Bug 385207

Summary: PPC64, generate_store_FPRF() generates too many Iops
Product: [Developer tools] valgrind Reporter: Carl Love <cel>
Component: vexAssignee: Julian Seward <jseward>
Status: CLOSED FIXED    
Severity: normal    
Priority: NOR    
Version First Reported In: 3.14 SVN   
Target Milestone: ---   
Platform: Other   
OS: Linux   
Latest Commit: Version Fixed/Implemented In:
Sentry Crash Report:
Attachments: replace body of generate_store_FPRF with C helper function.

Description Carl Love 2017-09-29 18:00:00 UTC 
A number of the power 9 instructions call generate_store_FPRF() to calculate
the floating point condition code value.  The function generates a lot of Iops.
If there are about 6 or instructions in the basic block that use this function, it exhausts the temporary memory.
Comment 1 Carl Love 2017-09-29 18:01:42 UTC 
Created attachment 108093 [details]
replace body of generate_store_FPRF with C helper  function.

replace the generate_store_FPRF function body with a C Helper function.
Comment 2 Julian Seward 2017-10-03 10:04:22 UTC 
+#define V128_FRACTION_MASK 0x0000FFFFFFFFFFFFULL  /* upper 64-bit fractional mask */
+
+ULong generate_C_FPCC_helper( ULong size, ULong src_hi, ULong src ) {
+   UInt NaN, inf, zero, norm, dnorm, pos;
+   UInt bit0, bit1, bit2, bit3;

* nit: { on its own line, in the house style

* Please rename 'size' here to 'irType', because that's what it really is



+   if ( size == Ity_I16 ) {
+      frac_part = I16_FRACTION_MASK & src;
+      exp_mask = I16_EXP_MASK;
...
+     exp_mask = V128_EXP_MASK;
+     exp_part = exp_mask & src_hi;
+     sign_bit = src_hi >> 63;
+   }

Add

  else {
    vassert(0);
  }

so as to cause the system to fail if any other value is passed for irType.


OK to land with the above points fixed.
Comment 3 Carl Love 2017-10-03 17:27:13 UTC 
patch committed:  commit acdeb75d2a58f4f3910ddaf9b2bc2ec74378fa3a