For whatever the reason, x86 CallingConv::Fast (i.e. fastcc) was not passing scalar arguments in registers. This patch defines a new fastcc CC which is slightly different from the FastCall CC. In addition to passing integer arguments in ECX and EDX, it also specify doubles are passed in 8-byte slots which are 8-byte aligned (instead of 4-byte aligned). This avoids a potential performance hazard where doubles span cacheline boundaries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55807 91177308-0d34-0410-b5e6-96231b3b80d8

7 files changed, 53 insertions, 16 deletions

diff --git a/lib/Target/X86/X86CallingConv.td b/lib/Target/X86/X86CallingConv.td
index 32f53ab..549eb9b 100644
--- a/lib/Target/X86/X86CallingConv.td
+++ b/lib/Target/X86/X86CallingConv.td
@@ -341,3 +341,20 @@ def CC_X86_32_FastCall : CallingConv<[
   // Otherwise, same as everything else.
   CCDelegateTo<CC_X86_32_Common>
 ]>;
+
+def CC_X86_32_FastCC : CallingConv<[
+  // Promote i8/i16 arguments to i32.
+  CCIfType<[i8, i16], CCPromoteToType<i32>>,
+
+  // The 'nest' parameter, if any, is passed in EAX.
+  CCIfNest<CCAssignToReg<[EAX]>>,
+
+  // The first 2 integer arguments are passed in ECX/EDX
+  CCIfType<[i32], CCAssignToReg<[ECX, EDX]>>,
+
+  // Doubles get 8-byte slots that are 8-byte aligned.
+  CCIfType<[f64], CCAssignToStack<8, 8>>,
+
+  // Otherwise, same as everything else.
+  CCDelegateTo<CC_X86_32_Common>
+]>;
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index cd01c64..1cf1b41 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1100,6 +1100,8 @@ CCAssignFn *X86TargetLowering::CCAssignFnForNode(SDValue Op) const {
     return CC_X86_32_FastCall;
   else if (CC == CallingConv::Fast && PerformTailCallOpt)
     return CC_X86_32_TailCall;
+  else if (CC == CallingConv::Fast)
+    return CC_X86_32_FastCC;
   else
     return CC_X86_32_C;
 }
diff --git a/test/CodeGen/X86/2007-08-13-SpillerReuse.ll b/test/CodeGen/X86/2007-08-13-SpillerReuse.ll
index 8cc235b..edcb823 100644
--- a/test/CodeGen/X86/2007-08-13-SpillerReuse.ll
+++ b/test/CodeGen/X86/2007-08-13-SpillerReuse.ll
@@ -6,9 +6,9 @@
 @rtx_format = external global [116 x i8*]		; <[116 x i8*]*> [#uses=1]
 @rtx_length = external global [117 x i32]		; <[117 x i32]*> [#uses=1]
 
-declare fastcc %struct.rtx_def* @fixup_memory_subreg(%struct.rtx_def*, %struct.rtx_def*, i32)
+declare %struct.rtx_def* @fixup_memory_subreg(%struct.rtx_def*, %struct.rtx_def*, i32)
 
-define fastcc %struct.rtx_def* @walk_fixup_memory_subreg(%struct.rtx_def* %x, %struct.rtx_def* %insn) {
+define %struct.rtx_def* @walk_fixup_memory_subreg(%struct.rtx_def* %x, %struct.rtx_def* %insn) {
 entry:
 	%tmp2 = icmp eq %struct.rtx_def* %x, null		; <i1> [#uses=1]
 	br i1 %tmp2, label %UnifiedReturnBlock, label %cond_next
@@ -30,7 +30,7 @@ cond_true13:		; preds = %cond_next
 	br i1 %tmp22, label %cond_true25, label %cond_next32
 
 cond_true25:		; preds = %cond_true13
-	%tmp29 = tail call fastcc %struct.rtx_def* @fixup_memory_subreg( %struct.rtx_def* %x, %struct.rtx_def* %insn, i32 1 )		; <%struct.rtx_def*> [#uses=1]
+	%tmp29 = tail call %struct.rtx_def* @fixup_memory_subreg( %struct.rtx_def* %x, %struct.rtx_def* %insn, i32 1 )		; <%struct.rtx_def*> [#uses=1]
 	ret %struct.rtx_def* %tmp29
 
