//===---------------------------------------------------------------------===//
-Think about doing i64 math in SSE regs.
+Think about doing i64 math in SSE regs on x86-32.
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
-For this:
-
-#include <emmintrin.h>
-void test(__m128d *r, __m128d *A, double B) {
- *r = _mm_loadl_pd(*A, &B);
-}
-
-We generates:
-
- subl $12, %esp
- movsd 24(%esp), %xmm0
- movsd %xmm0, (%esp)
- movl 20(%esp), %eax
- movapd (%eax), %xmm0
- movlpd (%esp), %xmm0
- movl 16(%esp), %eax
- movapd %xmm0, (%eax)
- addl $12, %esp
- ret
-
-icc generates:
-
- movl 4(%esp), %edx #3.6
- movl 8(%esp), %eax #3.6
- movapd (%eax), %xmm0 #4.22
- movlpd 12(%esp), %xmm0 #4.8
- movapd %xmm0, (%edx) #4.3
- ret #5.1
-
-So icc is smart enough to know that B is in memory so it doesn't load it and
-store it back to stack.
-
-This should be fixed by eliminating the llvm.x86.sse2.loadl.pd intrinsic,
-lowering it to a load+insertelement instead. Already match the load+shuffle
-as movlpd, so this should be easy. We already get optimal code for:
-
-define void @test2(<2 x double>* %r, <2 x double>* %A, double %B) {
-entry:
- %tmp2 = load <2 x double>* %A, align 16
- %tmp8 = insertelement <2 x double> %tmp2, double %B, i32 0
- store <2 x double> %tmp8, <2 x double>* %r, align 16
- ret void
-}
-
-//===---------------------------------------------------------------------===//
-
__m128d test1( __m128d A, __m128d B) {
return _mm_shuffle_pd(A, B, 0x3);
}
//===---------------------------------------------------------------------===//
-We should compile this:
-#include <xmmintrin.h>
-typedef union {
- int i[4];
- float f[4];
- __m128 v;
-} vector4_t;
-void swizzle (const void *a, vector4_t * b, vector4_t * c) {
- b->v = _mm_loadl_pi (b->v, (__m64 *) a);
- c->v = _mm_loadl_pi (c->v, ((__m64 *) a) + 1);
-}
-
-to:
-
-_swizzle:
- movl 4(%esp), %eax
- movl 8(%esp), %edx
- movl 12(%esp), %ecx
- movlps (%eax), %xmm0
- movlps %xmm0, (%edx)
- movlps 8(%eax), %xmm0
- movlps %xmm0, (%ecx)
- ret
-
-not:
-
-swizzle:
- movl 8(%esp), %eax
- movaps (%eax), %xmm0
- movl 4(%esp), %ecx
- movlps (%ecx), %xmm0
- movaps %xmm0, (%eax)
- movl 12(%esp), %eax
- movaps (%eax), %xmm0
- movlps 8(%ecx), %xmm0
- movaps %xmm0, (%eax)
- ret
-
-//===---------------------------------------------------------------------===//
-
We should materialize vector constants like "all ones" and "signbit" with
code like:
In sse4 mode, we could use insertps to make both better.
+Here's another testcase that could use insertps [mem]:
+
+#include <xmmintrin.h>
+extern float x2, x3;
+__m128 foo1 (float x1, float x4) {
+ return _mm_set_ps (x2, x1, x3, x4);
+}
+
+gcc mainline compiles it to:
+
+foo1:
+ insertps $0x10, x2(%rip), %xmm0
+ insertps $0x10, x3(%rip), %xmm1
+ movaps %xmm1, %xmm2
+ movlhps %xmm0, %xmm2
+ movaps %xmm2, %xmm0
+ ret
+
//===---------------------------------------------------------------------===//
+We compile vector multiply-by-constant into poor code:
+
+define <4 x i32> @f(<4 x i32> %i) nounwind {
+ %A = mul <4 x i32> %i, < i32 10, i32 10, i32 10, i32 10 >
+ ret <4 x i32> %A
+}
+
+On targets without SSE4.1, this compiles into:
+
+LCPI1_0: ## <4 x i32>
+ .long 10
+ .long 10
+ .long 10
+ .long 10
+ .text
+ .align 4,0x90
+ .globl _f
+_f:
+ pshufd $3, %xmm0, %xmm1
+ movd %xmm1, %eax
+ imull LCPI1_0+12, %eax
+ movd %eax, %xmm1
+ pshufd $1, %xmm0, %xmm2
+ movd %xmm2, %eax
+ imull LCPI1_0+4, %eax
+ movd %eax, %xmm2
+ punpckldq %xmm1, %xmm2
+ movd %xmm0, %eax
+ imull LCPI1_0, %eax
+ movd %eax, %xmm1
+ movhlps %xmm0, %xmm0
+ movd %xmm0, %eax
+ imull LCPI1_0+8, %eax
+ movd %eax, %xmm0
+ punpckldq %xmm0, %xmm1
+ movaps %xmm1, %xmm0
+ punpckldq %xmm2, %xmm0
+ ret
+
+It would be better to synthesize integer vector multiplication by constants
+using shifts and adds, pslld and paddd here. And even on targets with SSE4.1,
+simple cases such as multiplication by powers of two would be better as
+vector shifts than as multiplications.
