Add support for passing SSE vectors in registers.

This patch adds support for 6 XMM registers on x86-64 which overlap with the F
and D registers and may hold 128-bit wide SIMD vectors. Because there is not a
good way to attach type information to STG registers, we aggressively bitcast in
the LLVM back-end.

4 files changed, 82 insertions, 12 deletions

diff --git a/includes/CodeGen.Platform.hs b/includes/CodeGen.Platform.hs
index 14642bd1c5..beff19601d 100644
--- a/includes/CodeGen.Platform.hs
+++ b/includes/CodeGen.Platform.hs
@@ -388,36 +388,54 @@ activeStgRegs = [
 #ifdef REG_D1
     ,DoubleReg 1
 #endif
+#ifdef REG_XMM1
+    ,XmmReg 1
+#endif
 #ifdef REG_F2
     ,FloatReg 2
 #endif
 #ifdef REG_D2
     ,DoubleReg 2
 #endif
+#ifdef REG_XMM2
+    ,XmmReg 2
+#endif
 #ifdef REG_F3
     ,FloatReg 3
 #endif
 #ifdef REG_D3
     ,DoubleReg 3
 #endif
+#ifdef REG_XMM3
+    ,XmmReg 3
+#endif
 #ifdef REG_F4
     ,FloatReg 4
 #endif
 #ifdef REG_D4
     ,DoubleReg 4
 #endif
+#ifdef REG_XMM4
+    ,XmmReg 4
+#endif
 #ifdef REG_F5
     ,FloatReg 5
 #endif
 #ifdef REG_D5
     ,DoubleReg 5
 #endif
+#ifdef REG_XMM5
+    ,XmmReg 5
+#endif
 #ifdef REG_F6
     ,FloatReg 6
 #endif
 #ifdef REG_D6
     ,DoubleReg 6
 #endif
+#ifdef REG_XMM6
+    ,XmmReg 6
+#endif
 #else /* MAX_REAL_SSE_REG == 0 */
 #ifdef REG_F1
     ,FloatReg 1
@@ -569,6 +587,14 @@ globalRegMaybe (DoubleReg 6)            =
                                           Just (RealRegSingle REG_D6)
 #  endif
 # endif
+#if MAX_REAL_SSE_REG != 0
+globalRegMaybe (XmmReg 1)               = Just (RealRegSingle REG_XMM1)
+globalRegMaybe (XmmReg 2)               = Just (RealRegSingle REG_XMM2)
+globalRegMaybe (XmmReg 3)               = Just (RealRegSingle REG_XMM3)
+globalRegMaybe (XmmReg 4)               = Just (RealRegSingle REG_XMM4)
+globalRegMaybe (XmmReg 5)               = Just (RealRegSingle REG_XMM5)
+globalRegMaybe (XmmReg 6)               = Just (RealRegSingle REG_XMM6)
+# endif
 # ifdef REG_Sp
 globalRegMaybe Sp                       = Just (RealRegSingle REG_Sp)
 # endif
diff --git a/includes/stg/MachRegs.h b/includes/stg/MachRegs.h
index 6dc81f63c1..76bdb1fc21 100644
--- a/includes/stg/MachRegs.h
+++ b/includes/stg/MachRegs.h
@@ -167,12 +167,12 @@
 #define REG_D5    xmm5
 #define REG_D6    xmm6
 
-#define REG_SSE1    xmm1
-#define REG_SSE2    xmm2
-#define REG_SSE3    xmm3
-#define REG_SSE4    xmm4
-#define REG_SSE5    xmm5
-#define REG_SSE6    xmm6
+#define REG_XMM1    xmm1
+#define REG_XMM2    xmm2
+#define REG_XMM3    xmm3
+#define REG_XMM4    xmm4
+#define REG_XMM5    xmm5
+#define REG_XMM6    xmm6
 
 #if !defined(mingw32_HOST_OS)
 #define CALLER_SAVES_R3
@@ -199,13 +199,13 @@
 #define CALLER_SAVES_D6
 #endif
 