 cond_next32:		; preds = %cond_true13, %cond_next
@@ -56,7 +56,7 @@ cond_true47:		; preds = %bb
 	%tmp52 = getelementptr %struct.rtx_def* %x, i32 0, i32 3, i32 %i.01.0		; <%struct..0anon*> [#uses=1]
 	%tmp5354 = bitcast %struct..0anon* %tmp52 to %struct.rtx_def**		; <%struct.rtx_def**> [#uses=1]
 	%tmp55 = load %struct.rtx_def** %tmp5354		; <%struct.rtx_def*> [#uses=1]
-	%tmp58 = tail call fastcc %struct.rtx_def* @walk_fixup_memory_subreg( %struct.rtx_def* %tmp55, %struct.rtx_def* %insn )		; <%struct.rtx_def*> [#uses=1]
+	%tmp58 = tail call  %struct.rtx_def* @walk_fixup_memory_subreg( %struct.rtx_def* %tmp55, %struct.rtx_def* %insn )		; <%struct.rtx_def*> [#uses=1]
 	%tmp62 = getelementptr %struct.rtx_def* %x, i32 0, i32 3, i32 %i.01.0, i32 0		; <i32*> [#uses=1]
 	%tmp58.c = ptrtoint %struct.rtx_def* %tmp58 to i32		; <i32> [#uses=1]
 	store i32 %tmp58.c, i32* %tmp62
@@ -79,7 +79,7 @@ bb73:		; preds = %bb73, %bb105.preheader
 	%tmp92 = getelementptr %struct.rtvec_def* %tmp81, i32 0, i32 1, i32 %j.019		; <%struct..0anon*> [#uses=1]
 	%tmp9394 = bitcast %struct..0anon* %tmp92 to %struct.rtx_def**		; <%struct.rtx_def**> [#uses=1]
 	%tmp95 = load %struct.rtx_def** %tmp9394		; <%struct.rtx_def*> [#uses=1]
-	%tmp98 = tail call fastcc %struct.rtx_def* @walk_fixup_memory_subreg( %struct.rtx_def* %tmp95, %struct.rtx_def* %insn )		; <%struct.rtx_def*> [#uses=1]
+	%tmp98 = tail call  %struct.rtx_def* @walk_fixup_memory_subreg( %struct.rtx_def* %tmp95, %struct.rtx_def* %insn )		; <%struct.rtx_def*> [#uses=1]
 	%tmp101 = getelementptr %struct.rtvec_def* %tmp81, i32 0, i32 1, i32 %j.019, i32 0		; <i32*> [#uses=1]
 	%tmp98.c = ptrtoint %struct.rtx_def* %tmp98 to i32		; <i32> [#uses=1]
 	store i32 %tmp98.c, i32* %tmp101
diff --git a/test/CodeGen/X86/2008-02-22-ReMatBug.ll b/test/CodeGen/X86/2008-02-22-ReMatBug.ll
index fd97b61..f78d526 100644
--- a/test/CodeGen/X86/2008-02-22-ReMatBug.ll
+++ b/test/CodeGen/X86/2008-02-22-ReMatBug.ll
@@ -4,7 +4,7 @@
 
 	%struct.quad_struct = type { i32, i32, %struct.quad_struct*, %struct.quad_struct*, %struct.quad_struct*, %struct.quad_struct*, %struct.quad_struct* }
 
-define fastcc %struct.quad_struct* @MakeTree(i32 %size, i32 %center_x, i32 %center_y, i32 %lo_proc, i32 %hi_proc, %struct.quad_struct* %parent, i32 %ct, i32 %level) nounwind  {
+define  %struct.quad_struct* @MakeTree(i32 %size, i32 %center_x, i32 %center_y, i32 %lo_proc, i32 %hi_proc, %struct.quad_struct* %parent, i32 %ct, i32 %level) nounwind  {
 entry:
 	br i1 true, label %bb43.i, label %bb.i
 
@@ -39,12 +39,12 @@ bb34:		; preds = %CheckOutside.exit20.i
 	%tmp15.reg2mem.0 = sdiv i32 %size, 2		; <i32> [#uses=7]
 	%tmp85 = sub i32 %center_y, %tmp15.reg2mem.0		; <i32> [#uses=2]
 	%tmp88 = sub i32 %center_x, %tmp15.reg2mem.0		; <i32> [#uses=2]
-	%tmp92 = tail call fastcc %struct.quad_struct* @MakeTree( i32 %tmp15.reg2mem.0, i32 %tmp88, i32 %tmp85, i32 0, i32 %hi_proc, %struct.quad_struct* null, i32 2, i32 0 ) nounwind 		; <%struct.quad_struct*> [#uses=0]
+	%tmp92 = tail call  %struct.quad_struct* @MakeTree( i32 %tmp15.reg2mem.0, i32 %tmp88, i32 %tmp85, i32 0, i32 %hi_proc, %struct.quad_struct* null, i32 2, i32 0 ) nounwind 		; <%struct.quad_struct*> [#uses=0]
 	%tmp99 = add i32 0, %hi_proc		; <i32> [#uses=1]
 	%tmp100 = sdiv i32 %tmp99, 2		; <i32> [#uses=1]
-	%tmp110 = tail call fastcc %struct.quad_struct* @MakeTree( i32 %tmp15.reg2mem.0, i32 0, i32 %tmp85, i32 0, i32 %tmp100, %struct.quad_struct* null, i32 3, i32 0 ) nounwind 		; <%struct.quad_struct*> [#uses=0]
+	%tmp110 = tail call  %struct.quad_struct* @MakeTree( i32 %tmp15.reg2mem.0, i32 0, i32 %tmp85, i32 0, i32 %tmp100, %struct.quad_struct* null, i32 3, i32 0 ) nounwind 		; <%struct.quad_struct*> [#uses=0]
 	%tmp122 = add i32 %tmp15.reg2mem.0, %center_y		; <i32> [#uses=2]
-	%tmp129 = tail call fastcc %struct.quad_struct* @MakeTree( i32 %tmp15.reg2mem.0, i32 0, i32 %tmp122, i32 0, i32 0, %struct.quad_struct* null, i32 1, i32 0 ) nounwind 		; <%struct.quad_struct*> [#uses=0]
-	%tmp147 = tail call fastcc %struct.quad_struct* @MakeTree( i32 %tmp15.reg2mem.0, i32 %tmp88, i32 %tmp122, i32 %lo_proc, i32 0, %struct.quad_struct* null, i32 0, i32 0 ) nounwind 		; <%struct.quad_struct*> [#uses=0]
+	%tmp129 = tail call  %struct.quad_struct* @MakeTree( i32 %tmp15.reg2mem.0, i32 0, i32 %tmp122, i32 0, i32 0, %struct.quad_struct* null, i32 1, i32 0 ) nounwind 		; <%struct.quad_struct*> [#uses=0]
+	%tmp147 = tail call  %struct.quad_struct* @MakeTree( i32 %tmp15.reg2mem.0, i32 %tmp88, i32 %tmp122, i32 %lo_proc, i32 0, %struct.quad_struct* null, i32 0, i32 0 ) nounwind 		; <%struct.quad_struct*> [#uses=0]
 	unreachable
 }
diff --git a/test/CodeGen/X86/coalescer-commute3.ll b/test/CodeGen/X86/coalescer-commute3.ll
index ea75637..7d4a80a 100644
--- a/test/CodeGen/X86/coalescer-commute3.ll
+++ b/test/CodeGen/X86/coalescer-commute3.ll
@@ -2,7 +2,7 @@
 
