Merge pull request #130 from ThePortlandGroup/nv_stage

Pull 2017-07-07T07-12 Recent NVIDIA Changes

Author: sscalpone

tools/flang2/docs/xflag.n

@@ -5317,6 +5317,8 @@ The -ffast-math command-line option is present.
 Enable tuning code for -Minline.
 .XF "221:"
 This sets the maximum caller function size into which to Minline.
+.XF "222:"
+Functions whose size if smaller than this value will get inlined by Minline.
 
 .XF "248:"
 OpenMP Threadprivate TLS/TPvector implementation control.
@@ -5332,6 +5334,3 @@ Set number of bigbuffers for multi-buffer memory management for AMD GPU.
 (moved to 202)
 .XF "251:"
 (NOT available - check declaration in global.h for flg.x[], all compilers)
-
-.XF "260:"
-Functions whose size if smaller than this value will get inlined by Minline.

tools/flang2/flang2exe/bih.h

@@ -68,7 +68,7 @@ typedef struct {
                                */
       unsigned parsect : 1;   /* bih belongs to a parallel section */
       unsigned ujres : 1;     /* bih contains ujresidual start & count info */
-      unsigned sse : 1;       /* bih contains sse code */
+      unsigned simd : 1;       /* bih contains simd code */
 
       unsigned ldvol : 1; /* bih contains a load from volatile space */
       unsigned stvol : 1; /* bih contains a store to volatile space */
@@ -190,7 +190,7 @@ typedef struct {
 #define BIH_PARLOOP(i) bihb.stg_base[i].flags.bits.parloop
 #define BIH_PARSECT(i) bihb.stg_base[i].flags.bits.parsect
 #define BIH_UJRES(i) bihb.stg_base[i].flags.bits.ujres
-#define BIH_SSE(i) bihb.stg_base[i].flags.bits.sse
+#define BIH_SIMD(i) bihb.stg_base[i].flags.bits.simd
 #define BIH_LDVOL(i) bihb.stg_base[i].flags.bits.ldvol
 #define BIH_STVOL(i) bihb.stg_base[i].flags.bits.stvol
 #define BIH_ASM(i) bihb.stg_base[i].flags.bits.gasm

tools/flang2/flang2exe/bihutil.c

@@ -227,7 +227,7 @@ merge_bih_flags(int to_bih, int fm_bih)
   BIH_QJSR(to_bih) = BIH_QJSR(to_bih) | BIH_QJSR(fm_bih);
   BIH_INVIF(to_bih) = BIH_INVIF(to_bih) | BIH_INVIF(fm_bih);
   BIH_NOINVIF(to_bih) = BIH_NOINVIF(to_bih) | BIH_NOINVIF(fm_bih);
-  BIH_SSE(to_bih) = BIH_SSE(to_bih) | BIH_SSE(fm_bih);
+  BIH_SIMD(to_bih) = BIH_SIMD(to_bih) | BIH_SIMD(fm_bih);
   BIH_RESID(to_bih) = BIH_RESID(to_bih) | BIH_RESID(fm_bih);
   BIH_VCAND(to_bih) = BIH_VCAND(to_bih) | BIH_VCAND(fm_bih);
   BIH_MIDIOM(to_bih) = BIH_MIDIOM(to_bih) | BIH_MIDIOM(fm_bih);

tools/flang2/flang2exe/iltutil.c

@@ -337,8 +337,8 @@ dump_ilt(FILE *ff, int bihx)
     fprintf(ff, " PARLOOP");
   if (BIH_UJRES(bihx))
     fprintf(ff, " UJRES");
-  if (BIH_SSE(bihx))
-    fprintf(ff, " SSE");
+  if (BIH_SIMD(bihx))
+    fprintf(ff, " SIMD");
   if (BIH_LDVOL(bihx))
     fprintf(ff, " LDVOL");
   if (BIH_STVOL(bihx))

tools/flang2/flang2exe/mwd.c

@@ -4793,7 +4793,7 @@ db(int block)
   putbit("parsect", BIH_PARSECT(block));
   putbit("resid", BIH_RESID(block));
   putbit("ujres", BIH_UJRES(block));
-  putbit("sse", BIH_SSE(block));
+  putbit("simd", BIH_SIMD(block));
   putbit("ldvol", BIH_LDVOL(block));
   putbit("stvol", BIH_STVOL(block));
   putbit("task", BIH_TASK(block));