Fixing stvl/stvr.

src/alloy/frontend/ppc/ppc_emit_altivec.cc

@@ -265,18 +265,16 @@ int InstrEmit_stvlx_(PPCHIRBuilder& f, InstrData& i, uint32_t vd, uint32_t ra,
   //       we could optimize this to prevent the other load/mask, in that case.
   Value* ea = CalculateEA_0(f, ra, rb);
   Value* eb = f.And(f.Truncate(ea, INT8_TYPE), f.LoadConstant((int8_t)0xF));
-  Value* new_value = f.LoadVR(vd);
   // ea &= ~0xF
   ea = f.And(ea, f.LoadConstant(~0xFull));
+  // v = (old & ~mask) | ((new >> eb) & mask)
+  Value* new_value = f.Permute(f.LoadVectorShr(eb), f.LoadZero(VEC128_TYPE),
+                               f.LoadVR(vd), INT8_TYPE);
   Value* old_value = f.ByteSwap(f.Load(ea, VEC128_TYPE));
-  // v = (new >> eb) | (old & (ONE << (16 - eb)))
+  // mask = FFFF... >> eb
-  Value* v = f.Permute(f.LoadVectorShr(eb), f.LoadZero(VEC128_TYPE), new_value,
+  Value* mask = f.Permute(f.LoadVectorShr(eb), f.LoadZero(VEC128_TYPE),
-                       INT8_TYPE);
+                          f.Not(f.LoadZero(VEC128_TYPE)), INT8_TYPE);
-  v = f.Or(
+  Value* v = f.Or(f.And(old_value, f.Not(mask)), f.And(new_value, mask));
-      v, f.And(old_value,
-               f.Permute(f.LoadVectorShl(f.Sub(f.LoadConstant((int8_t)16), eb)),
-                         f.Not(f.LoadZero(VEC128_TYPE)),
-                         f.LoadZero(VEC128_TYPE), INT8_TYPE)));
   // ea &= ~0xF (handled above)
   f.Store(ea, f.ByteSwap(v));
   return 0;
@@ -301,16 +299,16 @@ int InstrEmit_stvrx_(PPCHIRBuilder& f, InstrData& i, uint32_t vd, uint32_t ra,
   //       we could optimize this to prevent the other load/mask, in that case.
   Value* ea = CalculateEA_0(f, ra, rb);
   Value* eb = f.And(f.Truncate(ea, INT8_TYPE), f.LoadConstant((int8_t)0xF));
-  Value* new_value = f.LoadVR(vd);
   // ea &= ~0xF
   ea = f.And(ea, f.LoadConstant(~0xFull));
+  // v = (old & ~mask) | ((new << eb) & mask)
+  Value* new_value = f.Permute(f.LoadVectorShr(eb), f.LoadVR(vd),
+                               f.LoadZero(VEC128_TYPE), INT8_TYPE);
   Value* old_value = f.ByteSwap(f.Load(ea, VEC128_TYPE));
-  // v = (new << (16 - eb)) | (old & (ONE >> eb))
+  // mask = ~FFFF... >> eb
-  Value* v = f.Permute(f.LoadVectorShl(f.Sub(f.LoadConstant((int8_t)16), eb)),
+  Value* mask = f.Permute(f.LoadVectorShr(eb), f.Not(f.LoadZero(VEC128_TYPE)),
-                       new_value, f.LoadZero(VEC128_TYPE), INT8_TYPE);
+                          f.LoadZero(VEC128_TYPE), INT8_TYPE);
-  v = f.Or(v, f.And(old_value,
+  Value* v = f.Or(f.And(old_value, f.Not(mask)), f.And(new_value, mask));
-                    f.Permute(f.LoadVectorShr(eb), f.LoadZero(VEC128_TYPE),
-                              f.Not(f.LoadZero(VEC128_TYPE)), INT8_TYPE)));
   // ea &= ~0xF (handled above)
   f.Store(ea, f.ByteSwap(v));
   return 0;

src/alloy/frontend/ppc/test/bin/instr_stvl.dis

@@ -7,3 +7,7 @@ Disassembly of section .text:
 0000000000100000 <test_stvl_1>:
   100000:	7c 64 05 0e 	stvlx   v3,r4,r0
   100004:	4e 80 00 20 	blr
+
+0000000000100008 <test_stvl_2>:
+  100008:	7c 64 05 0e 	stvlx   v3,r4,r0
+  10000c:	4e 80 00 20 	blr

src/alloy/frontend/ppc/test/bin/instr_stvl.map

@@ -1 +1,2 @@
 0000000000000000 t test_stvl_1
+0000000000000008 t test_stvl_2

src/alloy/frontend/ppc/test/bin/instr_stvr.dis

@@ -7,3 +7,7 @@ Disassembly of section .text:
 0000000000100000 <test_stvr_1>:
   100000:	7c 64 2d 4e 	stvrx   v3,r4,r5
   100004:	4e 80 00 20 	blr
+
+0000000000100008 <test_stvr_2>:
+  100008:	7c 64 05 4e 	stvrx   v3,r4,r0
+  10000c:	4e 80 00 20 	blr

src/alloy/frontend/ppc/test/bin/instr_stvr.map

@@ -1 +1,2 @@
 0000000000000000 t test_stvr_1
+0000000000000008 t test_stvr_2

src/alloy/frontend/ppc/test/instr_stvl.s

@@ -7,3 +7,13 @@ test_stvl_1:
   #_ REGISTER_OUT r4 0x1040
   #_ REGISTER_OUT v3 [BE74FCBD, BD912ABA, BF317BBB, BF2D135F]
   #_ MEMORY_OUT 00001040 BE74FCBD BD912ABA BF317BBB BF2D135F
+
+test_stvl_2:
+  #_ MEMORY_IN 00001040 00010203 04050607 08090A0B 0C0D0E0F
+  #_ REGISTER_IN r4 0x1044
+  #_ REGISTER_IN v3 [F0F1F2F3, F4F5F6F7, F8F9FAFB, FCFDFEFF]
+  stvlx v3, r4, r0
+  blr
+  #_ REGISTER_OUT r4 0x1044
+  #_ REGISTER_OUT v3 [F0F1F2F3, F4F5F6F7, F8F9FAFB, FCFDFEFF]
+  #_ MEMORY_OUT 00001040 00010203 F0F1F2F3 F4F5F6F7 F8F9FAFB

src/alloy/frontend/ppc/test/instr_stvr.s

@@ -11,3 +11,13 @@ test_stvr_1:
   #_ REGISTER_OUT v3 [BE74FCBD, BD912ABA, BF317BBB, BF2D135F]
   #_ MEMORY_OUT 00001040 BE74FCBD BD912ABA BF317BBB BF2D135F
   #_ MEMORY_OUT 00001050 00000000 00000000 00000000 00000000
+
+test_stvr_2:
+  #_ MEMORY_IN 00001040 00010203 04050607 08090A0B 0C0D0E0F
+  #_ REGISTER_IN r4 0x1044
+  #_ REGISTER_IN v3 [F0F1F2F3, F4F5F6F7, F8F9FAFB, FCFDFEFF]
+  stvrx v3, r4, r0
+  blr
+  #_ REGISTER_OUT r4 0x1044
+  #_ REGISTER_OUT v3 [F0F1F2F3, F4F5F6F7, F8F9FAFB, FCFDFEFF]
+  #_ MEMORY_OUT 00001040 FCFDFEFF 04050607 08090A0B 0C0D0E0F