-#define CALLER_SAVES_SSE1
-#define CALLER_SAVES_SSE2
-#define CALLER_SAVES_SSE3
-#define CALLER_SAVES_SSE4
-#define CALLER_SAVES_SSE5
+#define CALLER_SAVES_XMM1
+#define CALLER_SAVES_XMM2
+#define CALLER_SAVES_XMM3
+#define CALLER_SAVES_XMM4
+#define CALLER_SAVES_XMM5
 #if !defined(mingw32_HOST_OS)
-#define CALLER_SAVES_SSE6
+#define CALLER_SAVES_XMM6
 #endif
 
 #define MAX_REAL_VANILLA_REG 6
diff --git a/includes/stg/Regs.h b/includes/stg/Regs.h
index fd1577e71a..10ae2851ac 100644
--- a/includes/stg/Regs.h
+++ b/includes/stg/Regs.h
@@ -81,6 +81,12 @@ typedef struct {
   StgDouble 	  rD4;
   StgDouble 	  rD5;
   StgDouble 	  rD6;
+  StgWord128 	  rXMM1;
+  StgWord128 	  rXMM2;
+  StgWord128 	  rXMM3;
+  StgWord128 	  rXMM4;
+  StgWord128 	  rXMM5;
+  StgWord128 	  rXMM6;
   StgWord64       rL1;
   StgPtr 	  rSp;
   StgPtr 	  rSpLim;
@@ -270,6 +276,42 @@ GLOBAL_REG_DECL(StgDouble,D6,REG_D6)
 #define D6 (BaseReg->rD6)
 #endif
 
+#if defined(REG_XMM1) && !defined(NO_GLOBAL_REG_DECLS)
+GLOBAL_REG_DECL(StgWord128,XMM1,REG_XMM1)
+#else
+#define XMM1 (BaseReg->rXMM1)
+#endif
+
+#if defined(REG_XMM2) && !defined(NO_GLOBAL_REG_DECLS)
+GLOBAL_REG_DECL(StgWord128,XMM2,REG_XMM2)
+#else
+#define XMM2 (BaseReg->rXMM2)
+#endif
+
+#if defined(REG_XMM3) && !defined(NO_GLOBAL_REG_DECLS)
+GLOBAL_REG_DECL(StgWord128,XMM3,REG_XMM3)
+#else
+#define XMM3 (BaseReg->rXMM3)
+#endif
+
+#if defined(REG_XMM4) && !defined(NO_GLOBAL_REG_DECLS)
+GLOBAL_REG_DECL(StgWord128,XMM4,REG_XMM4)
+#else
+#define XMM4 (BaseReg->rXMM4)
+#endif
+
+#if defined(REG_XMM5) && !defined(NO_GLOBAL_REG_DECLS)
+GLOBAL_REG_DECL(StgWord128,XMM5,REG_XMM5)
+#else
+#define XMM5 (BaseReg->rXMM5)
+#endif
+
+#if defined(REG_XMM6) && !defined(NO_GLOBAL_REG_DECLS)
+GLOBAL_REG_DECL(StgWord128,XMM6,REG_XMM6)
+#else
+#define XMM6 (BaseReg->rXMM6)
+#endif
+
 #if defined(REG_L1) && !defined(NO_GLOBAL_REG_DECLS)
 GLOBAL_REG_DECL(StgWord64,L1,REG_L1)
 #else
diff --git a/includes/stg/Types.h b/includes/stg/Types.h
index d6bdc9042b..ccc06a175b 100644
--- a/includes/stg/Types.h
+++ b/includes/stg/Types.h
@@ -83,6 +83,8 @@ typedef unsigned long long int StgWord64;
 #error cannot find a way to define StgInt64
 #endif
 
+typedef struct { StgWord64 h; StgWord64 l; } StgWord128;
+
 /*
  * Define the standard word size we'll use on this machine: make it
  * big enough to hold a pointer.