+
+//===---------------------------------------------------------------------===//
+
+We compile this:
+
+__m128i
+foo2 (char x)
+{
+ return _mm_set_epi8 (1, 0, 0, 0, 0, 0, 0, 0, 0, x, 0, 1, 0, 0, 0, 0);
+}
+
+into:
+ movl $1, %eax
+ xorps %xmm0, %xmm0
+ pinsrw $2, %eax, %xmm0
+ movzbl 4(%esp), %eax
+ pinsrw $3, %eax, %xmm0
+ movl $256, %eax
+ pinsrw $7, %eax, %xmm0
+ ret
+
+
+gcc-4.2:
+ subl $12, %esp
+ movzbl 16(%esp), %eax
+ movdqa LC0, %xmm0
+ pinsrw $3, %eax, %xmm0
+ addl $12, %esp
+ ret
+ .const
+ .align 4
+LC0:
+ .word 0
+ .word 0
+ .word 1
+ .word 0
+ .word 0
+ .word 0
+ .word 0
+ .word 256
+
+With SSE4, it should be
+ movdqa .LC0(%rip), %xmm0
+ pinsrb $6, %edi, %xmm0
+
+//===---------------------------------------------------------------------===//
+
+We should transform a shuffle of two vectors of constants into a single vector
+of constants. Also, insertelement of a constant into a vector of constants
+should also result in a vector of constants. e.g. 2008-06-25-VecISelBug.ll.
+
+We compiled it to something horrible:
+
+ .align 4
+LCPI1_1: ## float
+ .long 1065353216 ## float 1
+ .const
+
+ .align 4
+LCPI1_0: ## <4 x float>
+ .space 4
+ .long 1065353216 ## float 1
+ .space 4
+ .long 1065353216 ## float 1
+ .text
+ .align 4,0x90
+ .globl _t
+_t:
+ xorps %xmm0, %xmm0
+ movhps LCPI1_0, %xmm0
+ movss LCPI1_1, %xmm1
+ movaps %xmm0, %xmm2
+ shufps $2, %xmm1, %xmm2
+ shufps $132, %xmm2, %xmm0
+ movaps %xmm0, 0
+
+//===---------------------------------------------------------------------===//
+rdar://5907648
+
+This function:
+
+float foo(unsigned char x) {
+ return x;
+}
+
+compiles to (x86-32):
+
+define float @foo(i8 zeroext %x) nounwind {
+ %tmp12 = uitofp i8 %x to float ; <float> [#uses=1]
+ ret float %tmp12
+}
+
+compiles to:
+
+_foo:
+ subl $4, %esp
+ movzbl 8(%esp), %eax
+ cvtsi2ss %eax, %xmm0
+ movss %xmm0, (%esp)
+ flds (%esp)
+ addl $4, %esp
+ ret
+
+We should be able to use:
+ cvtsi2ss 8($esp), %xmm0
+since we know the stack slot is already zext'd.
+
+//===---------------------------------------------------------------------===//
+
+Consider using movlps instead of movsd to implement (scalar_to_vector (loadf64))
+when code size is critical. movlps is slower than movsd on core2 but it's one
+byte shorter.
+
+//===---------------------------------------------------------------------===//
+
+We should use a dynamic programming based approach to tell when using FPStack
+operations is cheaper than SSE. SciMark montecarlo contains code like this
+for example:
+
+double MonteCarlo_num_flops(int Num_samples) {
+ return ((double) Num_samples)* 4.0;
+}
+
+In fpstack mode, this compiles into:
+
+LCPI1_0:
+ .long 1082130432 ## float 4.000000e+00
+_MonteCarlo_num_flops:
+ subl $4, %esp
+ movl 8(%esp), %eax
+ movl %eax, (%esp)
+ fildl (%esp)
+ fmuls LCPI1_0
+ addl $4, %esp
+ ret
+
+in SSE mode, it compiles into significantly slower code:
+
+_MonteCarlo_num_flops:
+ subl $12, %esp
+ cvtsi2sd 16(%esp), %xmm0
+ mulsd LCPI1_0, %xmm0
+ movsd %xmm0, (%esp)
+ fldl (%esp)
+ addl $12, %esp
+ ret
+
+There are also other cases in scimark where using fpstack is better, it is
+cheaper to do fld1 than load from a constant pool for example, so
+"load, add 1.0, store" is better done in the fp stack, etc.
+
+//===---------------------------------------------------------------------===//