 	%struct.quad_struct = type { i32, i32, %struct.quad_struct*, %struct.quad_struct*, %struct.quad_struct*, %struct.quad_struct*, %struct.quad_struct* }
 
-define fastcc i32 @perimeter(%struct.quad_struct* %tree, i32 %size) nounwind  {
+define  i32 @perimeter(%struct.quad_struct* %tree, i32 %size) nounwind  {
 entry:
 	switch i32 %size, label %UnifiedReturnBlock [
 		 i32 2, label %bb
@@ -10,8 +10,8 @@ entry:
 	]
 
 bb:		; preds = %entry
-	%tmp31 = tail call fastcc i32 @perimeter( %struct.quad_struct* null, i32 0 ) nounwind 		; <i32> [#uses=1]
-	%tmp40 = tail call fastcc i32 @perimeter( %struct.quad_struct* null, i32 0 ) nounwind 		; <i32> [#uses=1]
+	%tmp31 = tail call  i32 @perimeter( %struct.quad_struct* null, i32 0 ) nounwind 		; <i32> [#uses=1]
+	%tmp40 = tail call  i32 @perimeter( %struct.quad_struct* null, i32 0 ) nounwind 		; <i32> [#uses=1]
 	%tmp33 = add i32 0, %tmp31		; <i32> [#uses=1]
 	%tmp42 = add i32 %tmp33, %tmp40		; <i32> [#uses=1]
 	ret i32 %tmp42
diff --git a/test/CodeGen/X86/combine-lds.ll b/test/CodeGen/X86/combine-lds.ll
index 5e0ad99..a78a042 100644
--- a/test/CodeGen/X86/combine-lds.ll
+++ b/test/CodeGen/X86/combine-lds.ll
@@ -1,7 +1,6 @@
-; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 | grep movsd
-; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 | grep mov | count 1
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 | grep fldl | count 1
 
-define fastcc double @doload64(i64 %x) nounwind  {
+define double @doload64(i64 %x) nounwind  {
 	%tmp717 = bitcast i64 %x to double
 	ret double %tmp717
 }
diff --git a/test/CodeGen/X86/fastcc.ll b/test/CodeGen/X86/fastcc.ll
new file mode 100644
index 0000000..13068ba
--- /dev/null
+++ b/test/CodeGen/X86/fastcc.ll
@@ -0,0 +1,19 @@
+; RUN: llvm-as < %s | llc -mtriple=i686-apple-darwin | grep mov | grep ecx | grep 0
+; RUN: llvm-as < %s | llc -mtriple=i686-apple-darwin | grep mov | grep xmm0 | grep 16
+
+@d = external global double		; <double*> [#uses=1]
+@c = external global double		; <double*> [#uses=1]
+@b = external global double		; <double*> [#uses=1]
+@a = external global double		; <double*> [#uses=1]
+
+define i32 @foo() nounwind {
+entry:
+	%0 = load double* @d, align 8		; <double> [#uses=1]
+	%1 = load double* @c, align 8		; <double> [#uses=1]
+	%2 = load double* @b, align 8		; <double> [#uses=1]
+	%3 = load double* @a, align 8		; <double> [#uses=1]
+	tail call fastcc void @bar( i32 0, i32 1, i32 2, double 1.000000e+00, double %3, double %2, double %1, double %0 ) nounwind
+	ret i32 0
+}
+
+declare fastcc void @bar(i32, i32, i32, double, double, double, double, double)