add and test stack load/store

Change-Id: Ie0d29e31bd8c156ecd46cd658b5a4c53d8d2e11d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/282115
Reviewed-by: Herb Derby <[email protected]>
Commit-Queue: Mike Klein <[email protected]>

Author: Mike Klein

src/core/SkVM.cpp

@@ -1934,6 +1934,18 @@ namespace skvm {
         this->byte(mod_rm(Mod::Indirect, src&7, dst&7));
     }
 
+    void Assembler::stack_load_store(int prefix, int map, int opcode, Ymm ymm, int off) {
+        VEX v = vex(0, ymm>>3, 0, rsp>>3/*i.e. 0*/,
+                    map, 0, /*ymm?*/1, prefix);
+        this->bytes(v.bytes, v.len);
+        this->byte(opcode);
+        this->byte(mod_rm(mod(off), ymm&7, rsp/*use SIB*/));
+        this->byte(sib(ONE, rsp/*no index*/, rsp));
+        this->bytes(&off, imm_bytes(mod(off)));
+    }
+    void Assembler::vmovups(Ymm dst, int off) { this->stack_load_store(0, 0x0f, 0x10, dst,off); }
+    void Assembler::vmovups(int off, Ymm src) { this->stack_load_store(0, 0x0f, 0x11, src,off); }
+
     void Assembler::vmovq(GP64 dst, Xmm src) {
         int prefix = 0x66,
             map    = 0x0f,
@@ -1986,7 +1998,7 @@ namespace skvm {
                     map, 0, /*ymm?*/0, prefix);
         this->bytes(v.bytes, v.len);
         this->byte(opcode);
-        this->byte(mod_rm(Mod::Indirect, dst&7, rsp));
+        this->byte(mod_rm(Mod::Indirect, dst&7, rsp/*use SIB*/));
         this->byte(sib(scale, index&7, base&7));
     }
 
@@ -2082,7 +2094,7 @@ namespace skvm {
                     map, mask, /*ymm?*/1, prefix);
         this->bytes(v.bytes, v.len);
         this->byte(opcode);
-        this->byte(mod_rm(Mod::Indirect, dst&7, rsp));
+        this->byte(mod_rm(Mod::Indirect, dst&7, rsp/*use SIB*/));
         this->byte(sib(scale, ix&7, base&7));
     }
 

src/core/SkVM.h

@@ -144,6 +144,7 @@ namespace skvm {
         void vbroadcastss(Ymm dst, Xmm src);
         void vbroadcastss(Ymm dst, GP64 ptr, int off);  // dst = *(ptr+off)
 
+        void vmovups  (Ymm dst, int  imm);   // dst = *(sp + imm)
         void vmovups  (Ymm dst, GP64 ptr);   // dst = *ptr, 256-bit
         void vpmovzxwd(Ymm dst, GP64 ptr);   // dst = *ptr, 128-bit, each uint16_t expanded to int
         void vpmovzxbd(Ymm dst, GP64 ptr);   // dst = *ptr,  64-bit, each uint8_t  expanded to int
@@ -152,6 +153,7 @@ namespace skvm {
         enum Scale { ONE, TWO, FOUR, EIGHT };
         void vmovd(Xmm dst, Scale, GP64 index, GP64 base);   // dst = *(base + scale*index),  32-bit
 
+        void vmovups(int  imm, Ymm src);     // *(sp + imm) = src
         void vmovups(GP64 ptr, Ymm src);     // *ptr = src, 256-bit
         void vmovups(GP64 ptr, Xmm src);     // *ptr = src, 128-bit
         void vmovq  (GP64 ptr, Xmm src);     // *ptr = src,  64-bit
@@ -270,6 +272,8 @@ namespace skvm {
 
         // *ptr = ymm or ymm = *ptr, depending on opcode.
         void load_store(int prefix, int map, int opcode, Ymm ymm, GP64 ptr);
+        // *(sp+off) = ymm or ymm = *(sp+off), depending on opcode.
+        void stack_load_store(int prefix, int map, int opcode, Ymm ymm, int off);
 
         // Opcode for 3-arguments ops is split between hi and lo:
         //    [11 bits hi] [5 bits m] [6 bits lo] [5 bits n] [5 bits d]

tests/SkVMTest.cpp

@@ -1402,6 +1402,24 @@ DEF_TEST(SkVM_Assembler, r) {
         0xc5,     0x79,   0xd6,  0b00'111'010,
     });
 
+    test_asm(r, [&](A& a) {
+        a.vmovups(A::ymm5,   0);
+        a.vmovups(A::ymm5,  64);
+        a.vmovups(A::ymm5, 128);
+
+        a.vmovups(  0, A::ymm5);
+        a.vmovups( 64, A::ymm5);
+        a.vmovups(128, A::ymm5);
+    },{
+        0xc5,0xfc,0x10,0x2c,0x24,
+        0xc5,0xfc,0x10,0x6c,0x24,0x40,
+        0xc5,0xfc,0x10,0xac,0x24,0x80,0x00,0x00,0x00,
+
+        0xc5,0xfc,0x11,0x2c,0x24,
+        0xc5,0xfc,0x11,0x6c,0x24,0x40,
+        0xc5,0xfc,0x11,0xac,0x24,0x80,0x00,0x00,0x00,
+    });
+
     test_asm(r, [&](A& a) {
         a.movzbl(A::rax, A::rsi, 0);   // Low registers for src and dst.
         a.movzbl(A::rax, A::r8,  0);   // High src register.