch) {
-  // Move the data from fs to t8.
-  mfc1(t8, fs);
-  Cvt_d_uw(fd, t8, scratch);
+  if (IsFp64Mode()) {
+    // On FP64Mode we can simple do convertion from long.

[paul.l..., 2015/11/19 01:23:38]

Nits: fix spelling and word order in this sentence. Start with: In. Then 'do
simple conversion'

[Alan Li, 2015/11/20 17:57:55]

Done.

+    cvt_d_l(fd, fs);
+  } else {
+    mfc1(t8, fs);
+    Cvt_d_uw(fd, t8, scratch, fs);
+  }
 }
 
 
-void MacroAssembler::Cvt_d_uw(FPURegister fd,
-                              Register rs,
-                              FPURegister scratch) {
+void MacroAssembler::Cvt_d_uw(FPURegister fd, Register rs, FPURegister scratch,
+                              FPURegister fs) {
   // Convert rs to a FP value in fd (and fd + 1).

[paul.l..., 2015/11/19 01:23:38]

Remove the '(and fd + 1); from the comment. While that is true for fp32 mode, it
is well-known, and it is not true for fp64 mode. 

[Alan Li, 2015/11/20 17:57:55]

Done.

-  // We do this by converting rs minus the MSB to avoid sign conversion,
-  // then adding 2^31 to the result (if needed).
-
   DCHECK(!fd.is(scratch));
   DCHECK(!rs.is(t9));
   DCHECK(!rs.is(at));
 
-  // Save rs's MSB to t9.
-  Ext(t9, rs, 31, 1);
-  // Remove rs's MSB.
-  Ext(at, rs, 0, 31);
-  // Move the result to fd.
-  mtc1(at, fd);
-
-  // Convert fd to a real FP value.
-  cvt_d_w(fd, fd);
-
-  Label conversion_done;
-
-  // If rs's MSB was 0, it's done.
-  // Otherwise we need to add that to the FP register.
-  Branch(&conversion_done, eq, t9, Operand(zero_reg));
+  Label msb_clear, conversion_done;
+  // For a value which is < 2^31, regard it as a signed positve word.
+  Branch(&msb_clear, ge, rs, Operand(zero_reg), USE_DELAY_SLOT);
+  lui(at, 0x41F0);  // FP value: 2^32, in delay slot.
 
-  // Load 2^31 into f20 as its float representation.
-  li(at, 0x41E00000);
+  // For unsigned word A which is greater than 2^31, consider this fact:
+  // A = 2^32 + reinterpret_cast<int32>(A)
+  // Thus we can first regard A as signed word and cast it into double,
+  // then add 2^32 on the result to complete the casting.

[paul.l..., 2015/11/19 01:23:38]

Comment is too wordy, and strangely worded, we don't need 'consider this fact'
or 'thus'. How about:

For unsigned inputs > 2^31, we convert to double as a signed int32, then add
2^32 to move it back to unsigned value in range 2^31 .. 2^32-1.

[Alan Li, 2015/11/20 17:57:55]

Done.

   mtc1(zero_reg, scratch);
   Mthc1(at, scratch);
-  // Add it to fd.
+
+  if (fs.is(no_freg)) {
+    mtc1(rs, fd);
+    cvt_d_w(fd, fd);
+  } else {
+    cvt_d_w(fd, fs);
+  }
+
+  Branch(USE_DELAY_SLOT, &conversion_done);
   add_d(fd, fd, scratch);
 
+  bind(&msb_clear);
+  if (fs.is(no_freg)) {
+    mtc1(rs, fd);
+    cvt_d_w(fd, fd);
+  } else {
+    cvt_d_w(fd, fs);
+  }
+
   bind(&conversion_done);
 }