<h5>Syntax:</h5>
<pre>
- <result> = load <ty>* <pointer>[, align <alignment>]
- <result> = volatile load <ty>* <pointer>[, align <alignment>]
+ <result> = load <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>]
+ <result> = volatile load <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>]
+ !<index> = !{ i32 1 }
</pre>
<h5>Overview:</h5>
marked as <tt>volatile</tt>, then the optimizer is not allowed to modify the
number or order of execution of this <tt>load</tt> with other
volatile <tt>load</tt> and <tt><a href="#i_store">store</a></tt>
- instructions. </p>
+ instructions.</p>
<p>The optional constant "align" argument specifies the alignment of the
operation (that is, the alignment of the memory address). A value of 0 or an
alignment results in an undefined behavior. Underestimating the alignment may
produce less efficient code. An alignment of 1 is always safe.</p>
+<p>The optional !nontemporal metadata must reference a single metatadata
+ name <index> corresponding to a metadata node with one i32 entry of
+ value 1. The existance of the !nontemporal metatadata on the
+ instruction tells the optimizer and code generator that this load is
+ not expected to be reused in the cache. The code generator may
+ select special instructions to save cache bandwidth, such as the
+ MOVNT intruction on x86.</p>
+
<h5>Semantics:</h5>
<p>The location of memory pointed to is loaded. If the value being loaded is of
scalar type then the number of bytes read does not exceed the minimum number
<h5>Syntax:</h5>
<pre>
- store <ty> <value>, <ty>* <pointer>[, align <alignment>] <i>; yields {void}</i>
- volatile store <ty> <value>, <ty>* <pointer>[, align <alignment>] <i>; yields {void}</i>
+ store <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] <i>; yields {void}</i>
+ volatile store <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] <i>; yields {void}</i>
</pre>
<h5>Overview:</h5>
alignment results in an undefined behavior. Underestimating the alignment may
produce less efficient code. An alignment of 1 is always safe.</p>
+<p>The optional !nontemporal metadata must reference a single metatadata
+ name <index> corresponding to a metadata node with one i32 entry of
+ value 1. The existance of the !nontemporal metatadata on the
+ instruction tells the optimizer and code generator that this load is
+ not expected to be reused in the cache. The code generator may
+ select special instructions to save cache bandwidth, such as the
+ MOVNT intruction on x86.</p>
+
+
<h5>Semantics:</h5>
<p>The contents of memory are updated to contain '<tt><value></tt>' at the
location specified by the '<tt><pointer></tt>' operand. If
def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
+// MOVNT Support
+// Like 'store', but requires the non-temporal bit to be set
+def nontemporalstore : PatFrag<(ops node:$val, node:$ptr),
+ (st node:$val, node:$ptr), [{
+ if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
+ return ST->isNonTemporal();
+ return false;
+}]>;
+
+def alignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
+ (st node:$val, node:$ptr), [{
+ if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
+ return ST->isNonTemporal() && !ST->isTruncatingStore() &&
+ ST->getAddressingMode() == ISD::UNINDEXED &&
+ ST->getAlignment() >= 16;
+ return false;
+}]>;
+
+def unalignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
+ (st node:$val, node:$ptr), [{
+ if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
+ return ST->isNonTemporal() &&
+ ST->getAlignment() < 16;
+ return false;
+}]>;
+
def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
"prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
// Non-temporal stores
-def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
+def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
"movntps\t{$src, $dst|$dst, $src}",
[(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
+let AddedComplexity = 400 in { // Prefer non-temporal versions
+def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
+ "movntps\t{$src, $dst|$dst, $src}",
+ [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
+
+def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
+ "movntdq\t{$src, $dst|$dst, $src}",
+ [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
+
+def : Pat<(alignednontemporalstore (v2i64 VR128:$src), addr:$dst),
+ (MOVNTDQ_64mr VR128:$src, addr:$dst)>;
+
+def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
+ "movnti\t{$src, $dst|$dst, $src}",
+ [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
+ TB, Requires<[HasSSE2]>;
+
+def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
+ "movnti\t{$src, $dst|$dst, $src}",
+ [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
+ TB, Requires<[HasSSE2]>;
+}
+
// Load, store, and memory fence
def SFENCE : PSI<0xAE, MRM7r, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
[(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
// Non-temporal stores
-def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
- "movntpd\t{$src, $dst|$dst, $src}",
- [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
-def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
- "movntdq\t{$src, $dst|$dst, $src}",
- [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
-def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
+def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
+ "movntpd\t{$src, $dst|$dst, $src}",
+ [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
+def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
+ "movntdq\t{$src, $dst|$dst, $src}",
+ [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
+def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
"movnti\t{$src, $dst|$dst, $src}",
[(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
TB, Requires<[HasSSE2]>;
+let AddedComplexity = 400 in { // Prefer non-temporal versions
+def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
+ "movntpd\t{$src, $dst|$dst, $src}",
+ [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
+
+def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
+ "movntdq\t{$src, $dst|$dst, $src}",
+ [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
+
+def : Pat<(alignednontemporalstore (v4i32 VR128:$src), addr:$dst),
+ (MOVNTDQmr VR128:$src, addr:$dst)>;
+}
+
// Flush cache
def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
"clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,